Introduction

We live in a critical time for democracy around the world.

Over the last few decades, a narrative of inevitable, gradual progress towards a more democratic world has become first strained, and then just incompatible with the observable facts. As the 2025 V-Dem Democracy report said, we are seeing "weakening of democracy in some established liberal democracies, breakdown of democracy in countries that were democratic for most of the 21st century, as well as deepening of autocracy in already autocratic states". We need to be realists about the current situation, without being fatalistic. That same V-Dem report also points out that autocracy isn't a one-way train, that there are countries around the world bucking the trend, either gradually or dramatically. There remains work to do everywhere, and every small win makes a difference.

A key question for both technologists and pro-democracy actors is how technology fits into this picture, when (in different forms) it sits on both sides of the fight. We talk about what we do as pro-democracy tech because we believe that technology can strengthen and defend democracy — but also because we recognise that there is clearly such a thing as "anti-democracy tech". Modern autocracies are not just a resurgence of old ideas, but are interconnected movements that make significant use of digital technology in how they operate at home, and in attacks on democracies elsewhere. Technology provides new methods of democratic participation, deliberation and collaboration. It also creates new forms of oppression, control, and warfare. Part of winning the big fight is making sure we properly understand, utilise and/or resist these different approaches to technology.

Mark Carney identifies the key question of this period as being the balance between strategic autonomy and the efficiency gains of interdependence. How do we reduce the risk of our global interdependence being weaponised against us, without creating a brittle and poorer world where everyone is duplicating capacities instead of collaborating and building on each other's work?

This is a question to which the pro-democracy civic tech movement is well placed to contribute. Around the world there is a huge amount of reliance on technology provided by tech giants, but also strong practical traditions of open source work and collaboration across borders. We run TICTeC because we believe that in order to have the most impact, we are necessarily interconnected. There is value in building and learning from each other's work, both practically in terms of code and frameworks, but also in sharing ideas, inspiration, and ways to have a bigger impact in changing landscapes.

Our goal with this report is to explore different approaches to using technology in the service of democracy, and present practical guidance on making the most of the tools available. New AI technologies create potentially interesting opportunities for pro-democracy tech, while at the same time contributing to shifts in the landscape that require adaptation in themselves. We draw out examples from projects that exist and are doing good work, as well as where changing circumstances introduce new challenges to our collective work.

About mySociety and TICTeC

For the last 20 years mySociety has been one of the world's leading civic tech organisations, and for the last 10 years we have run the TICTeC (The Impacts of Civic Technology) conference and programme. From our base in the UK, running services used by millions of people, we work at the cutting edge of using technology to help people participate in decisions that make an impact in their lives.

We are part of a wider community of people working to create pro-democracy technology in different societies and contexts. We actively work to bring this community together through our TICTeC conference and networks. Our convening work helps people share code, patterns, and inspiration.

We believe that having an anchor in the global perspective helps civic tech organisations fight their local battles. It brings back tools and ideas to use at home, and reframes what can be an isolated experience chipping away at big problems with connections to a global community doing the same things.

In the last few years, we've been exploring the idea of ‘pro-democracy tech' as a way to define and bring together different areas of global civic tech, and connect clearly with the need to build better connections between technologists and the wider pro-democracy movement.

This report is an extension of that work, building on the presentation and conversations we've been having at TICTeC and with the wider community on what the current challenges are, and the wide range of approaches that are being explored.

What is pro-democracy tech?

Pro-democracy tech describes digital tools aimed at realising and defending democratic values.

A key motivation of this approach is that authoritarianism is not standing still, and is learning how to use technology to extend its surveillance and control over people. Democracy's reaction to this needs to be not to reject technology, but to use it to evolve and compete. Democracy needs to be fast, effective and popular, while not conceding that the only way to do this is by becoming more centralised and authoritarian itself.

Within this, there are two key activities:

• Defensive democratic tech – defending the open society: anti-corruption, anti-misinformation, etc.
• Constructive democratic tech – empowering technologies that build democratic fibre and capacity: participation and deliberation, community tools, civic response tech.

These are interconnected, and not hard divides. Defensive approaches safeguard the openness which democracy needs to function, while constructive approaches build the capacity of the engine of democratic progress.

There are tools and approaches that apply to both. Access to Information laws are both vital anti-corruption tools, and part of capacity building through lowering costs of accessing information. Democratic transparency organisations (parliamentary monitoring and similar) are both about increasing anti-corruption surveillance, and transforming the capacity and connections of democratic institutions.

Where they differ is in their approaches to new technological tools. Defensive democratic tech is in an arms war with autocrats. We have to keep moving and innovating to stay in place. It is reactive against a well-financed opponent, and needs to understand how to bend tools (often developed by those with deep pockets and their own motives) to democratic purposes.

Constructive democratic tech is less of a zero-sum game. Just as there are technological approaches that make authoritarianism much more effective, there are approaches that make democracy much more effective. Here the enemy is less tangible but omnipresent: inertia, low expectations, and a belief that things can't be better. The goal of this approach is developing civic capacity, where the extraordinary capacity and cognitive diversity in a democracy are fully enabled to work together to solve the big problems of the age.

Pro-democracy tech is a big umbrella covering many different kinds of projects, but we think it is important to have conversations bridging these areas, showing how different ideas fit into the big picture, and where practical tools and skills that are common across different approaches can be supported and encouraged.

What this report covers

This report is intended as a practical document for current civic tech practitioners, pro-democracy organisations interested in making more use of tech, and funders interested in the intersection of tech and democracy.

This report covers a lot of ground: it's split into four thematic areas sections with chapter listings below to help you decide which areas are most relevant to your work.

We want this report to be a practical jumping-off point for those seeking to understand more about this space. To learn more, explore the TICTeC Knowledge Hub which contains 10 years of TICTeC talks and in-depth material.

Pro democracy tech: practical approaches and challenges

These sections show how pro-democracy tech has been operationalised, and give examples of practical approaches to bridging technologists with people working in deliberative democracy and anti-corruption.

This section will be of most interest to people trying to understand how to connect the wider group of people doing pro-democracy work (in either democratic defence or democratic innovation) with the use of technology.

The arms race: approaches to defensive pro-democratic tech. Defensive pro-democratic tech is often in an arms race. New technology and innovation is required to keep pace, but often these same approaches are useful to those who seek to undermine democratic values. Getting ahead requires efficiencies in use of new technologies: building infrastructure, tools, and approaches that can empower many different actors at once. But at the same time, this creates new trade-offs and funding challenges in dependence on shared infrastructure. This chapter discusses obstacles and approaches to cybersecurity and anti-corruption approaches to defensive tech.

The new wave: Constructive democratic tech: technology can be used to support new forms of democracy, but this needs careful alignment between technology choices and the values the democratic innovation puts emphasis on. This chapter explores how tech approaches to deliberative democracy have adapted to what its practitioners value, providing a model for thinking about how technology can support different kinds of democratic approaches.

Communities of practice

This section covers what we've learned convening communities of practice around two different areas: Access to Information/Freedom of Information and parliamentary monitoring/democratic transparency. As well as convening the broad civic tech community through the TICTeC conference, we also wanted to experiment more with how we could convene and support projects working in similar areas across the world. We took a different approach for each community, looking for where we could best add unique value (especially from the technology side), and constructively overlap with other networks.

These chapters will be of most interest to people working/interested in Access to Information or parliamentary monitoring, but also those who are interested in different approaches to convening and sharing learning.

-

Access to Information as democratic infrastructure: in this chapter, Access to Information is reframed from an individual legal right into essential democratic infrastructure needed to make power legible and accountable in increasingly automated, outsourced, and opaque systems. This is supported by technical infrastructure tools like Alaveteli, collective and journalistic practice, and evolving approaches to measurement of performance and effectiveness.

Supporting global democratic transparency: in this chapter, we explore how we can support global democratic transparency by expanding beyond traditional parliamentary monitoring to a wider set of "democratic transparency projects". We believe that a focus on transferable technical innovation (including AI) fills a gap that helps projects adopt new technologies, while managing risks in an era of democratic backsliding.

Shaping the landscape

Civic tech projects need to balance when they need to be flexible (adapting to local contexts and the distribution channels that exist), and when they need to try to change that context: investing and laying down infrastructure that makes their (and others) work easier.

This section explores changing methods of distribution, how projects can use intermediaries to reach their audience, the role of foundational digital infrastructure in shaping what's possible, and how democratic actors should be seeking to influence and build that infrastructure.

Digital infrastructure: Foundations for democratic futures: Civic tech is often about building infrastructure, but rarely described in those terms. This chapter explores overlaps and connections between existing work and ideas of digital public infrastructure. Digital infrastructure has the capacity to either strengthen or erode democracy depending on its design, governance, and funding. As such, pro-democracy actors must deliberately build and sustain "democratic" infrastructure, while making a better case for its long-term impact.

Distributing democracy: Having impact through reaching mass audiences: in this chapter, we explore why scaling civic tech impact now requires strategies beyond "build a great website"— adapting to the decline of the open web and choosing deliberate ways to reach mass audiences through platforms, influencers, chatbots/AI, or alternative channels like email and print.

Using technology effectively

This is an obvious point, but successful pro-democratic technology involves making good choices in how to use technology effectively. These chapters explore frameworks and examples of approaches to AI technologies, as well as other considerations in making effective and sustainable technology choices.

Evaluating and understanding AI: in this chapter, we explore what modern AI (especially LLMs) can and can't do, when hype and black-box risks mean we should adopt it cautiously, and how to decide and evaluate when AI is a good fit. We propose a practical question framework, and disciplined model testing ("evals"), while weighing open vs closed models and long-term sustainability.

Sustainable tech: balancing ease of development with user experience: in this chapter, we explore the tension between the benefits of out-of-the-box solutions for civic tech teams, with the device constraints of potential users. We explore how pro-democracy tech teams can make sustainable technology choices under tight constraints; designing for low-spec users, selectively using performance/security-focused languages, and exploring low-cost approaches to LLMs.

The arms race: approaches to defensive pro-democratic tech

At TICTeC 2024, Marci Harris (Popvox Foundation) called her talk "Overcoming the Red Queen problem: ensuring democracy keeps pace with emerging technologies." The Red Queen is a character from the children's book Alice in Wonderland, who teaches Alice that "it takes all the running you can do, to keep in the same place."

This is one of the challenges for defensive approaches to pro democracy tech: it's an arms race, where new technology and innovation is required to keep pace, but often these same approaches are useful to those who seek to undermine democratic values. Getting ahead requires efficiencies in use of new technologies: building infrastructure, tools, and approaches that can empower many different actors at once.

In this chapter, we draw out two different sets of conversations we've been having at TICTeC conferences. One is about cyber attacks on civic tech and civil society actors themselves: how this imposes new costs, and how both procedural and technological approaches can help mitigate these risks.

The other is looking at uses of technology for anti-corruption purposes. This is an area where thinking about the infrastructure underpinning projects is especially important, with good access to data being an essential prerequisite for projects, and where there are already influential data standards that make transferring approaches and thinking between areas easier. AI technologies lower the cost of certain kinds of anti-corruption projects — but importantly, these still need to rest on a foundation of good data, and have clearly identified and articulated routes to impact.

Digital threats to pro democracy actors

Pro-democracy actors and democratic states are under increasing risk of significantly damaging cyberattacks. This includes both attacks from within countries, where digital threats are part of the toolset available to repressive states; and the international picture, where foreign cyberattacks on infrastructure and organisations are a hybrid warfare tactic.

We ran a TICTeC seminar, Defending democracy defenders, talking about different aspects of digital threats to civic tech and pro-democracy projects. Pro-democracy work might be a direct target (spear phishing, DDOS attacks), an opportunistic target (scattergun approach to phishing or vulnerability detection), or just generally have raised costs in terms of cyber-defence to mitigate risk.

A key message from this seminar is that defence against these kinds of digital threats needs both human and technological approaches — and a strong connection between the two to ensure that best security practices keep pace with changing threats, and that tools to mitigate risks are usable and effective by their target audiences.

Digital threats are not an isolated issue and fit into a general landscape of pressure and threats that shrink civic space. They are a growing tactic (among many offline ones) of repressive regimes and other threat actors that fit into this wider picture. Technical civil society organisations can be better placed than most to understand cyber-security risks. The Engine Room talked about their Cybersecurity Assessment Tool (CAT), where an explicit goal is to help non-technical CSOs understand and manage risk. This is based on a questionnaire format that helps lead into specific recommendations around managing broader risks, and how to operationalise risk management from different kinds of security approaches. The long term goal is that cybersecurity thinking comes not in response to a specific issue, but fits into broader strategy and organisational practice.

Digital tools can be part of how pro-democracy actors manage risk, but the important work is in building strong feedback loops between the users and potential creators of those tools. A speaker from the International Republication Institute (IRI) talked about their evolving perspective on how to support internet freedom tools, helping circumvent shutdowns, censorship or surveillance. In their programme they moved over time from an approach that was focused on the tools, to one that made much more use of user testing to build trust in tools for users, and to help developers ensure real world fit. Effective use of tech requires fixing coordination problems between funders, implementers and users of technology. Pro-democracy tech is not just "tech", but an attitude and approach to making sure that technology is available and useful to the pro-democracy actors who creators believe would benefit from it.

We also heard from Cloudflare about their Project Galileo, which provides free access to Cloudflare's infrastructure to "humanitarian organisations, civil society, and human rights defenders", providing technical protection against DDoS attacks and other approaches to cyber security. The idea is to reduce the amount of technical expertise small organisations need to be able to mitigate the risk of attack.

With a disclaimer that we benefit from this programme (with one of our high traffic services sitting behind Cloudflare's system), in terms of the big picture it's also worth noting we ran this seminar a day after the November Cloudflare outage which itself had a widespread effect on a range of web services. The price of reduced risk of targeted attack on any one organisation is a collective exposure to dependence on a small number of infrastructure providers.

Here the wider picture matters: we're not reducing the redundancy of the web for no reason; we're doing it because both the idea of the open web and those that most benefit from its democratising spirit are under attack. We have to make choices between different sets of risks to stay in the race. This is a problem we see more generally through democratic defence and anti-corruption work: we need constant innovation and investment just to maintain capacity. But also, efficiencies of scale (either through common infrastructure or sharing innovation) are an important part of trying to control costs while having more impact.

Use of technology in anti-corruption

Civic tech tools and approaches are an important part of anti-corruption work, especially at the international level. Technology is particularly effective for joining the dots between datasets and public data that cover different domains.

At TICTeC 2025 Reinaldo Chaves from Abraji, the Brazilian Association of Investigative Journalists, talked about how they have developed civic tech tools as an evolution of traditional investigative journalism. They have built tools like CruzaGrafos to join up different datasets and make relationships across them visible. This has led to more than 200 stories produced by Brazilian journalists using their platforms and data. Abraji creates an infrastructure and advocacy layer that helps journalism go further; working in this area they've also been part of important legal victories against SLAPPs and access to public spending data.

One form of work in this area is creating the data structures and infrastructure required to do good analysis. The Open Contracting Data Standard (OCDS) defines a common standard for procurement information, an ideal for the format in which governments release information. If they don't adhere to it, other forms of disclosures can be mapped to this data model. Having a standard approach makes it easier to combine datasets across jurisdictions and share common tools in the data analysis process.

The Budeshi tool, used in Nigeria and presented by Nnenna Eze from PPDC at TICTeC in 2024, links procurement data to public services. It initially converts procurement information in PDFs into OCDS formats to make analysis clearer, and supports five states to produce OCDS-compliant procurement portals. Free, open source tools like Cardinal, from the Open Contracting Partnership, which help map data quality issues and automate the incorporation of red flag detection, can be more easily moved between jurisdictions because of the underlying common format.

Similarly, the Beneficial Ownership Data Standard (BODS) defines a common standard for beneficial ownership data (disclosures of who ultimately owns companies). Moving away from structured data, OCCRP's Aleph toolset brings a huge amount of information together in one place to make it easy for journalists and investigators looking at a particular individual or company to find related relevant documents.

But the use of civic tech is not just about analysis of big government datasets; it can also help connect people's individual experiences with the bigger picture. In Thailand, having information in public helped people connect things they can see locally (a new set of lampposts) with the underlying procurement practices, revealing corruption through inflation in the price of lampposts. Ushahidi (a civic tech reporting framework) has been the basis of a number of projects around the world helping people report instances of corruption (such as needing to pay bribes). This is helping to create new data that turns people's individual experiences into a collective picture that can inform collective advocacy and analysis.

Use of AI technologies

In the anti-corruption space, AI technologies (machine learning and large language models) can accelerate and simplify certain kinds of analysis tasks. A briefing from the U4 Anti-Corruption Resource Centre covers a range of possible cases, including both machine learning and large language models — but also highlights some of the unique drawbacks of these approaches.

Machine learning

Machine learning approaches are useful when there is unstructured data (eg satellite pictures) that in principle contains information relevant to the desired structured output (eg the location of illegal mines), and there is a good training set describing connections between the unstructured data to the structured data.

In anti-corruption, this has been used for remote sensing and risk score analysis. For instance a 2018 Global Witness/Data Kind/NRGI project tried to use machine learning to identify mines from satellite images. This was successful in finding known mines, but with a high false positive rate. The WFF 2024's Forest Foresight project similarly explored identifying illegal deforestation activities through an ML approach.

Another use is to create risk scores based on a wide range of factors that might point to potential corruption or illegal activities. For instance, OjoPúblico developed a model to create risk scores for illegal timber sales in Peru. Risk scores are a useful way of expressing the kinds of outputs from machine learning, where no individual feature may be indicative on its own.

It should be noted that information trained on existing known illegal activities will be replicating the assumptions and tools that led to that investigation, making it more vulnerable to changing approaches, and possibly drawing assumptions based on irrelevant factors (for instance, if enforcement was stronger in one region, the model might learn that a project in that region should be treated as higher risk).

Large language models

Large language models (LLMs) are especially helpful for creating structured labels for unstructured data. Transparency International UK created an Open Access tool to bring information about lobbying into one place (as seen in their TICTeC 2025 presentation), both for their own analysis and so that other interested parties can understand what is happening. This process uses public disclosures, but where those are missing fills in gaps with volunteers and freedom of information, by providing a template to make requests covering specific authorities.

In this case, LLM approaches are used to classify meetings and identify those in particular areas (eg climate, health, defence). This helps create structured categories from messy data, and enables trend analysis over time. Drilling into particular areas (such as climate lobbying in Scotland), LLM approaches help pull out common themes in the arguments being made: such as where corporate lobbyists make arguments for delaying radical change, through talking about "energy mix" and local employment, for example.

Limits of technology

Anti-corruption analysis often has an "and then what" question that effective approaches need to address. While there may be a number of potential risk factors that narrow the area of search, investigations that can be taken all the way to prosecution require clear documentation rather than nebulous risk scores. Tools designed for use by enforcement teams need to meet a high standard. In the UK, software used by the Serious Fraud Office had problems with documents being inconsistently identified — a potentially significant problem for the legal process's document disclosure requirements.

This means that analysis might (correctly) identify likely corruption, but not be able to lead all the way through to enforcement action if there is not the capacity to deal with the outputs of the system. For instance the Alice procurement investigation system in Brazil fed into a system with "chronic staff shortages [that] left auditors overloaded with risk-alert emails, unable to triage everything, so potential red flags went unreviewed". Similarly, civil society projects in Brazil looking at representative spending disclosures shifted to a "name and shame" strategy only after trying to engage the wider accountability system and realising it was ineffective. As such, Albert Sanchez-Graells, writing for OCP, argues that we need to be realistic about what AI represents: improvements in efficiency around specific tasks, but not a revolution that can overcome fundamental resource limitation, or replace investment in the basics.

This means the challenge for civil society AI tools is not just in the creation and practical use, but in finding plausible routes to impact, whether this is making reports that feed into oversight systems, making a wider argument that the problem exists, or stating the case for proper resourcing and laws to combat an issue.

New challenges

Changing technological capabilities bring new capacity to anti-corruption work, but also pushback to successes, and the use of technology to enable corruption. LLMs can facilitate fraud in procurement by creating synthetic identities and fraudulent documentation, disguising risk indicators and making it harder for both automated and human checks to recognise fraudulent processes.

An ongoing challenge in data availability is the trade-off between privacy risks and the benefits of open data. Anti-corruption datasets (especially beneficial ownership datasets) can necessarily contain personal information of large numbers of people. While there are legitimate trade-offs here, this is also a useful argument for people seeking to conceal connections and ownership across borders. The Court of Justice of the European Union (CJEU) ruled that indiscriminate public access to beneficial ownership (BO) information was unlawful, setting back open registers of information in Europe. This judgement did allow that journalists and civil society groups can have a legitimate interest in accessing the data, but the way in which this decision is operationalised in practice can vary.

A more volatile world is leading to remilitarisation and increases in military procurement, which poses challenges for anti-corruption approaches. A Transparency International Defence & Security report argues that a norm of ‘defensive spending exceptionalism' means there are attempts (and successes) in pushing defense procurement outside normal transparency requirements for (often broadly defined) national security reasons. Shifts in procurement activity need new approaches to analysis of data and advocacy.

Conclusion

The picture that emerges across both defensive cybersecurity and anti-corruption civic tech is that Red Queen dilemma: maintaining capacity requires constant innovation against active opponents.

Building those capacities cost-effectively involves investment in common data approaches and tooling. But the same efficiencies that let small organisations survive and have impact, also create collective points of dependence. The challenge is how to balance investment (and funding) for higher level infrastructure, with ensuring that it is adequately connected to local capacity to have the strongest impact.

The new wave: Constructive democratic tech

One of the really exciting progressions we've seen at TICTeC over the last decade has been an expansion from technology that allows people to be "citizen sensors" for problems (making a street issue report, reporting corruption) to deliberative tech that builds on this function and also helps citizens collaborate and deliberate on how society should respond and act.

This is the "constructive" wing of pro-democracy technology, where an overlap between democratic innovation and civic tech is exploring new potential ways in which technology can facilitate democracy — and how it can unlock the one of the key benefits of democracy over other systems of government: it can make use of not just the labour, but the ingenuity and creativity of its citizens.

There are lots of different schools of thought on what "better democracy" looks like, and different groups will pull on technology in different ways to make their case. This chapter looks at the "democratic innovation" space, and how it selectively rejects and accepts tech approaches. This includes people and groups interested in participatory democracy (where decisions are taken by individuals and groups more directly), which includes the wave of interest in participatory budgeting after its initial use in Porto Alegre. There is also a newer grouping interested in deliberative democracy, where the focus is on the quality of discussion and decision-making, rather than mass participation.

Deliberative approaches to democracy

Citizens' assemblies and deliberative democratic approaches are (similar to participatory budgeting), an example of a fast policy: spreading rapidly around the world in different contexts and adapting to different local conditions and challenges. There are also interesting experiments with transnational citizens assemblies as democratic devices that can span national borders.

A key belief of deliberative democracy is that people are directly collectively competent to make decisions that affect their lives, but lack the knowledge and structures to do so. Deliberative practitioners are concerned with creating those decision making structures. As part of this, there is opposition to elitist forms of democracy, with a weak form that sees deliberative approaches as a useful or necessary corrective element to electoral democracy, and a strong form that seeks to replace electoral democracy with new forms altogether.

It draws on alternative democratic ideas of lotteries and turn-taking rather than elections and civic intermediaries. Filtering decisions through small "mini-publics" changes the nature of a number of democratic challenges. Ignorance of a policy area is an addressable problem for 100 people, but not for 100 million. The approach is to shrink the problem so you can fit it in a room, get the right answers, and then scale up the results.

There are a number of critiques that can be made of this approach, but what we're interested in here is how it interacts with technology. This represents an excellent example of thinking through how to align technological approaches with a specific theory of democracy, and in turn how the availability of new tools can make that theory of democracy more practically achievable.

The gulf between civic tech and deliberative democracy

In a 2024 TICTeC talk, Matt Ryan talked about the challenges people working in democratic innovation perceived in working with civic tech, both philosophical (equality vs scale) and practical (tech-solutionism versus messy human discussion). Part of this was that the different approaches of these groups reflected different ideas about what was valuable about democracy.

A key use of technology is that it lowers transaction and information costs, and helps processes scale up to larger numbers more cheaply. Civic tech practitioners, who see the value of this in improving democracy, will generally believe in versions of democracy where scale is important. For instance, involving more people in processes, or helping more people have information, is an important part of the democracy they want to live in.

Deliberative practitioners do not value scale in the same way, and often see the trade-offs that scale requires as opposed to the kind of democracy they want to build. Reduced publication costs have created social media: is this unquestionably good for our democracies? Lowering costs to entry helps more people participate, but it does not change that some have a lot more resources, and ability to participate, or improve the quality of debate in the public square.

Instead, deliberative practitioners will focus on equality and quality of discussion at the expense of scale, creating expensive processes to find and convene a representative cross-section of society, and not just lower the barrier for anyone to enter. Where civic tech approaches might be useful was in the neglected "scale up the result" point of the process, helping conclusions and arguments reach the biggest possible audience.

On the practical side there are questions about human versus technological infrastructure: do we need to be building deep pools of convening and facilitation knowledge inside and outside of public institutions, or do we need to be procuring a piece of software? Your answer will depend a lot on how much you see the challenges of democracy as an information problem to be fixed with better distribution, or as a human one that is fixed by people talking to each other.

Ryan found two groups among his interviewees. One is constrained pessimists, who believe that tech in participation tends to fail or harm by concentrating power among already-advantaged groups. The other is critical optimists, who believe tech can help, but only with careful design, mixed online and offline models, and serious attention to inclusion, user research, and institutional follow-through

This latter group provides the ground for a synthesis view, where the role technology plays in the process needs to reflect and amplify the kind of democracy this group is arguing for.

Democracy as difference - the Polis approach

Polis is a digital framework and service that is well aligned with the beliefs of deliberative democrats, and approaches scale in a way that is more compatible with their concerns about inclusion. It has been popularised in part via its use by the vTaiwan process, which is of interest to both the civic tech and democratic innovation communities.

Polis is essentially polling software, but it is uninterested in how many people choose each side; rather in identifying clusters in responses to statements (to understand the different opinion grounds at play). It encourages users to submit new and variations on statements, and highlights statements that are shared between groups.

This is a use of technology that matches the beliefs of deliberative democrats: where deliberation and discussion is not a clash of different interest groups, but a sharing of difference in the hope of synthesis and reconciliation.

Seeing difference between groups as a core element is resistant (if not completely immune) to problems of unrepresentative users. The key signal that is being pulled out is not the size of clusters, but that they exist at all — and what is of interest are the potential areas of overlap and consensus between divergent views. This is a philosophy based on the premise that finding consensus (rather than one side ‘winning') is the goal of the process, and it is supported by a range of technological tools and approaches.

Use of LLMs in deliberation

New AI technologies help both to build on existing machine learning approaches, and potentially to lower some of the costs of more in-depth deliberative approaches (either by supporting existing processes more cheaply, or by getting more value out of in-person processes).

Sammy McKinney identified a range of potential uses for AI in deliberative processes, supporting recruitment and planning, learning, deliberation/decision making, and follow up. This section focuses on some live examples.

A key use of LLMs is to help better understand clusters by assigning human-understandable labels. Polis 2.0, launched at TICTeC 2025, uses this approach, where the fundamental clusters remain based on an understandable methodology, and the LLM-generated cluster labels link back to the source material they are drawing their descriptions from. Zkorum's Agora Citizen Network similarly uses LLMs for labelling on top of a version of the Polis approach, creating more easily parsable meaning from the cluster information.

A more speculative application (but under a range of active investigations) is using LLMs to move deeper into the deliberation process. There is the potential for cheap transcription and clustering to pull out threads and themes from unstructured discussions. At TICTeC 2025, Demos talked about their programme Waves, exploring a hybrid participation approach. The goal is a broad and lightweight stage with thousands of online participants (on a Remesh platform), combined with deep deliberation stages with a narrower range of participants (also through an online platform Psi).

Both of these stages will be assisted by ML/LLM tools that help surface and understand the conversations happening: in the first stage, helping to cluster and theme contributions and detect consensus and division; in the second, helping to manage conversations and mapping which arguments people found effective. Over the top of this, they are exploring a "glue" layer that moves results between different processes, and creates an LLM-assisted question answering tool that ingests all the information consumed and produced from the project, to help participants and local authority staff ask questions about the process involved.

These uses of AI technologies are experimental and need good evaluation approaches to ensure that they deliver enough value to be useful to the overall process. Approaches to transcription can have variable quality (especially in different languages) — does even inaccurate transcription help add value to processes if it helps draw out and make explorable otherwise lost deliberation? As methods develop, having clear criteria for useful technology helps keep integrations of different approaches practically grounded.

Polis the organisation vs Polis the idea

At a TICTeC seminar, Sitra's Sanna-Kaisa Saloranta talked about their approach to building deliberative infrastructure in Finland. Here, while the idea of Polis is key, it has needed important localisation. Integration has been about building redundancy (supporting an ecosystem rather than a single provider), making it usable by officials (hosted in-country in a way that is GDPR compliant), and adapting to the local environment: mobile first UI, and renaming the software to Voxit - Polis being too similar to the Finnish word for 'police' (poliisi).

In a similar approach, Zkorum's Agora Citizen Network re-implements the core "Polis" clustering approach (a python library called Red Dwarf) to allow them to build new features and ideas around it.

What we see here are a few different examples where the Polis approach is increasingly separated from the original organisation, reflecting the continued usefulness of open source in helping ideas, and the software that embodies them, adapt to more contexts. The future of this approach is one where the democracy-as-difference approach is implemented in many different forms, and is adapted into a range of tools rather than a single provider.

Intersection with wider conflicts

Recognition that some forms of technology are helpful to certain kinds of democratic projects and not others means that those working at the intersection of democracy and technology need to carefully think about what they mean by democracy, and which ideologies the technology they create might be appealing to.

For instance, part of the appeal of Twitter's Birdwatch/X's Community Notes (heavily based on the Polis methodology) was the idea that moderation could be shifted to the community, getting rid (on ideological or cost grounds) of other forms of content moderation (while the initial design anticipated different approaches working together). Isolated, it is a success story about how technology can practically bridge divides between groups. In context, hate speech on X went up over the first period Community Notes was live. It's an obvious point, but innovative approaches are not enough if positive goals are not widely enough shared to support them correctly.

Conclusion

Successful adoption of technology by people working in democratic innovation relies on an understanding of, and work within, a shared theory of democracy. Having a clear sense of what is valued and what is rejected helps in applying technology to new democratic approaches; understanding how to bridge gaps between different interested groups; and practically make new forms of democracy easier to reach for as a tool.

Access to Information as democratic infrastructure

Access to Information (ATI) is a mechanism through which individual citizens can request information held by public bodies. The many organisations that make up our ATI Community of Practice show that across the globe, ATI is more than just a legal right — it's a form of vital democratic infrastructure. ATI is a set of practices and norms that make meaningful civic scrutiny achievable, and we believe that technology can help make that scrutiny faster, deeper and available to more people.

Alaveteli is mySociety's open source codebase that is used in over 30 countries around the world. Our code makes ATI laws easier for requesters and authorities to engage with in practice: publishing requests and responses openly, standardising workflows, creating reusable public records, and generating data about how transparency systems function. By turning individual requests into shared public resources, Alaveteli lowers the barrier to entry for ATI at a citizen, community and country level.

Our recent series of webinars brought together associates from around the world, including those who run Alaveteli-powered ATI sites as well as ATI sites that use different codebases; journalists; academics and campaigners. Across each event, speakers consistently framed ATI not as a narrow transparency mechanism, but as a foundational condition for democratic oversight. This shift reflects broader changes in how we are governed; public decision-making is now distributed across digital systems, outsourced suppliers, automated processes, and data infrastructures that are rarely visible to the public by default. In this context, ATI is an essential tool for civil society actors seeking to ‘lift the lid' on government decision-making, helping us understand how big and small decisions are made, who designs systems, and where responsibility lies.

This infrastructural role of ATI is apparent globally, though perhaps particularly visible in the parts of the global south where formal transparency institutions may be weak, but civic actors have developed practical, adaptive approaches to accessing and reconstructing public information. In states with all kinds of democratic backdrops, effective civil actors treat ATI as part of a wider data ecosystem; combining legal requests, crowdsourced data, investigative methods, and digital tools to create accountability where it would not otherwise exist.

Seen in this light, ATI is an essential tool for democratic scrutiny in an increasingly automated and fragmented state.

ATI and algorithmic power: making automated decisions legible

One of the most pressing challenges for ATI today is the rise of algorithmic decision-making in the public sector. Across the world, automated systems are now used to assess welfare eligibility, flag potential fraud, control traffic, support policingand manage borders. These systems fundamentally alter how power operates, and often do so without corresponding transparency.

Taking a global perspective, there's a need to recognise that the adoption of digital and automated systems is unevenly distributed, between and within countries. While some public institutions are deploying advanced machine learning systems, others are still struggling with the basics of digitisation, data management, and interoperability. In many places practical challenges remain for public management offices attempting to coordinate even basic digital records across institutions.

In this context, algorithmic decision-making does not replace existing governance problems so much as layer on top of them, often amplifying opacity and accountability gaps rather than resolving them.

The politics of opacity

At our event on AI decision making, all three of our speakers raised concerns that transparency around AI is actively deteriorating. Jake Hurfurt of Big Brother Watch spoke of the decreasing transparency of our public bodies when it comes to AI-based systems, and the root causes of it: a lack of technical expertise among smaller authorities and the contracting of technology from private suppliers.

Public authorities increasingly rely on private vendors whose systems are treated as proprietary, using this justification to refuse requests. FOI officers may lack the technical expertise needed to interpret requests about algorithms, while institutions themselves may we wrote down a series of ques not fully understand the systems they deploy.

This creates a dangerous accountability gap. When neither the public nor the state can explain how decisions are made, democratic control is undermined. Our speakers argued that ATI must be explicitly recognised as part of AI governance frameworks, rather than treated as an afterthought.

From documents to systems

Traditional research or investigation frameworks may assume that relevant information exists in discrete, retrievable documents. Algorithmic governance disrupts this assumption. Key information may be spread across procurement contracts, vendor documentation, internal emails, risk assessments, training datasets, and codebases, each of which may sit inside, outside or in between private and public bodies. As a result they may be incomplete, contested, or protected on grounds of commercial confidentiality. Last year mySociety Chief Executive Louise Crow wrote about the need to extend FOI legislation in the UK to cover government-procured AI that is used in decision making.

Academic analyses of the use of automated decision-making in welfare and fraud detection have utilised ATI as a process of reconstruction rather than simple retrieval. Understanding how an algorithm functions often requires submitting dozens of requests over time and triangulating partial disclosures. No single response provides clarity; insight emerges only through cumulative analysis.

ATI as a tool for exposing harm

By forcing disclosure of risk indicators, decision rules, and internal evaluations, ATI users have been able to show how automated systems disproportionately target already marginalised groups. Alaveteli sites around the world mean that you don't have to be a trained journalist or academic to bring these issues to light.

Dr Morgan Currie from the University of Edinburgh explained that much of her research into the Department of Work and Pensions (DWP)'s use of automation for fraud detection was based on analysing FOI requests made by others on WhatDoTheyKnow, which are public for everyone to see.

The Suspicion Machines investigation by Lighthouse Reports brought together ATI requests across multiple European countries and ultimately uncovered how automated welfare fraud detection systems relied on risk-scoring models that disproportionately targeted already marginalised groups, including women, migrants, and low-income households. Crucially, these impacts were not apparent from official descriptions of the systems themselves, but only emerged once disclosed data was analysed comparatively and at scale.

In our webinar, Gabriel Geiger from the Suspicion Machines investigative team highlighted his four-step process for building up a comprehensive picture using ATI:

1. Find out the basics from the press office first, such as: what type of AI software is being used? What are the names of the software programmes? What type of data do they use? Is there a private vendor involved?
2. Request standard documents: data protection impact assessment, human rights impact assessment, database codebook
3. Request more detailed information: caseworker manual, full set of variables, performance evaluation of the model
4. Finally, request the most sensitive documents: the source code, training data, and machine learning model

Through this layered approach, ATI enables external actors to identify patterns that institutions themselves may not recognise, or may prefer not to acknowledge. ATI is complementary to other scrutiny mechanisms such as internal audits and regulatory oversight, offering a citizen-powered early-warning mechanism that grows more and more necessary and powerful as governments are procuring more and more AI-powered decision-making tools.

Measuring ATI: from legal compliance to lived experience

A key question that has come up regularly through the convening of our global community of practice is the question: what is a good ATI system? How can civil society in any country effectively advocate for ATI improvements? What matters more, the law, or the implementation of the law?

Strong ATI laws do not always guarantee meaningful access. Countries with robust statutory frameworks may still suffer from administrative resistance, under-resourced institutions, or political hostility toward transparency. Conversely, some weaker legal environments exhibit high levels of practical innovation driven by civil society, resulting in powerful ATI-driven investigations.

We invited a range of speakers to tell us more about how we can measure ATI in practice, and how civil society actors can use both legal and practical analysis of their ATI systems as part of their advocacy work.

Green arrow up = good, red arrow down = bad

The RTI Rating of global ATI legislation plays an important role in driving political attention and reform by making Access to Information legislation visible and comparable across the globe. As the founder of the rating tool, Toby Mendel from the Centre for Law and Democracy, explained in a TICTeC webinar, even when governments dispute the technical details of their score, "they still pay attention when they see that their neighbour is doing better than them."

By introducing a clear, comparative benchmark, the RTI Rating creates a form of reputational pressure that encourages engagement with ATI standards, sparks public debate, and provides civil society with a credible tool for advocacy. However, what happens when a country is good on paper but bad in reality?

How long is a piece of string?

The Centre for Law and Democracy is now working on new projects that measure ATI in practice, but this isn't easy for a number of reasons. Firstly, Mendel argued that real-world access is shaped less by formal rules and more by administrative culture, capacity, and discretion. Institutions operating under the same law can behave very differently, and many of the most serious barriers to access (informal discouragement, selective responsiveness, unanswered requests, or pressure on requesters) are effectively invisible in official data. These factors make it difficult to capture an accurate picture of how ATI systems actually function.

He also highlighted the challenges of scale, consistency, and definition. Meaningful measurement requires large volumes of comparable requests over time, which is resource-intensive and politically sensitive.

At the same time, compliance does not necessarily equal success: Maria Žuffová from the European University Institute argued that a response can meet legal requirements while still failing to provide usable information. Access fails in practice when information is released in formats that are incomplete, poorly structured, or impossible to reuse, even if deadlines are met. Žuffová highlighted that in data-driven governance, meaningful ATI depends on machine-readability, completeness, and accessibility, and warned that narrow metrics risk masking systemic failures by treating minimal compliance as success.

Giovanni Esposito from the Université Libre de Bruxelles focused on how ATI performance varies widely within countries, even under the same legal framework. He emphasised that transparency is shaped by organisational culture and administrative practice rather than law alone. Some institutions develop habits of openness, while others systematically delay or resist disclosure, and these patterns persist over time. Esposito argued that measuring ATI in practice therefore requires institution-level analysis and longitudinal data, rather than national averages, if it is to reveal where transparency is genuinely working and where it is breaking down.

What can we measure?

Arguably the single most important factor influencing how ATI works in practice is the strength and independence of the local Information Commissioner or oversight body. Where commissioners are genuinely independent, adequately resourced, and willing to enforce the law, public bodies take Access To Information obligations seriously because there are real consequences for non-compliance.

Conversely, even strong ATI laws tend to be ignored when oversight bodies are weak or politically constrained. Effective Information Commissioners shape behaviour across the system, not only through appeals and enforcement, but by setting norms that frame Access to Information as a democratic obligation rather than an administrative inconvenience. In our recent evidence to the Scottish Parliament we highlighted this point: even though Scotland's system is already comparatively strong, a more empowered Commissioner can only be a good thing for the healthy functioning of the law.

Civic tech's role in providing access and analysis

As governance becomes more data-driven, ATI measurement must evolve. A response that arrives on time but is unreadable, unstructured, or incomplete may technically comply with the law while failing any meaningful standard of access.

Civic tech platforms that publish ATI requests play a crucial role here. By generating structured data about response times, refusal rates, and disclosure practices, we can help to identify patterns of compliance and resistance that would otherwise remain anecdotal.

Crowdsourcing and community-led ATI: rebalancing power

Crowdsourcing and community-led ATI initiatives illustrate how Access to Information can be re-imagined as a collective democratic practice rather than an individual legal transaction.

Producing information the state does not

Communities often seek information that governments do not systematically collect or disclose. Crowdsourcing allows citizens to generate datasets about service delivery failures, environmental harms, or procurement practices that are otherwise invisible. Nnenna Eze from PPDC in Nigeria told us how the organisation supports residents to make requests about how and when services failed, how frequently issues occurred, and which areas were most affected, creating a body of evidence that can be used to engage authorities.

This approach shifted accountability conversations away from anecdotal complaints toward documented patterns, demonstrating how community-led ATI can surface critical public-interest information that the state does not systematically collect.

Trust, legitimacy, and participation

Community-led ATI also changes the relationship between citizens and information. When people participate in collecting and validating data, they are more likely to trust it and use it for collective action. This participatory dimension is often absent from formal and individual ATI processes, which can feel opaque or adversarial.

mySociety supports Climate Emergency UK (CE UK) to build the Council Climate Action Scorecards. Rather than relying solely on national datasets, CE UK mobilises volunteers to collect, verify, and update information about local climate action plans, council commitments, and delivery against stated targets through a combination of national data, website analysis and FOI requests. This distributed model allows CE UK to build a comprehensive picture of climate governance across hundreds of local authorities, while also engaging citizens directly in accountability work.

By combining crowdsourced data collection with clear frameworks and public-facing analysis, CE UK demonstrates how community participation can strengthen access to information and enable scrutiny of complex, decentralised policy areas such as climate action. This process has inspired organisations in other countries including Switzerland and Canada to create similar initiatives.

Verification and collective intelligence

A common concern about crowdsourced data is reliability. Speakers at ATI Network events addressed this by highlighting verification practices rooted in collective intelligence: peer review, cross-checking with official sources, and local knowledge. Marzena Błaszczyk from Citizens Watchdog Network Poland described how local knowledge plays a crucial role in validating information during politically sensitive moments, such as elections, where communities are often best placed to notice gaps or anomalies in official processes. These methods challenge assumptions that institutional data is inherently more reliable than citizen-generated information.

Complementarity, not replacement

Community-led ATI can complement formal accountability mechanisms by filling gaps, setting agendas, and sometimes prompting new forms of formal disclosure. In several cases discussed, community datasets were later used by journalists, regulators, or public officials themselves.

Since January 2019, German ATI organisation FragDenStaat's TopfSecret project has encouraged 34,000+ users to make 71,000+ requests for food hygiene inspection reports. As well as publishing them, as standard, on their ATI portal, they also made an interactive map, allowing citizens to see where serious hygiene violations had occurred. Stefan Wehrmeyer, FragDenStaat founder, explained that publishing this information collectively (rather than delivering it privately to individual requesters) shifted incentives for both businesses and regulators, increasing public pressure for compliance and improving transparency around food safety. The success of this project continues: - the platform is still running and some authorities are now proactively publishing their region's food hygiene data.

Scaling through technology

Emerging uses of AI and automation (such as classifying submissions or supporting multilingual participation) may offer opportunities to scale community-led ATI without losing its participatory character. However, our speakers emphasised the need for caution to ensure that technology supports, rather than overrides, community agency.

ATI, journalism, and the politics of public impact

Access to Information arguably has its greatest democratic impact when it is actively used to inform journalism that reaches the public. In her webinar, Maria Vitória Ramos of Fiquem Sabendo focused less on ATI as a legal mechanism and more on ATI as a practical journalistic method; one that requires skills, support, coordination, and cultural change if it is to produce meaningful accountability.

Maria's contribution emphasised that simply having a right to information is not enough. The democratic value of ATI depends on whether journalists know how to formulate effective requests, recognise newsworthy disclosures, and reuse information strategically.

What is Fiquem Sabendo?

Fiquem Sabendo is a Brazilian civil society organisation that works to make Access to Information usable and impactful for journalists. Rather than positioning ATI as a specialist legal tool, the organisation focuses on integrating it into everyday journalistic practice.

Maria described how Fiquem Sabendo supports journalists by helping them draft requests, understand ATI law, interpret responses, and identify stories hidden in administrative data. Much of this work is educational and relational: building confidence, demystifying bureaucratic processes, and encouraging journalists to see ATI as a normal part of reporting rather than an exceptional tactic.

The virtuous circle of FOI news stories

There's nothing like a captivating news story that uses FOI to promote the use of FOI to other journalists, and Maria gave us this classic example from Brazil:

One journalist made two really simple FOIA requests. He asked the health authority how many denials of transportation for transplanting organs they received, and then he also asked the Air Force for all the flight logs for politicians. By matching these two data sets, he realised that in three years, over 150 lives were lost due to an absolute lack of priority in the use of public resources for the organs. And the day after he published his news piece, the government created a regulation obliging the Air Force to always have a plane ready to go for transplanting organs. In one year, almost 300 lives were saved.

ATI as a journalistic practice, not a legal formality

A central theme of the session was the gap between the formal availability of ATI and its practical use in newsrooms. In many contexts, ATI rights may exist on paper but be unevenly used, particularly by journalists working under time pressure, in the context of legal uncertainty, or limited institutional support. As a result, large volumes of potentially valuable information remain unexamined.

The work of organisations such as Fiquem Sabendo illustrates how this gap can be addressed by treating ATI as an everyday reporting tool rather than a specialist legal exercise. Starting with the basics is worth it. Helping journalists to draft effective requests, understand disclosure obligations, and recognise the news value of administrative data can significantly increase the real-world impact of ATI frameworks. This approach shifts the focus from compliance to use, and from access in principle to access in practice.

Practical strategies for making ATI useful to journalists

A key insight from the webinar was that successful ATI-driven journalism depends as much on timing and distribution as on disclosure itself. One practical strategy discussed was the importance of maintaining a calendar of events that journalists already care about, such as elections, budget announcements, major policy milestones, or anniversaries of significant decisions. Planning ATI requests around these moments increases the likelihood that disclosed information will be relevant, timely, and publishable, rather than arriving too late to shape coverage.

Another lesson concerned the value of building and maintaining direct relationships with journalists through simple but consistent communication channels. Developing a mailing list of reporters interested in ATI allows organisations to share disclosures, explain why particular information matters, and flag emerging patterns across multiple requests. Rather than expecting journalists to search for information themselves, this approach brings material directly into their workflows, lowering the barrier to use.

Together, these practices highlight that ATI impact is not automatic. Information needs to be anticipated, curated, and actively circulated if it is to translate into public-facing journalism. For pro-democracy tech actors, investing in these seemingly low-tech coordination strategies can be just as important as building platforms or tools.

Implications for pro-democracy tech

Taken together, these themes suggest that pro-democracy tech can play a crucial role by supporting training, coordination, reuse, and sustainability, particularly in contexts where journalism is under-resourced but essential to democratic life.

ATI, in this framing, functions as a shared investigative infrastructure. Its democratic value depends on how effectively it is integrated into the everyday work of those who turn information into public understanding.

Conclusion

Throughout this chapter, ATI emerges not as a narrow legal right or technical mechanism, but as a core component of democratic infrastructure. ATI enables scrutiny in systems that are increasingly complex, automated, and fragmented. It allows civil society, journalists, and communities to understand how decisions are made, where responsibility lies, and whose interests are being prioritised, even when power is exercised through data, algorithms, and outsourced technologies rather than visible political processes. To play our part in celebrating and championing the use of Access to Information as democratic infrastructure, we've built a new destination for ATI success stories.

The case studies and debates explored in this report and highlighted on this new showcase page demonstrate that ATI is increasingly collective rather than individual. ATI platforms allow civil society to surface patterns that no single requester could identify alone. These approaches rebalance power by turning Access to Information into a participatory democratic practice, grounded in local knowledge and collective verification. Far from undermining reliability, such methods often produce more robust and trustworthy data than institutional sources alone.

Journalism plays a central role in this ecosystem. ATI reaches its widest public impact when information is transformed into stories that people can understand and act on. Relatively simple requests, strategically timed and combined, can expose systemic failures and lead directly to policy change. But this impact is not automatic. It depends on skills, coordination, reuse, and sustained support. These are all areas where pro-democracy tech can make a decisive difference.

Where ATI works well, it is supported by institutional enforcement, civic capacity, cultural norms of openness, and practical tools that lower barriers to use. Where it fails, the problem is rarely the absence of a legal framework alone, but weak implementation, limited resources, or deliberate opacity. Pro-democracy tech has a role to play across all of these factors, highlighting good and bad practice, demonstrating how things can be different.

Looking ahead, the challenges facing ATI are likely to intensify. Governments are deploying more automated decision-making systems, relying more heavily on private vendors, and generating ever-larger volumes of data. At the same time, political pressure on journalists and civil society actors is increasing in many parts of the world. In this context, ATI must be recognised as a non-negotiable component of democratic governance in the digital age, not an optional add-on. ATI can help us tackle today and tomorrow's problems, if we let it.

For practitioners, this means continuing to invest in practical capacity, collaboration, and experimentation. For funders, it means supporting ATI not just as a legal reform agenda, but as a living infrastructure — one that connects technology, law, journalism, and community power. And for the pro-democracy tech field as a whole, it means treating Access to Information as foundational: a mechanism that makes other forms of democratic innovation possible.

Supporting global democratic transparency

Supporting global democratic transparency is less about exporting a single parliamentary monitoring model and more about strengthening shared capacity and imagination. There are many routes to impact (through citizens, intermediaries, and changes to institutions themselves), and different projects will have different approaches. What we want to do is broaden the set of ideas and projects that parliamentary monitoring projects can draw from, and make it easier to adopt new technical approaches in their work.

This is a critical time for transparency and public participation. Recent years have seen many new and exciting approaches to Parliamentary Monitoring Organisations (PMOs), against a context of growing challenges to global democracy that is shifting the nature of the work. Democratic backsliding in the modern era rarely means the abolition of Parliaments, but that they become less important arenas of power, and are instrumentalised into wider, less democratic, regimes. PMOs in these contexts might both be drawn into wider anti-corruption work, and, with a more future-orientated perspective, focus on the restoration of functional democratic institutions.

A key part of mySociety's work internationally has been supporting the development of parliamentary monitoring. We're proud of running TheyWorkForYou, one of the oldest online parliamentary monitoring projects, and proud that this twenty year old service continues to be a source of ideas and inspirations for new projects globally.

We've seen parliamentary monitoring develop and mature, including the development of a wide range of organisations around the world, and regional networks, such as the Red Latinoamericana por la Transparencia Legislativa (RLTL) and African Parliamentary Monitoring Organizations Network (APMON).

Over the last few years of TICTeC, one of the questions we've wanted to explore is how best to support PMOs around the world without duplicating work happening elsewhere. As part of that we've been conveying a community of practice around parliamentary monitoring. Our conclusion has been to go both broader and narrower.

Broader, in trying to be more inclusive of projects that are fundamentally similar to PMOs, but not based around Parliaments. We've created a new umbrella term of Democratic Transparency Projects (DTPs) to capture this set.

But we also found a niche in being narrower, to avoid overlapping with regional networks while making sense of a global scale. Here we wanted to focus on technical innovation and techniques that might be applied, or learned from, between jurisdictions. We see an especial value in examining how cutting edge approaches (such as AI technologies) present new angles of attack on old problems, but bring up new questions on organisations' approach to risk, rather than purely technical matters.

Democratic transparency

We were keen not to narrow down the explorations we made in our TICTeC Community of Practice around democratic transparency. We identified interesting projects that felt philosophically similar to PMOs (making use of similar tools, and with similar theories of change), but that were not focused on legislative bodies — were there lessons to be applied from them?

Arguably the concept of PMOs (as technical projects) has over-focused on parliaments, because this approach aligns well with existing ideas of democratisation. PMOs came to prominence as the intersection of new tech-literate civic society actors ("civic tech") with older post-Cold War ideas on what deepening and consolidation of democracy progress looks like. This had a strong focus on parliaments themselves, but with ideas of openness to the public and civil society. The height of this era of a global idea of PMOs can be seen in the 2012 Declaration on Parliamentary Openness, which positioned PMOs as a key part of the global drive towards greater transparency, seeking both greater access to parliamentary data, but also to help drive dialogues on internal parliamentary reform.

Today, ideas of democratic development have expanded to include deliberative democracy and direct democratic approaches such as participatory budgeting. This reflects both broadening perspectives, but also the exploring of additional avenues in light of stalled progress or democratic backsliding in focusing on parliaments.

Expanding from PMOs, we can see a wider group of Democratic Transparency Projects (DTPs). The general definitions here is:

Democratic Transparency Projects are projects that [rework / create] public information about [democratic institutions / politicians] to improve [transparency / accountability / standards / efficiency].

The important assumptions of this definition are:

• DTPs will generally build on existing data from one or many institutions, but can also create new analysis or data.
• Their focus is democratic institutions – generally elected representatives, on parliaments or governments, but could include scrutiny/transparency of citizens assemblies.
• Their role is not passive: there is a goal to change the democratic system they are a part of. While the theory of change may differ, the goal is not just to record, but make things different. This might be through active campaigning, or that creating public visibility in itself creates pressure for change.

Routes to impact

There are three groups of users DTPs/PMOs leverage to effect change on the democratic system:

• Citizens: they create better feedback loops between citizens and representatives; better alignment between principal (citizen) and agent (representatives) - help align elite behaviour with citizen needs.
• Civil society: they give new tools to infomediaries (journalists, academic, CSOs) so they can understand, share information, and take action.
• Institutional: they create internal efficiencies for the representative institution (parliaments, etc) by making their own information more accessible/inspiring improvements.

These are mutually reinforcing, in that use of their outputs by a variety of actors helps stabilise support and minimise backlash to their work.

Different types of projects choose to focus on different priorities, but a recent NDI/Directorio Legislativo report on parliamentary monitoring identified (at least) seven possible areas of work :

• Legislator directory and activity monitoring
• Public engagement with MPs
• Legislative tracking
• Discourse quality monitoring
• Budget and public finance tracking
• MP asset and conflict of interest monitoring
• Transcription and interpretation

Expanding beyond PMOs, we can see how these functions can be adapted to new contexts; but also projects that add new ideas beyond these to the toolkit. Rather than parliaments, these kinds of projects might focus on non-legislative bodies (for instance, projects monitoring the executive) or where a project is covering a mesh of related institutions.

Going wide – Querido Diário

Querido Diário is a project that aims to bring together and make searchable the government gazettes for every city in Brazil (of which there are over 5,000). The goal is to create a national level database of decisions made in every municipality.

This evolved out of Open Knowledge Brazil(OKB)'s national level projects looking at public spending data. There often isn't great data at the national/federal level, but this gets worse the more local you go. However, if information is technically available but horribly fragmented, this is something that civic tech can try and address from the outside. As there isn't an API of government decisions available, OKB are building it themselves.

When you're trying to build a project covering hundreds or thousands of different institutions, you have to do more work further down the value chain just bringing the data together before you can analyse it. For instance, in the UK, we've ended up being the holders of the best list of local authorities, because we needed that to power our local climate policy analysis. The uses of that base layer are a bit abstract, but it is the foundation that is required for highly impactful services.

In the case of Querido Diario, Diários do Clima also builds on top of this base layer to create a service specifically looking at new environmental and climate regulations in all municipalities covered. Having this information in one place makes cross-city comparisons of climate action possible. As well as making gazettes easier to search for local journalists or civil society, this dataset enables subject area journalists and researchers to do new cross cutting analysis.

Going wide means investing lower down the value chain so others can build accountability work on top. Lowering the cost of accessing all local information helps people and organisations with subject matter expertise do work that would otherwise be unviable or incomplete. The scope of Querido Diario shows the challenges of scale, but also the big rewards of joining up the data.

Going deep – Datos que hacen Ciudad

An alternative to covering lots of municipalities is to build a richer service catering to one.

Datos que hacen Ciudad's goal is to create better information about Santiago de Cali, a city in Colombia. The project includes familiar PMO approaches of displaying information about representatives, but is also consciously aimed towards getting better information to those representatives. The theory of change here is "If we give our leaders more information, it will lead to better decisions." Through councillors sharing information on their areas of focus/problem areas, Datos que hacen Ciudad can both provide that information to citizens, and shape the information to be sourced and created.

One of the complexities of local governance is that it is best understood as a patchwork of different institutions. This project makes that complexity an advantage — it's a partnership of different local institutions, pulling on resources and knowledge from different places, with a key anchor in the university.

Going deep helps create tighter local feedback loops and decision-maker-facing information as a route to institutional change. The feedback loop between decision makers and different groups helps create highly localised information. Data collection and analysis catering to the exact needs of decision makers can be more sensitive to local patterns than generalised national data.

New variations

Another reason to take an expansive view of democratic transparency is that it helps bring into scope projects by organisations that are clearly adjacent to PMO work.

At TICTeC 2025, we saw presentations from WeVis and Mzalendo: both organisations with Parliamentary monitoring platforms, talking about engaging citizens in constitutional change, and governmental promise tracking respectively. Based on connections made at TICTeC, WeVis is currently replicating a Directorio Legislativo project around politically exposed persons (PEP) data — that builds on their PMO expertise but takes it in a new direction. This reflects that organisations who create PMOs often have interests in democratic and governmental processes that go beyond parliaments, and taking the wider view in how we organise convenings helps highlight approaches to transparency that could be more widely shared and emulated.

Supporting technical innovation

One of the challenges for international democratic transparency is how to best support the transfer of technical knowledge and innovation between places.

Our current approach is less to support the transfer of whole frameworks (we do this more in Access To Information through the Alavateli software), but in helping to demonstrate practical uses of technology that might inspire new users or approaches in different contexts.¹ Lowering the technical barriers to entry does not always mean projects are automatically viable, as technical issues are not the only significant limit on PMOs' success, with wider organisational sustainability linked to funding and variations in how supportive or hostile the surrounding political environments are.

This TICTeC-based innovation approach grows out of hard-won lessons about how to assist in the technical development of PMOs. The initial idea was that parliamentary monitoring software itself was adaptable and transferable (and so the costs of implementing could be reduced in each case). This was challenged by the problem of adaptations in each case, even when superficially countries shared a "Westminster" parliamentary system. Parliamentary procedure, MP responsibilities, attendance requirements — all meant that significant adaptation was needed over a core approach.

From mySociety's perspective, ideas based first on shared software (generalising TheyWorkForYou), then common components developed among a consortium (Poplus), then common data/scrapers to support local developers (EveryPolitican) all ran into the problem that specific work was required in each context to adapt general tools. In some places, projects can build on top of already digitised resources; others require a lot more initial setup work. For instance, Odekro in Ghana had to start with a scanner and OCR software, before they could even consider putting parliamentary documents online. While there may be useful tools to share and lessons to learn, the significant costs of the project are in the adaptation of general concepts to the specifics of the local context.

While the fruits of projects to develop common technical components and data infrastructure do continue to be used in a range of approaches around the world, through TICTeC we are seeing successful adoption of approaches between areas, where one project recognises an approach of another that can be adapted to their goals, even if no actual code is shared between the two.

Application of AI technologies to democratic transparency

Use of LLMs to build adaptive scrapers

Data that is fragmented and not in useful formats is a fairly common civic tech problem. The solution is to write a ‘scraper' that reads the content of a website and has a process for converting it to a more structured format.

This works well when dealing with only a few sources (e.g. the memberships of the UK's four parliaments only needs a few different scrapers), or where a common format is being used.

A problem mySociety ran into when looking at memberships of cross party groups was that information was distributed across many different individual websites in individual formats.

Rather than a traditional scraper, we built an agentic AI/LLM approach that is more flexibly able to extract memberships from websites. This pairs the use of LLMs to navigate a website and extract information, with automated checks that the extracted information was actually present on the original page. Practically, this was an effective way of getting the information we needed that turned a very hard problem into one that we can dependably run regularly.

Aligning transcripts and video

In 2008 mySociety worked with the BBC to explore how to make parliamentary video searchable. At the time, the best approach available was a crowdsourcing approach to reconcile timestamps and speeches: around 400 volunteers aligned 160k speeches. Ultimately this process had big technical overhead in video storage, and a lot of manual work was required to keep the two feeds together.

New approaches make this possible at scale with much less manual effort. As part of our TICTeC Community of Practice around Parliamentary Monitoring we ran a session on video and transcripts, hearing from OpenParliament TV about their approach.

The organisation currently covers the German Bundestag, but is interested in expanding their approach to more countries. The platform is a combined video/transcript search application, where a user can search for individual speeches, and view video at that point in the record.

There's a lot of moving parts that make this work, but the core work is in how the video and the transcript are aligned. The transcript is converted to computer-generated audio, and then the generated audio is matched against the real audio. The matching uses an adapted version of the open source aeneas framework for ‘forced alignment' of text and speech.

The reason for the approach is that speech-to-text is often specifically bad at generating the punctuation of written language, while one of the key things in syncing transcripts to video is finding the start and end of blocks. This can still run into difficulties when what is being said is just not present in the transcript, but generally it can resolve problems by matching items on either side. From an international perspective, this is also interesting in that it's an approach that works better across different languages than speech-to-text approaches.

Open Parliament TV does not host the video themselves. Offline they need to process it to extract and match audio, but online their player links to the videos as hosted by the Parliament. In the long term, they want a workflow that sends videos to an internet archive as a backup. This is another useful function of democratic transparency projects: to act as civic redundancy against backsliding access to democratic materials. Another compelling example of this was when Abrimos.info's copy of an official procurement contract database in Guatemala was accessed hundreds of thousands of times during a change of government when the official database was taken down, amid fears of suppressed evidence of corruption (TICTeC 2025 presentation).

Transcribing from video

Going beyond reconciliation of video and transcripts is extracting transcripts where just the video is available, either because no transcript is produced or there is a significant delay between the availability of video and the edited transcript.

For instance in Taiwan, there are twenty day delays between video being released and the documents of those meetings becoming available. OpenFun are exploring automated transcription as a way of making useful information available before documents are released. Tainan Sprout have explored automated transcription to get around long delays via their Tainan City Council WatchBot, but found a combination of OpenAI's Whisper's performance in Mandarin + local/political vocabulary produced transcription errors and misleading text, which made answers based on the data less accurate.

For CRTA's Open Parliament project in Serbia, a challenge arises where committee sessions have videos but no transcripts. This is especially important as more substantial content has moved from the plenary to (less observed) committee stages. The challenge here is to chain together several different approaches to produce useful transcription in Serbian; currently different models are used to identify the words spoken, and separately reconcile this to speakers, using OpenAI to glue back together the results.

CalMatters's Digital Democracy,a PMO tracking the California legislature, uses the legal requirement for the Senate and Assembly to post video online within 24 hours as the foundation of their services. The video is sent to an online transcription service (using AI approaches), followed by human correction to the automated transcript. The amount of manual review needed to link speakers and text has been reduced through the use of voice print matching (building on the Resemblyzer library) to identify who is speaking, which can be reconciled with transcription of what was said.

Video-first approaches have a number of different overheads compared to cases where the official transcript is available, both in accurate transcription and labelling of speakers. Official parliamentary transcripts often contain more information than can be heard aloud, imposing a consistency on the parliamentary record that is very useful as a building block connecting it to other data. But where transcripts are not available, the lowering costs of automated transcript and speech recognition is making the creation of "good enough" transcripts possible at scale.

Summaries of content

Another potential use of LLM approaches is to produce summaries of parliamentary content. Parliamentary debates can have a structure that is difficult to perceive, or is so formal that it makes them hard for people to understand.

Automated summaries have the potential to help more people understand what has been happening in a Parliament, and navigate the original content more effectively. CRTA uses LLM approaches to summarise on top of their automated transcripts, and create briefings for the media that can be turned around in a day. This kind of approach might be especially important when key communication channels for the PMO are already distinct from their main website. For instance Odekro in Ghana is investigating the sharing of AI-generated summaries of parliament documents via WhatsApp, SMS or email, and bots that allow people to engage with the topics they're interested in.

LLM approaches can be used to help join different sources of data through "fuzzy" matches between fields in each source, enriching understanding of the content and drawing connections between different areas of parliamentary content. Automated analysis can also help draw attention to potential outliers (eg connections between activities and declared financial interests) that might merit further investigation.

But the key question in this kind of automated production of information is accuracy. Summaries always represent a decision about what something is really about, what aspects are important, and which can be reduced. Parliamentary proceedings are a high stakes environment for accurate coverage, where summaries might be politically contestable, and parliamentarians potentially litigious (especially as part of wider challenges to a parliamentary monitoring project).

One approach to balance the benefits of automated summaries and analysis with the risk of publishing them is to restrict access to expert users. Calmatters' Digital Democracy produces "AI Tip Sheets", which explicitly recognise that certain kinds of automated content (in this case mostly outlier detection in the data they hold) exist in a complicated "sometimes useful, sometimes inaccurate" box. These tip sheets are not publicly available, but California-based journalists can get access. The idea is that these provide useful tips to local journalists covering local politicians, while still needing extra work to validate that the story leads somewhere.

As Calmatters' CEO Neil Chase puts it: "The part that we're not exposing to the public is what I equate to the kind of thing that's in a reporter's notebook. They are notes. They are not facts. They're not things we can publish." Gated access to "buyer beware" information helps support a wider ecosystem of infomediaries, and get value out of new technologies, while minimising the reputational and legal risk of inaccurate information.

With use of LLM content, there are specific questions about the accuracy of outputs which can be validated and evaluated — but the biggest question for PMOs is around their approach to organisational risk. The question is not so much a binary "do we use this or not?" but "is there an approach with an appropriate trade-off between risk and reward?" A clear focus on both the problem being solved and the route to impact helps ground decisions about technology use.

Conclusion

Democratic transparency projects rest on the idea that sharing information about democratic institutions helps shape citizen and civic relationships with democracy, and how those institutions internally function. Supporting global democratic transparency is less about exporting a single model and more about strengthening a shared capability. Regional networks are best placed to help strengthen practical skills in running PMO/DTPs: what's missing at a global level is the flow of ideas and approaches to innovative technical approaches.

New AI-assisted techniques shift what's feasible by lowering the cost of turning unstructured democratic information into something searchable and actionable. They also raise in new ways longstanding questions about trade-offs between accuracy and transparency — and the importance of building approaches that work for the long term. The civic redundancy that democratic transparency projects can provide helps lock in transparency progress, and shape the wider argument for asking for more from democratic institutions.

Digital infrastructure: foundations for democratic futures

Technology is not neutral: some systems make it easier to concentrate power, while others make it easier to distribute it. That difference is not only in how tools are used, but in what gets built in the first place at the bottom of the stack: what are the choices made in the infrastructure that everyone must build on and pass through?

The respective availability of different technologies is a fact on the ground that shapes the society we live in. In this framing, pro-democratic technology is technology that makes different forms of democracy easier in practice, building the connective fibre and transparency that democratic life depends on. By investing in that infrastructure, we can make democratic life easier and more robust.

That's why today's choices about digital infrastructure matter for democracy. As states and companies expand foundational digital systems, those "base layer" decisions quietly shape what is possible higher up: whether people can contest decisions, organise, and hold power to account, or whether systems default toward surveillance, lock-in, and central control.

A running thread through years of TICTeC conferences has been that of civic tech projects that either are consciously designed as infrastructure, or require infrastructure to be created to achieve their goals. In the last few years a new field has emerged around the idea of digital public infrastructure (DPI), as a foundational layer of government and private services, but with arguments about what the scope of this idea is, and an important question of what ‘public' means in practice.

This chapter explores a simple question: what would it mean to build infrastructure that makes democratic life easier to imagine and put into practice? We argue that pro-democracy approaches have to work from the foundations upward: influencing mainstream DPI so it is governed for the public good, while also building an explicitly democratic layer of infrastructure (the civic democratic stack) both inside and outside the state that provides resilience against backsliding.

This chapter is useful for civic tech practitioners for a grounding in how best to think about (and pitch) infrastructure projects, but also for helping democracy advocates in seeing digital infrastructure as an important area of advocacy and work.

What is infrastructure?

Infrastructure (both physical and digital) has a distinct set of features that make it potentially very impactful, but that at the same time make it a complicated proposition to spend money on.

Infrastructure does not directly have impact, but is required to do everything required by the modern state/market economy. It involves spending upfront to make a set of activities (travel, transactions) much cheaper. This ideally unlocks wider benefits of scale/specialisation.

Infrastructure is obviously important, but the diffuse nature of its benefits makes normal cost-benefit analysis really tricky. The downstream benefits of infrastructure are difficult to attribute back to that infrastructure. Many of the impacts of "making something much much cheaper" will be something you didn't anticipate ("spillover benefits").

As a result, financing infrastructure is complicated. Many (but not all) costs are upfront, while benefits are spread over time. Recovering costs from users is difficult, because the cost of providing infrastructure to any one user can be effectively zero, but the benefits are unevenly spread among users. You need funding that takes the long view and is comfortable with these fuzzy benefits. As such, public funding is one of the better aligned funding sources for infrastructure.

Digital Public Infrastructure

Digital Public Infrastructure is an umbrella term for the idea that there are foundational digital services (such as digital identity, digital payments, and address infrastructure) that should be created to enable government services, commercial services and economic growth.

It's a term intended in part to build support for funding from governments and international bodies, and so benefits from being a broad concept that can be adjusted to different circumstances. When we dig into the details, there are lots more questions about what DPI should mean. Is it wider than identity and payment? What does the ‘public' bit mean here? Does ‘Visa/Mastercard' count as underlying digital infrastructure? Does the identity system need to be publicly owned or can it be the crypto-orb? There's a quick reference in the UNDP description to being "governed for the public good", but what does this look like in practice?

The public question is discussed in this paper by Mariana Mazzucato, David Eaves and Beatriz Vasconcellos, who try to push the DPI discussion towards one in which governments have a stronger role. There's a clear reason to have an eye on when "public infrastructure" is not publicly owned. A key element of foundational infrastructure is that everything goes through it. A reason private corporations want to run your infrastructure is that they want to make it hard to switch away from them, and to charge you rent forever.

The idea of DPI connects to democracy and democratisation in a few ways. There is a foundational argument that, even if the platform itself is not inherently "democratic", the very act of creating state capacity creates the possibility of effective public services. These build consent to the state and have a stabilising effect on democratic regimes.

Democracy also appears in the "governance" question for these DPI services. The kind of multi-stakeholder approaches necessary to bed-in DPI and make shared infrastructure a success implies, if not democracy directly, then the kind of distribution of power that is well aligned with democratic processes, and an obstacle to consolidation of power. These kinds of arguments try to both de-politicise DPI (by maintaining that the core services are so foundational that they sit below these kinds of concerns) and asserting that this kind of base layer is important to successful democracy further up the stack.

For pro-democracy tech perspectives, this implies three key questions:

• How can we push democratic principles and concerns into the foundational base layer?
• What do explicitly democratic DPI (D-DPI) services look like?
• What should the relationship be between "civic" and "state" D-DPI?

Democracy into the base layer

For infrastructure to serve democratic societies, the argument made needs to be wider than simply improving government efficiency. As Iivi Riivits-Arkonsuo argues when talking about Estonia, "[s]tate legitimacy depends not merely on efficiency but also on the cohesion and ethical integrity of the institutions sustaining the system". The reality can be that foundational digital services can support authoritarian regimes as well as democratic ones.

As demonstrated by India's Aadhaar system, an ID service is enormously useful in giving access to a range of public and private services and enabling widespread participation in the economy, but requires a common key to track people across services. The system is now moving towards a privacy-preserving position that allows people to validate identity using tokens/virtual IDs without sharing the underlying ID.

For an authoritarian regime, surveillance can be an intention, rather than an emergent side-effect. Centralised frameworks for identity and payments create new opportunities for state surveillance, for instance China's centralised ID for authentication on a variety of web services. A DPI system that does not engage with this question at the start can lay foundations that might be equally suitable to several kinds of regimes, including bad ones.

Another principle that is important to get in at the ground level is that the platforms are interested in the experiences of users, both as users of the platform, and as citizens. In a 2025 TICTeC Seminar, Richard Gevers (Head of Service Design and Delivery at the Digital Services Unit) talked about the importance of DPI being built that can carry two way communication: voice as well as services. He argues DPI-era services are a chance to bake in participatory mechanisms and feedback instead of treating services as "done" once delivered.

Even though the foundational services may not relate to democratic processes directly, DPI can say a lot about democratic values by their orientation to their users. Does it see them as customers, or co-producers of services? Is it concerned with helping them maximise their rights and opportunities, or does it arrive with the orientation of a state seeking to control and surveil? Asking these kinds of questions about the basics helps build infrastructure that is supportive of democratic futures.

Democratic DPI services

Another approach to Democratic DPI is in the implementation of services that directly deliver democratic functions, enabling direct or indirect participation in governance.

For electoral democracy, they might run through electoral registration, to full electronic delivery of elections. But there is also the argument that part of the stack here should be the inclusion of participation services. In Finland, Sitra sees part of its role as an innovation funder to help create deliberative democracy infrastructure, supporting a GDPR compliant version of Pol.is (Voxit), with experiments, pilots, and building the institutional capacity of different providers to run versions of the Pol.is process. A similar approach, in Scotland, is taken by CivTech, which is run as part of the Scottish Government and has supported the development of common participation software for public authorities.

These different approaches to democratic services, which might be run by either the state or civil society, raise an important question: what is the space for infrastructure that comes from outside the state? When is this a good fit, and how should different approaches work together?

The civic democratic stack

In practice, we can think of a lot of civic tech/pro democracy tech as both building and encouraginga civic democratic stack. This is a combination of state and private/third sector infrastructure, where non-state run infrastructure fits into one of the following categories:

• Innovation/no-one responsible - What the state could provide, but does not yet. Outside actors can develop a DPI approach that has become possible, but where for various reasons the state is not investing in itself (often lacking a specific institution with the relevant scope).
• Accountability infrastructure - The state might not provide good (or even half-hearted) infrastructure to hold itself to account. Here, extra work by civic infrastructure builders can enable other accountability actors to go further.
• Civic insurance - having some overlap between state and civic capacity is not redundant, but a vital guard against backsliding. Public authorities can (and do) delete their FOI/ATI disclosure logs, but can't delete third party public databases.

For the innovation track, there is a good alignment of the civic tech philosophy of lobbying by demonstration, and a concern of third-party funders in a plausible eventual exit. Here, outside funders can help build capacity, validate approaches, with a plausible handover to state bodies at a future date (which might take some considerable time). This might include aspects that are established state functions in other jurisdictions (e.g. access to legal/parliamentary information), or experimental infrastructure that has not yet made the case of being a core state function (and this is part of the argument for doing so), e.g. participation/deliberation infrastructure.

Accountability infrastructure has a clearer "this should not be run by the state" rationale, which means it is in the difficult position of trying to create infrastructure without plausible current or future alignment to public funding (this is a similar position to other accountability tools like public interest journalism). A civic democratic stack works best both when the state is active and interested in supplying core components when it is best placed to do so; but also when recognising that a stack that fulfills a wider democratic interest involves important components outside state control. A compelling argument for funding these components needs to make the case for the unique properties and rewards of creating infrastructure as a way of empowering many projects at once.

Making the argument for funding infrastructure

A civic tech complaint that has been voiced repeatedly over the years is the reluctance of funders to provide money for maintenance and ongoing projects. Funders generally (with many honourable exceptions) like new ideas because they want to use their limited resources to fund innovation that ideally demonstrates its worth and then goes on to find other sources of funding.

From this perspective, maintaining infrastructure seems to have quite a large opportunity cost in supporting innovation. If something is going to structurally find it difficult to be self-supporting, you will find yourself facing the same funding problem in five years' time.

In some cases the desire to fund projects that benefit a specific group will select against infrastructure, because so much of the overall value of a project is in the spillover benefits. The spillover benefits to infrastructure are part of what makes investment in infrastructure so powerful. But if you score applicants based on "what percentage of the benefit reaches the groups we care about" infrastructure projects will score poorly.

Legal Information Institutes are a good example of the rewards of funding infrastructure: they are consciously infrastructure that do not achieve outcomes directly, but have a transformative effect on the wider legal system, levelling a playing field and especially making legal information available to law students who would not otherwise have been able to access vital information for understanding the legal system. But at the same time, the resource is also used by those who would previously have paid for access to expensive legal services. Social justice aims can be accomplished by investing in foundational infrastructure, but the beneficiaries of the project are likely to also include winners in the current system.

When done well, infrastructure and maintenance is not a money pit: it's a way of spending money to achieve goals. Foundational infrastructure is hugely pro-innovation because it means people can build on it to focus on the interesting bit of their idea, rather than failing because the basics were not there. It is easier to grow a lot of seeds if someone has built an irrigation system.

For those that want to build and run this kind of infrastructure, it's important to have a clear eye on how the value created may be different to other kinds of projects, and how best to capture and explain that impact when talking to funders.

Measuring impact

Infrastructure poses problems around finding a common language for describing and measuring impact. For instance, nonprofit organisations that directly make something happen can say "your donation makes [X] [things]" happen, making a tangible connection between even small donations and impact. Infrastructure does not work in units like that. The cost to one additional FOI/ATI request sent through an Alaveteli site is zero. But the aggregate cost of the infrastructure and staff required to make that infrastructure exist is high.

The IIPP/Bennett Institute paper on assessing the value of DPI suggests impact measurements. These break down in similar ways to that we've described mySociety's impact to funders, and might be a more useful wider framing for pro-democracy tech to adopt.

Their three groups are Direct, Dynamic and Market/Society Shaping.

For instance, a parliamentary monitoring feature like TheyWorkForYou's email alerts can be described like this:

• Direct: Citizens are able to get messages about their MPs parliamentary activity,
• Dynamic: The same feature is used to support free parliamentary monitoring for a range of organisations
• Market/Society Shaping: Parliamentary activity is much more public, tied into a range of other information systems. Smoother flow of information inside government, more difficult market for paid parliamentary monitoring, more equal access to information.

Democracy from top to bottom

We need more people invested in democratic futures to be interested in digital infrastructure.

The decisions that are made at the very bottom of the stack shape what is possible all the way up. Decisions about DPI — what's included, its goals, governance and safeguards — strongly benefit from people in the room making the case for baking in democratic ideals as close to the foundations as possible.

On the technical side, we also benefit from crossing the streams between civic tech's interest in open source, and arguments about sovereignty, vendor lock-in and dependence on Big Tech. With AI, this extends to thinking about investment in public AI models as a way of creating resilience in our usage of new technologies.

Moving beyond the foundations, building infrastructure is the one the most effective ways of empowering a diverse group of people and institutions to have access to information and be able to act on it. Civic tech creators need to talk in the language of infrastructure, its unique features and drawbacks, to make the case for funding truly impactful work that can make a difference at scale.

On the other side of this, funders should think about how infrastructure can fit into their wider portfolio of work. This applies especially where there are useful synergies and alignment between grantees that mean infrastructure will also enable different grantees to do what they are best positioned to do.

Distributing democracy: having impact through reaching mass audiences

One of the distinctive features of digital tech is the ability to scale. While each new user is not costless, it is much cheaper than in non-digital approaches, and it is possible to reach a mass audience more cheaply and easily than it would be through non-digital methods.

Civic tech can still scale — but the routes to mass reach have shifted over time. A historically successful approach of publishing useful resources and accumulating a search footprint has become less reliable as attention shifts into closed social platforms and as search itself is reshaped by AI summaries and chatbot interfaces. The idea of "build it and they will come" was never good advice, but did sometimes work. Now positive discovery feedback loops are weaker, and more duplication and explicit effort is needed in distribution.

Mass reach matters for three reasons. First, it has direct public benefit: more people can be informed, supported, and empowered. Second, it can be system-shaping: widely used services become part of the democratic ecosystem, opening doors to influence institutions and reforms. Third, reach underpins sustainability: it supports fundraising, partnerships, and the ability to demonstrate impact. But having mass reach is not just a matter of producing good work: it means understanding the nature of the digital ecosystem in different contexts, and how this has changed over time.

In 2025, we held a TICTeC seminar learning from our community about different approaches people have taken to reaching a mass audience. This chapter builds on that seminar to explore the problem, and different approaches and strategies that are available. We begin by describing the decline of the open web as a growth channel, then compare a set of strategic responses: producing native video content ("be the influencer"), partnering with creators ("influence the influencer"), building chat-based access points, and attempting to shape large AI intermediaries. We also explore, alongside, the deliberate choice to opt out of these channels and invest in more resilient routes such as email, membership, or offline products. Across each of these we examine the trade-offs in cost, control and sustainability.

The rise and fall of the open web

Historically, our own route to service growth and impact has exploited the affordances of an ecosystem that is slowly disappearing. Many of mySociety's successful services have built userbases through a combination of weak network effects and effective search platforms. This means that while the sites work from day one (you can make an FOI/ATI request), they become more useful through other people using them (get information others have requested without needing to do so yourself). These kinds of publicly published reports/requests importantly have an impact in search engine placement: as the footprints of the site increase and cover an increasing number of keywords, it's easy for people to find the site to use it for similar purposes.

There are a few assumptions this is based on: mass access to the internet, an open web where people can freely access all websites, and search engines that work cooperatively with that open web, rather than extractively.

One of the key changes of the last 20 years has been an enormous growth in internet use (with an estimate of 68% of the world online in 2024). But this has not been a steady growth of the internet that existed in 2005: more and more internet usage happens in the semi-open and closed ecosystems of different social media platforms.

Sometimes this is structurally engineered, such as through internet access packages that make access to social media cheaper than the general web (in some cases, when the infrastructure is being funded by social media companies like Meta).

In other cases, there has been a shift over time, where platforms that originally grew out of the open web become more closed and self-contained. BlueSky is a relatively small social network, but at this point is unusual in not actively punishing (through deprioritising, etc) posts that contain links to other sites. Most social networks now need content to be entirely within the format of the site, with elaborate workarounds necessary if you want to include links (‘link in bio' for Instagram; links a few items down in the thread will be accessible on X, etc). The purpose of this is to keep users on the site, increasing the amount of time spent there overall, and the value of ads that can be shown to users.

The growth of AI chatbots poses a new challenge as these are both a logical successor to social media (where the algorithm now creates as well as curates content) and direct competition with traditional search. As a result, search engines have adopted the same tactics and become increasingly extractive, displaying content to their users directly in AI summaries, rather than encouraging clicking through to get more information.

As a result, online journalism and projects like Wikipedia are reporting big decreases in traffic as content is served directly through chatbots. This is not necessarily all downside: content taken from low-traffic sites might find a bigger audience, but this happens increasingly on the platform's terms, without providing clear evidence of this use after the fact. Without that direct link, it is harder for these projects to be sustainable: for example, in Wikipedia's case if there are "fewer visits to Wikipedia, fewer volunteers may grow and enrich the content, and fewer individual donors may support this work". Similarly, a move from primarily text-based social media (that can contain links to content) to primarily video-based content (which actively makes it hard to link to other content off the platform) means that information can travel in new forms, without linking back to support the creation of that work.

These are challenges that existing projects need to adapt to, as old models for growing an audience are increasingly inapplicable to new projects. Word of mouth and building a search footprint look very different when both of those activities are substantially intermediated by algorithms controlled by private companies.

Approaches to video and AI

When planning how they will reach mass audiences, modern pro-democracy tech projects need to have a considered approach to both video-based algorithmic social media (TikTok and similar) and AI chatbots (ChatGPT and similar).

There are different approaches that can be taken to this.

For video-based social media:

• Be the influencer: investing resources in employing people who can navigate social media
• Influence the influencer: providing resources and access to experts to influencers with existing big reach.

For AI:

• Build your own chatbot: in both AI and non-AI variants, online and instant messenger chatbots can present useful information through a familiar source.
• Influence the chatbot: the big platform LLMs are not unchangeable, and can be influenced through data, advocacy and regulations.

And an important option related to either channel:

• Ignore these growth channels and do something else: make investments in other approaches that build sustainability.

A considered approach might not be "we have to invest in this area", but "we have weighed up the options and decided to do something else".

Becoming the influencer

One approach to reaching people on new platforms is to produce content in a way that works well on that platform.

ParliamentWatch in Uganda is an example of a parliamentary monitoring organisation that has actively invested in creating video content to reach audiences they would not reach through other means. Their website is an informational core for a wider range of activity that supports different audiences where the expectation is that only a small number of the people they seek to inform will actually directly use it. The site is integrated through direct relationships with civil society groups, a Media Accountability Network that connects with over 70 news editors in local radio stations (which are in turn a key source of public information), and a video approach to target Gen Z - important in a country where a large proportion of the population is under 18.

With funding from the National Endowment for Democracy they invested in a multimedia studio to produce native video content, in short form on TikTok and a longer "Policy Digest" on YouTube. As Parliament Watch's Joseph Tahinduka put it:

"What we do is that if a bill has happened on the floor of Parliament, we break it down, and then we get the video running on social media, and we ensure that that video travels as far as it can. We do this for committee reports in Parliament. We do this for bills themselves, for statements for issues, when, let's say, certain agencies are appearing in Parliament, we run a poll on our social media platforms to ensure that citizens can send in their comments, which we then transmit to our contacts."

Parliament monitoring in the open internet era benefited from two big gains in efficiency: automated approaches to search and data crunching that reduced the work of producing accessible analysis; and easy access to a large audience. For PMOs in contexts where that audience is elsewhere, that first efficiency is still incredibly important,making it easier and faster to create information that can be distributed through other channels. As Tahinduka put it: "We know people rarely visit these things. So what do we do? We said, OK, we're going to find you where you are. We will not wait for you. We will ensure that wherever you are, we find you there."

An advantage of this approach is the ability to build channels and approaches that are coherent and wider recognition of the brand. But this also has significant costs. While the "authenticity" of low production values is rewarded by new social media, this is still not costless, and is much more time-consuming to produce than typing text based posts.

There is a "dark pattern" available here. Misinformation providers create content farms where misinformation is spoken by AI voices accompanying "background gameplay" videos (like Minecraft Parkour') that are visually frenetic and less likely to be swiped away. This is a relatively cheap approach that could be used by organised pro-democracy actors to convert text factional information; however its contribution to the degradation of the media ecosystem overall may make it an unattractive one.

Influencing the influencers

Another approach is not to create your own native content, but work to have indirect impact through influencers with existing audiences.

While TikTok and other video platforms make it hard to link back to other sources, in practice content on these short form video platforms has a strong umbilical cord to the more traditional media. Many highly viewed posts come from entrepreneurial creators who are just reading out things they read elsewhere for their audience, with the message being "I'm just like you except I've read a thing", making the messengers both relatable and informed. In turn, content on these sites become part of traditional news stories – they are connected ecosystems.

In this sense, influencers on video social media can be put in a similar box to traditional journalists. You do not need to replicate their skills and reach to gain benefits from them, but you do need to understand how they work, their incentives, and how you can best accomplish goals by working together.

In Georgia, ForSet ran a campaign to get people to register to vote using social media influencers (TICTeC 2025 presentation slides). Their method managed the inherent tension of this approach: trying to understand how to get the best from both sides, enabling creative freedom through strong early work on understanding alignment and values.

This process started with desk research and focus groups to understand where this audience are (TikTok, Instagram, YouTube) and what they actually want to hear about elections. From an initial pool of 65 candidates, they selected 17 influencers who shared democratic values, including non-political creators (eg lifestyle/skincare etc) to break the bubble and reach new audiences.

The campaign gave influencers broad creative freedom but with guidance and mentoring from ForSet and election experts, so content stayed accurate and safe but still felt authentic. The campaign experimented with different approaches, including informational content, but also actions and trends related approaches ("_Get ready with m_e to go vote") Over the course of the campaign, 17 content creators published 145 videos on social media, bringing 9.4 million views, 405k likes, 40k shares, 24k saves, 14k comments, an 8% engagement rate, and 1100+ user-generated videos around four trends that the influencers had seeded.

While there are clear benefits of this approach in avoiding the need to not just invest in a new medium, but to become good at the new medium, this is not as simple as "give TikTokers some fact sheets". The work required to manage this bridge between two approaches is complicated; it's a creative co-creation exercise to reach the necessary audience, paying for the intermediaries' professional expertise, ensuring appropriate transparency and keeping a careful eye on value alignment, as well as managing the risks to content creators themselves.

Building instant messenger chatbots

A huge amount of social and democratic activity takes place on instant messaging platforms like WhatsApp.

One idea about the future of this kind of social media activity was that these would be platforms that hosted chatbots run by a range of organisations (again, as a closed ecosystem that stopped people needing to venture onto the open web — and requiring a fee to be paid to the platform owner). While the platforms themselves are now interested in integrating their own AI chatbots directly, this is still an interesting approach to explore. A key question in the decision "should we make our own AI-powered chatbot?" is "how do we reach a wider audience than the big chatbots?" — and part of the answer to that is to make your version easier to access in more ways.

This approach can be used both by public bodies and civil society organisations. In the 2024 Mexican elections, the National Electoral Institute (INE) ran an WhatsApp service called Inésthat gave quick answers to key questions about voting — which documents were needed, etc — as well as connecting people to fact checking organisations. In Venezuela's 2024 elections, a group of academics and social activists explored AI-powered chatbots to give information and report electoral fraud via Facebook Messenger, Instagram and WhatsApp APIs. ShineYourEye(run by EiE Nigeria) has a core traditional parliamentary monitoring website, but is made more accessible via a WhatsApp chat bot, alongside social media campaigns. This enables people to access important information about candidates, and where polling areas are, in an app they already use and have wide access to.

The downside of this approach is exactly what was appealing to the platforms themselves: lock-in to the platform and the potential for charges to be levied (that said, for WhatsApp, ‘Service' messages are currently free to enable responses to user messages through the chatbot), but this might be a happy medium in terms of being accessible to a wider audience, while having control of the full interaction.

Influencing the big chatbots

There is a clear trend towards online content being intermediated by platform AI chatbots, where there may be specialist services in some areas, but generally (like search) most queries are going through a few major providers.

The problem for a pro-democracy chatbot embedded on a website or other platform is that it is competing with the much better known, multipurpose chatbot like ChatGPT. There is clear evidence that people are using these general tools as sources of political information. This multi-purpose chatbot might be worse at the intended task, but will continue to be used if what it returns keeps the user from discovering a more tailored or better informed approach. Here, an alternate goal is not to compete with these platforms, but to adopt a strategy to shape their responses. This is the area that is most speculative and subject to change as platforms shape their approaches.

Shaping the outputs of these platforms is an attainable policy goal. AI chatbots are a logical continuation of algorithmic social media (and can be thought of as ‘post-social media' that replicate some of the form of social media companies even while containing no actual people). Strategies that are effective for social media such as producing content that fits into the shape expected by the platform, which in turn shapes its outputs, are viable. At a bigger scale, regulatory pressure can successfully shape these outputs. Neither one of these approaches inherently leads to ‘good' outcomes, but reflect that as AI chatbots become an influential way of distributing information, a key part of getting information to a wide audience is attempting to shape how these platforms present content.

There are three main approaches that can be taken:

• Shape the training data
• Shape the search results
• Shape the prompt through regulatory pressure

As with SEO, the platforms themselves are trying to be resistant to attempts to influence their outputs, and so viable approaches change over time.

Training data

LLMs "learn" through exposure to massive amounts of data that shape their internal structure and responses to inputs.

For example: if we take the problem of "we want LLMs to be better at helping people write messages to their elected representatives", different kinds of LLMs draw from their training data in different ways.

• A basic LLM would produce an output based on its exposure to messages contained within its training data (which is likely to draw from a wider corpus of formal letters rather than a dataset of actual messages)
• A "thinking" model might first prompt itself to consider what a good message looks like, sharing some guidance based on exposure to writing guides in its training data, and this would then shape the actual message it writes.

For the first kind of model, creating a synthetic dataset of "good" messages could be used to fine-tune a model for this purpose, and could also be published to influence the behaviour of the big LLMs.

For the second, encouraging a range of different authoritative sources to adopt similar guides (not just linking to content, but republishing it) could get clearer information on better style into the training data — which in turn shapes the outputs of more sophisticated LLMs.

Currently this is better understood as a misinformation tactic, with Russian based disinformation approaches aimed at creating large amounts of internet content that in turn influence AI models.

This context means that chatbot creators are aware of the risk of overreliance on training data, and increasingly supplement it with additional tools like web search, and long prompts that try to minimise the risk of the training data being hijacked. This is not to say work at this level cannot be influential, just that training data is no longer the only pool of knowledge that LLMs draw on.

Search results

Search-based chatbots execute a search strategy (on a traditional search platform) before "word calculating" summaries based on the information retrieved. When used as a search engine, ChatGPT rewrites the query into a few more targeted searches before returning an aggregate summary.

This means that Search Engine Optimisation (SEO) is still important, but that the AI searches more like an expert user: not asking statements with a question mark, but looking for concrete keywords based on the query that it will then filter. Speculatively, as its review stage probably only pulls on a sample of the document, more webpages with less content (and a good selection of keywords) might be a more effective way of helping pages pass a "relevant to this query" test.

This is a more flexible and influenceable approach, but only if the chatbot correctly determines instances where it needs to search for better advice. On the example above (as of December 2025), if prompted with a request to write a message reporting mould in council-provided housing in the UK, and also asking it to ‘research how to best write this message', it will find guidance from a UK housing and homelessness charity, and draw on this for the template it creates (including, importantly, giving consent for the MP to act on the correspondent's behalf). Without being asked to research, it will not do this search, only pulling on its training data, and does not tend to include the consent point.

Again, the best documented examples of this approach are in disinformation: for instance, an AI-generated fake news site around the 2025 Bondi Beach shooting was quickly amplified by Grok.

The knowledge that people will obviously try to shape the results from search queries necessitates more instructions in the prompt on how to balance different approaches, which makes understanding and shaping the prompt important in shaping the overall system.

Shaping the prompt

Current LLM services are best understood as a series of shells around the core LLMs, with tools and prompts that shape how they interpret search results and the outputs of their own training data. This is important because it means that the chatbots present information to users that results from search strategies and prompts that are more directly controllable by the platform creators themselves than when they were solely reliant on training data.

What can currently be seen in ChatGPT's (5.2) approach to search is an attempt to draw on multiple sources where possible, avoiding a single source shaping too much of the output. The downside of this approach could be false equivalences between sources of different standards. But this is a strategy that OpenAI, the company that produces ChatGPT, has adopted rather than a requirement of the technology.

For instance, it is perfectly practically possible for OpenAI to prompt ChatGPT to prefer answers from certain websites in response to certain kinds of queries (eg answers from Electoral Commissions in response to election queries), and similarly to the way social media websites built civic tools into their platform, could be encouraged to do so.

But this greater customisation approach also shows one of the big risks of mixing democratic information and chatbots: the platform might not be interested in being neutral, but actively trying to push one side in a political debate. X's Grok AI chatbot is a conscious project to "rebalance" the internet by promoting the political agenda of the company's owner. Grok's insistence that Elon Musk is, personally, the top of every field is a clear demonstration of how shapeable these platforms are. The concern is not just what the company producing AI chatbots wants to do in political terms, but that companies are also exposed to pressure from government and other political actors trying to shape what "neutral" responses look like.

This should caution us at an individual project level against approaches that rely entirely on shaping the behaviour of big chatbots. But it also raises the importance of the issue of creating countervailing pressures towards good information on a regulatory level. Technologists can do practical work to show what this looks like, for example creating ‘eval' evaluation datasets that LLMs can be scored against for the provision of democratic information. These are both useful in creating and refining our own LLM approaches, and in demonstrating where big providers are failing.

Ignore and do something else

A final approach is to challenge the idea that needing to "pivot to video" or "use AI" is just something inevitable. It might be reasonable, but it can also be a prompt to consider other options.

This approach does not pretend that a shift has not happened, but for various reasons sees it as a bad idea to go with the flow. Recognising the challenges that adopting new approaches can bring, this is a choice to address the problem by doing something else.

A key reason to be cautious is that approaches to new platforms require investment, and this might be building a house on the sand. Algorithmic platforms may pull a bait-and-switch where having trapped you in their ecosystem, they then start to limit access to your own followers. Sometimes the data they provide on engagement is just bad, and high numbers have no connection to actual impact. Many organisations built big reaches on Twitter that became less and less reliable as a way of reaching people. Changes to algorithms mean hard-won followers and service-specific knowledge become less relevant.

Here, there is a solid fallback option in the original social network: email. Email newsletters are portable, and followers can be moved from service to service. While services like Substack attempt to build some aspects of the walled garden, ultimately lists can be moved out to other providers (eg Ghost) and this hybrid approach is part of the pitch of the platform. This is a robust approach that goes in and out of style but fundamentally remains in your control.

Similarly, if doing well on video social media requires time and resource investment, it's worth considering the opportunity cost against other things — especially when part of the concern is around threats to the current revenue model. Building on the email newsletters example (which can be developed into a gated, premium product), civic tech products could be invested in print newsletters and other physical products as a way of reaching specific audiences.

For instance, PMG in South Africa has in the past distributed a quarterly Spark newsletter to a range of official offices and democracy centres, and a print option of their monthly newsletter. In journalism, print models accompany online publications as part of a revenue strategy around membership and subscriptions. The UK based Bristol Cable produces a quarterly print newspaper; the US-based The Onion satirical website returned to print as part of a wider push towards a membership model. Here print is less the primary means of mass distribution, but an anchoring in an engaged group of subscribers who financially support the wider project.

The key point here is that decisions made should be deliberate, and with an eye for the opportunity costs of following the crowd. Rather than trying to influence AI crawlers with your content, it's an option to try and use Cloudflare or similar future services to block AI crawlers. Services can make it harder for the AI companies to intermediate your work, while adopting new approaches to reach audiences and build sustainability. This is not necessarily an easy option, or one that's guaranteed success — but if that's also true of all approaches, it's worth thinking clearly about the goals of your project, and what's really the best way of achieving them.

Conclusion

With the decline of the open web, there has been a disconnect between the efficiency gains in using technology to process and produce information, with the ease of discovery of this information by a mass audience.

A shift over the last decade is that a website is less likely to be the primary way information reaches a mass audience. The wider "service" needs to include both how the organisation directly reformats its information for a wider audience, or how it influences intermediaries to spread information (whether these are human or machine).

The exact forms this takes depend on a range of factors, and it needs a considered approach based on the local context and resources. No path represents guaranteed success, and a flexible approach is needed to make the most of the current possibilities, and understand how to lean into (or stand against) shifts in the landscape.

Evaluating and understanding AI

In the last few years of TICTeC we've seen projects involving machine learning (ML) or large language models (LLMs) rapidly move from a niche area to a whole field of work: exploring not just how these tools can be practically useful for pro-democracy work, but how we should approach democratic issues around their governance and use.

In this chapter, we distinguish between AI technologies and "AI" as a wider phenomenon, which includes hype and raised expectations that can push organisations towards approaches that are poorly aligned with their needs. As AI is a fast moving field, the goal of this chapter is to provide some central concepts and tools that are useful in evaluating and understanding claims about it.

This chapter covers practical tools for decision-making, a summary of our AI Framework, a production-versus-verification lens for assessing whether an LLM is a good fit for a task, and guidance on choosing models (open vs closed; large vs small) and evaluating them through structured "evals" and audits. Approaching new capabilities with a clear idea of the goal in mind is important in choosing the right technical tool for the task.

AI technologies

AI is a broad label, which is applied both to a wide set of machine learning approaches, and to the narrower set of large language model (LLM) generative AI approaches.

The key distinctive feature of these technologies is that the model itself is mostly a black box, created from training data. Unlike procedural computer code, it can't be evaluated on its logic; it can only be evaluated on its outputs.

LLMs are a kind of machine learning, but at such a scale that they have become their own set of generalisable language tools. While production use of machine learning might require fine-tuning a model with additional bespoke training data, the size of LLMs allow zero-shot approaches, where describing a problem can be enough to create a useful solution.

This is both a very flexible approach, but also the source of one of the key drawbacks with LLMs: there is no mechanical difference between the prompt (the instructions on what the model should do) and the context (the data to be manipulated or ongoing conversation). While there are efforts to reduce this problem, described as ‘guardrails', the technology is fundamentally vulnerable to input from untrusted sources (e.g. an instruction to read and evaluate a document might be overwritten language within the document that instructs it to return a different result) changing the underlying instructions (an approach known as ‘prompt injection').

The best way to think about them might be as a calculator for words: they are at their most effective when applying transformations to language. Other kinds of problems need to be restated as language problems, and some of the confusion about their capabilities is that while this approach is enormously powerful, it also has limits. While many problems can be restated as language problems, not all problems are best solved as language problems (leading to the situation where LLMs often have difficulty giving the right answers to maths problems). A lot of the language used (even "AI" itself) implies that there is some element or thinking or reasoning happening, but this is limited to the extent that thinking or reasoning can be restated as a problem of creating an internally consistent block of text.

LLMs have a different set of strengths and weaknesses to what we generally understand computer-based approaches to have. As Colin Megil put it in his TICTeC presentation on Pol.is "you have to squeeze the alien squid between other things". A lot of practical advances come from trying to wrap this alien core in frameworks that counteract some of the fundamental drawbacks of the technology.

For instance "thinking" models, which aim to create better alignment between input and output, do this by being prompted to produce intermediate outputs ("thoughts") on aspects of the problem and how this aligns with instructions, before producing an output that is coherent with those intermediate outputs, and so hopefully more coherent with the original input. Nothing about these steps means that the core LLM is itself smarter, just that the framework around it can be adapted to pump things in and out in a way that improves the final output.

Similarly, "agent" approaches are when the language models are shown how to format requests for use of other tools. These might be mechanical computer functions (such as running code or doing calculations), or an LLM with a separate prompt, which can then be fed back into the original context, to be "word calculated" to the final output. This can help adapt for weaknesses in LLMs' core approach (eg "if you see something that looks like a maths problem, defer to this external tool"), but can also enhance some of their fundamental vulnerabilities. An LLM with access to a tool that lets it read and send your emails might read an email that instructs it to send sensitive information from one email to another address.

As such, there are drawbacks with the technology that are likely to be fundamental and should make us intensely cautious about this kind of deep integration — and given that such integration has continued despite this risk, we need to understand the wider context of why technology has been pushed in inappropriate ways.

AI (the tool) vs AI (the phenomenon)

It is important to distinguish between AI technologies (such as machine learning, LLMs) and "AI" as a socio-technical phenomenon. When possible in this report we will prefer to talk about either machine learning (ML) or LLMs rather than AI if that is the more appropriate description.

LLMs and ML are practical technologies, but exist in a wider environment that is deeply opposed to the reality that there are boundaries on their usefulness, and where there are incentives to describe other (working well!) non-AI systems as AI, creating a confusing picture of what is happening.

This report is being written during a period where there is not just a lot of "AI" hype, but massive investment of capital in AI infrastructure and hugely subsidised/free access to LLMs. Big players with fundamentally unviable business models are attempting to establish a defendable position by converting current financial dominance into infrastructure dominance, and more widely through administrative and economic integration with governments and big businesses.

This is a competition that has seen search products released that are more expensive to run while returning inaccurate results, or video-generation approaches that have a clear risk of harm to individuals and society at the same time as having no obvious business model (for instance, OpenAI's Sora-2). There are explanations as to why this is happening, but the effect has been a dramatic challenge to the idea of aligning products with user needs. These companies cultivate the idea of AI as the big wave that cannot be missed, hoping that fear of missing out sidelines people in organisations and governments asking sensible, practical questions about the utility of these products for problems they actually have.

LLMs as a technology are likely to survive this moment (and one of the potential ways the bubble could burst is the development of smaller LLMs with less of an infrastructure requirement), but this makes it all the more important to focus now on the technology, understand more about sustainable approaches, and think hard about the alignment of problems that do have LLM solutions.

Questions to ask

To inform our own thinking, we wrote down a series of questions we wanted to ask ourselves about potential uses of AI.

This starts with what we thought should be an uncontroversial statement: We should use AI solutions when they are the best way of solving problems, are compatible with our wider ethical principles and reputation, and can be sustainably integrated into our work.

This leads to questions and considerations in six domains:

• Practical – does it solve a real problem for us or our users?
• Societal – does it plausibly result in the kind of social change we want, and have we mitigated change we don't want?
• Legal/ethical – does our use of the tools match up to our wider standards and obligations?
• Reputational – does using this harm how others view us or our services?
• Infrastructural – have we properly considered the costs and benefits over time?
• Environmental – have we specifically accounted for environmental costs?

You can read this in full in our Framework. We've also written some practical guidance on working with open models as an alternative to the major closed models (which we do use, but feel it is important to work in a way that is not dependent on them in the long term).

The feedback we've received on the Framework is that people have found it useful as a way of framing internal discussions, and especially to recapture a focus on the user needs of technological solutions, rather than a rush to apply the latest technological solution.

Our key message is that it is OK not to rush this, and that rushing might lead to approaches that are poorly aligned with what you're trying to do, or that are unsustainable in the long run and difficult to undo once embedded.

Is AI the right fit for a project?

In thinking about how and when to apply AI technologies, we need to understand the ways in which they are substantially different to how both people and traditional computing approaches work.

Generative AI has no concept of truth. It is designed to create outputs that are internally consistent, and this might or might not coincide with true things when the training data and context provide good grounding in true things to explore. The failure mode for this is false-positive hallucinations, where it will claim something exists or was said because doing so is internally consistent with the question.

As such, effective LLM workflows that produce useful outputs will involve separate production and validation processes. These may involve a combination of LLM approaches (eg verification tool use that is itself an LLM), but will be more reliable when part of the verification is a traditional "mechanical" process.

This gives us a useful way to triage how useful an AI approach might be for a certain problem: how hard/expensive is it currently to produce a solution, and how hard/expensive is to verify a correct solution to an acceptable standard. The grid below splits problems into four camps:

For this exercise, it is helpful to think about the purpose of the proposed approach. In some settings, such as when making internal tools for skilled operators, ‘to an acceptable standard' might include a larger amount of error. However, the same basic approach might be less suitable to make available to end users directly. So the same basic task might fall into different boxes depending on the wider context.

Significant ML/AI benefits (producing solution expensive / verifying solution cheap)

In this box, using non-AI means, it is harder to create a credible solution than it is to check if it is a valid solution. There are strong potential benefits here because an expensive process can be made cheaper, while the quality of the output can be checked through relatively cheap verification methods.

One of the biggest practical uses of AI has been in coding, because there are coding problems that fit in this box. The structure of computer code is often formally checkable (for at least syntax errors), and often there is a relatively short turnaround between "having code" and "checking code is effective". This isn't to say that all coding fits in this box, but enough that a clearly productive set of tools exists.

This box can be impactful even where access to models is relatively expensive, as a relatively small number of LLM users can have a big impact through the high-impact products of the work.

Weak AI benefits (at scale) (producing solution cheap / verifying solution cheap)

This is a set of problems where AI tools can be useful in doing things more efficiently, but not in activities that are either high-impact or that represent significant time savings. Benefits here are present while small on an individual level, but might become more significant at scale.

Many of the "copilot" and similar users fit into this box, e.g. letter writing and summarisation (where the user knows the original content). For this set of uses, just getting closer to a valid output can be useful as it reduces work to a point where a human can intervene and make tweaks to the result. Similarly if there are low standards needed for verification, verification costs are cheap and the outputs are useful anyway. For approaches in this box to be viable, LLM access needs to be cheap and widespread. The benefits may not be worthwhile on an individual level if access to LLMs is expensive.

For civic tech applications, these might represent new forms of helping people navigate bureaucracy. For instance, helping people format letters to representatives, or effective appeals when FOI requests are refused. Cheap processes at scale can also unlock new collective benefits. For instance, Muckrock uses LLMs to extract information and success/fail status from individual FOI responses. Doing this manually per request is easy for people, but requires lots of people to do the work to get the systematic benefits of the data. An AI approach drops the costs further, which produces a small benefit on an individual scale, but collectively creates useful systematic data.

This is also an area that engages our ‘societal' evaluation lens, in that the sum of small uses can overload or change the overall impacts of whole systems. For instance, AI-created objections to planning applications might overwhelm a system that assumes objections are much harder to make. A focus on individual effectiveness might have collective costs, and be short-lived if the system adapts to restrict/ration access in a new way.

Get a human to do it (producing solution cheap / verifying solution expensive)

In this box, LLM processes are producing outputs that cannot be quickly verified by automatic or human means.

Here, using an LLM for the initial solution might be less effective than having a human do it from the start. While tweaking an email that's slightly off-optimal is a cheap correction, adjusting an LLM written report (involving fact checking, correction, restructure, etc) might be more complicated than just having someone write the original work.

Looking at human processes, production and verification are often a bit more blurred together, because the skills required to produce the work are the same ones that suggest the work is valid. Trust in the quality of the output often comes from professional standards/past experience, rather than a formal process (although this is a sliding scale).

This is most clear in projects that need a high level of accuracy and confidence in the material produced. We talked to OpenFun about their project LawTrace (bringing together legislative information in Taiwan), and here they made a clear point of choosing not to use AI at all in this project. Having accurate information was far more important to users than any convenience AI could introduce.

Break down verification problem (producing solution expensive / verifying solution expensive)

In this box are problems that are expensive for a combination of reasons, and the combination of high costs can justify investment in trying to move it into a different box by decomposing the verification problem into smaller problems.

Are aspects of the problem (if not the whole problem) appropriate for mechanical checking? Is a custom verification process made easier via LLM code? Can a series of automatic/human checks be made more efficient with a clear verification workflow? Each individual improvement moves this closer to being a potentially strong use of ML/AI.

With investment in the verification process, this might move the problem closer to having weak/strong AI benefits, where LLM outputs can be derisked through cheap quality checks.

mySociety's APPG scraper seems to sit in this box. The goal is to get accurate lists of parliamentary group memberships from dozens of different websites. Our original thought was that this was a significant enough piece of work to be a crowdsourcing problem. When first considering an LLM approach, we thought it would be vulnerable to inventing lists of MPs because these are public figures, likely to be in the training data, and an LLM could conceivably invent a list of MPs. Our approach to this was to invest time in the verification step, where we could verify with code if the names extracted were actually listed. We can see another example of this in Pol.is - where category descriptions are linked back to concrete sources to facilitate easier double checking.

Through a sequence of different checks on LLM output, we can move problems towards being strong uses of LLMs, because they dramatically reduce the time needed to produce the solution, while the verification costs are manageable.

Evaluating the right model for a project

Having explored whether the problem is the right kind of problem for an LLM approach, the next question is which model is, practically speaking, the best fit. Here there are trade-offs between both the openness of the model and the size of the model.

Open vs closed models

Most discussion and usage of LLMs is focused on high profile closed models such as OpenAI's ChatGPT family, and Google's Gemini, which are widely available and integrated into a range of existing products and services.

Because these are closed models, access and hosting of the models is controlled by the companies that create them. This presents a dilemma for civic tech organisations who believe in open source, as important parts of the processes guiding the LLMs can be inaccessible, hidden within black boxes beyond their control. These may work well and be affordable today, but they create new risks. Specific models might become unavailable; there might be changes in pricing; and the lack of transparency in how they work — and therefore whether the function can be replicated elsewhere — creates lock-in to specific providers.

Open LLM models provide an alternative approach. In an issue familiar from open source licensing, there are different ways in which a model can be ‘open'. Open weights models have the final structure released and can be run on your own hardware (Meta's Llama model is an example of this). Fully open models have the underlying (openly licenced) training data released, as well as the recipes and evaluation systems used in their training. AI2's OLMo family of models and the recent Swiss AI institute's Apertus model are examples of these. Somewhere in between these are approaches like IBM's Granite models, where the model is released as open weights and the data was licensed so it could be trained on (addressing copyright issues) but is not publicly accessible.

Any kind of open weights model is practically appealing because they unlock new ways to work with private data without sharing it with third parties, and create more flexibility around infrastructure.

For more thinking about different approaches to Public AI, see Felix Sieker's (Bertelsmann Stiftung) 2025 TICTeC talk on breaking AI monopolies.

Model size

Broadly speaking there is a tradeoff that smaller models (which have fewer parameters) require fewer resources to run (and so can run on cheaper hardware/ more quickly on similar hardware), but are less capable than the larger version of the same model. The extent to which they are less capable varies because there are different approaches used to create smaller models (for instance, training a smaller architecture on the same data, or using knowledge distillation, where a larger model is used to train a smaller one). So whilst smaller models are less capable, they might be capable enough for a specific project for a trade-off in requirements or cost.

Evals as tests

The first practical stages in exploring model evaluation fit is to run sample sets of data and prompts against a few different models (or sizes) for a sense of the performance, and make basic adjustments to the prompt. Using a model agnostic framework like PydanticAI is helpful here - because it makes it easier to swap different models in rather than being locked into the API of a specific ecosystem.

The next step in formalising this is to create a testing/evaluation approach (for LLMs, generally called ‘evals') that tests a set of prompts and inputs, against expected results. This is useful not just in comparing models, but in creating a documented set of expectations that future maintainers can use to check that future changes to the model or prompt haven't broken a previously functional approach (a process known as regression testing). There are different practical approaches to this, but the PydanticAI framework has some built in helper functions (and the guidance is useful on the underlying concepts).

Testing approaches include:

• Deterministic Validation - a fairly traditional test that looks for exact content to be present (the exact same input should result in the expected output). This might be a useful approach for prompts where the goal is to extract structured data, checking whether output of the model matches expected extracted information.
• LLM-as-a-Judge Evaluation - this is for functions where part of the purpose is creating new content (for summaries, etc). Here, the LLM is evaluating qualities in the output that are harder to express as normal tests (e.g. accuracy) and the approach uses another LLM process to do a "word calculator" evaluation. For this, calls to more capable (and potentially more expensive) models might be useful in validating the performance of smaller models.
• Performance Testing - an important part of understanding the viability of an approach is evaluating how long a model takes to return a result. Newer ‘thinking' models especially might take longer to create the intervening steps and it is worth reviewing whether this is actually adding value to the output.

Once these tests are established, they serve the normal purpose of tests in allowing changes and optimisations to be made without introducing regressions. These kinds of tests also help explore whether switching to other models would have an impact on performance.

However, it's also worth noting that we would not always expect smaller models to perform as well as large ones, on exactly the same prompt. To get more effectiveness with a smaller model might require the prompt to give more examples of good practice than a larger one does. Small variations in prompts can lead to big variations in the results. At this point, it makes less sense to think of the prompt as "code" running on the model's "computer"; rather the prompt is a way of priming the model to act in the desired way. Different models may not react to this priming in the same way.

Evals as scrutiny

A key feature of LLM AI technologies is that they have a black box at their core, where it is difficult to clearly audit the processes they follow. This means the only way to understand their outputs is to examine them in systematic ways.

Separate, but related to the question of "how can we evaluate the impact of our own technology", is the evaluation of the impact of other AI technologies, especially as made available to the public, and so making it possible for third-party audits.

Here, creating evals works as a form of scrutiny and democratic activity in its own right: creating benchmarks that can be used to compare models to each other, but also to validate whether they perform well on the kind of democratic questions that users might reasonably ask. For example, there have been a few projects creating evals for chatbots in election contexts (see studies for Swiss, Bavarian and Hessian Elections, UK Elections and US elections). Production of evals can be a way for pro-democracy tech to evaluate current performance, and make arguments for better incorporation of democratic values into the core of these models.

At TICTeC 2025, we heard from Matt Stempeck, Eloise Gabadou and Yung-Shuan Wu about community-led AI audits, where the focus can be on the whole output, and building evidence that helps people affected by AI decision-making to understand, and organise around, systems making decisions about their lives. Audits help evaluate real-world outcomes and create an entry point for organising, policy pressure, or litigation.

Understanding approaches to auditing outcomes in practice are an important part of both building our own AI systems, and scrutinising the impacts of others.

Conclusion

AI technologies present a new set of useful and practical tools that can be (and already are being) applied to a range of pro-democratic applications.

But that said, "AI" as a phenomenon (and the huge amount of capital behind it) is both directly being applied to anti-democratic ends, and is more generally creating AI slop and deepfakes that undermine democratic society.

As such, we need to approach it with caution, and understand where we can make use of the tools, while trying to avoid dependence on approaches that may be unsustainable in the long run.

Sustainable tech: balancing ease of development with user experience

This report has often talked about "technology" in an abstract sense, but effective pro-democracy tech also needs an engagement with software development as a craft.

Running software projects can have enormous impact, but also requires consideration of the skills needed and maintenance required. At aTICTeC 2021 event, MIT GOV/LAB presented advice on when not to build technological approaches, and this remains a useful guide in avoiding complexity and looking for the simplest solution, sticking to popular and well established coding languages, which come with a wide pool of possible developer talent.

The goal of this chapter is to build on this advice, and help navigate a tension between the conditions of the people building tools and the conditions of the people using them. Civic tech teams are often operating under tight budgets and limited time, so the sensible default is to reach for the easiest out-of-the-box frameworks and services. Increases in the accessibility of creating code through these frameworks make it easier for teams with more subject matter expertise to create workable tools and approaches. But improvements in developer experience can have a trade-off with user experience. The users these tools are meant to serve are frequently working on older, lower-spec devices and unreliable connections, where those same frameworks can deliver a slow or fragile experience.

This chapter looks at practical ways to manage that trade-off. We discuss the reality of user device constraints: how performance inequality shapes what accessible services look like, and why designing for below-average hardware (and intermittent connectivity) changes the right technical choices from the outset. From there, we look at when lower-level languages can be a pragmatic sustainability tool by reducing infrastructure costs or creating approaches for locked-down or low-resource environments.

Similarly, taking a considered technical approach can help overcome trade-offs in using technologies like LLMs. Rather than treating LLMs as an always-on dependency, we discuss options for smaller models and using LLMs to generate work products (data, rules, or code) that can be run cheaply and maintained conventionally. Finally, we cover a different kind of recent shift: tools like DuckDB and Parquet that make large data analysis possible on modest hardware, reducing the need for infrastructure requirements for certain kinds of projects.The goal of this chapter is to present practical examples and thinking tools for considering where heavier technological investment might provide worthwhile impacts.

User device resource constraints

Making services work for everyone needs a focus on users with a below average resource device. A key requirement in building pro-democracy tech is understanding the devices that target users have access to. A website may be easy to create using an out-of-the-box framework, while being a struggle to use by the people who would benefit from it.

There is a gulf in computer capacity between developers creating services/web frameworks (with this being especially true for frameworks and services created in better connected countries), and those of their users. As a Paradigm HQ article put it: "The dominance of the Global North, reflected in higher Internet penetration and greater political and economic power, also means that digital products are still largely shaped by the priorities of wealthier nations." In these circumstances, even frameworks with wider support and developer knowledge might be inappropriate for a huge number of users.

Alex Russell maintains a ‘Performance Inequality Gap' blog series that advises on test conditions to ensure a usable experience for 75% of users. A key point here is the dominance of lower-spec Android phones: "Most phones — indeed, most computers — are 24+ month old Androids." Here the challenge for pro-democratic tech is to make sure there is a clear focus from the design stage in building approaches that work for their intended audience.

This doesn't have to mean old technologies, although ‘older' languages remain reliable options. Systems that support frontline workers need to assume periods without connection and support offline modes — modern cross-browser support for Service Workers that make it easier for web applications to continue to offer some functions when offline, avoiding the older need for separate applications for different platforms.

Use of lower-level languages

Programming languages operate on a rough spectrum between "good at instructing the computer what to do" versus "straightforward for humans to operate", where languages that are more human-friendly will often be making available a set of functions and libraries that have been written in "lower-level" languages in a simpler way.

There are a variety of reasons why these lower-level languages are often less appealing to work in (such as being more challenging to write code in, or there being less widespread developer knowledge). This section explores when specific requirements of the project suggest exploring whether a lower-level language such as Rust or Go can play a significant role in a successful project.

Reduce infrastructure costs by putting the work in upfront

Writing applications in lower-level or specialist languages can deliver huge improvements in performance, either reducing running costs or enabling high-performance software to run well on older hardware.

For instance, Minha Receita is a project created by Eduardo Cuducos that demonstrates an approach to creating high-traffic digital infrastructure at low cost. It is an API that provides access to information about Brazilian companies based on their unique identifier.

Brazilian Access to Information law requires that company information (including the unique CNPJ identification number) be made public in a machine readable format. However, the relevant information for a single company is published behind a CAPTCHA or scattered across separate CSV files, making it complicated to gather all information that is relevant to a single company. Minha Receita brings that information together and republishes it in a single, publicly accessible, API. This serves an average of 100-120 million requests per month, while costing only $30 a month to run.

The design goal of the service was to serve a huge amount of information very cheaply: as such, Cuducos learned and used Go to write the service. Go is a programming language created by Google, designed for efficiently serving web requests. This keeps the project a sustainable one-person project while enabling it to have a huge impact.

High security needs

A key feature of Rust is a focus on high performance while architecturally guaranteeing "memory safety" (how programs store and retrieve information while they are running). Memory-related vulnerabilities are a huge source of security problems (where a user can exploit errors to access or modify information they shouldn't have access to).

This has made Rust increasingly appealing as a way to create low level infrastructure and tools that are less likely to contain unknown vulnerabilities (increasing security, managing long term maintenance costs), but also has appeal for applications where user privacy is a crucial part of the setup.

CoverDrop is an approach to whistleblowing designed by a group of academics and then implemented together with UK newspaper the Guardian. It takes advantage of the widespread availability of the Guardian app to create a secure messaging system to allow people to whistleblow to journalists. All installations of the app are regularly sending "cover" messages (encrypted packets that contain no content), meaning that there is no pattern of metadata traffic that reveals "real" messages if a network is being monitored.

A lot of the backend of this approach is written in Rust, and a key part of the appeal is the underlying insurance of the memory-safe approach, given that any information sent is, by its nature, confidential, with big risks to whistleblowers if the system is purposely or accidentally compromised. As Daniel Hugenroth put it in their RustConf presentation :"It's all about making sure we can all sleep at night while working on such a sensitive system".

But memory-safety alone does not guarantee security, and an important part of a decision to introduce Rust is how maintainable it is by current and future developers. In their presentation the team also talked through decisions made in using the first Rust based approach internally at the Guardian.

Software for low-resource/locked down computers

Writing software in a lower-level language has higher set-up costs, but can have significant performance benefits for users. Because these languages produce compiled apps, they are exceptionally fast and can run on lower resources compared to applications that are running on top of frameworks. This is especially useful when the target audience does not have access to the latest devices —and government workers are often a key example of this.

Trafficking Free Tomorrow publishes Rust-based forensics software FolSum, to more effectively process and move collections of evidence and simplify the process of preparing evidence for prosecution of human trafficking cases. Here, the appeal of using Rust is to "empower investigators and prosecutors to do more with the machines that they already have". The compiled application runs well within restrictions that are common in law-enforcement settings: highly locked-down, low-resource, thin client machines. Fast start-up and good performance also help win over target users to the usefulness of the software.

Translating key computationally expensive functions

There are benefits of dropping into a lower-level language without the full software needing to be written in that language. Big increases in performance can be found by identifying slow processing steps and moving them into a faster language, and this can balance less developer knowledge of these languages against a focus on small but important calculations that rarely need to be changed. This is the approach taken by the pydantic data validation library, which wraps a Python interface around core validation and transformation functions in Rust.

For a civic tech example, mySociety is experimenting with where classification rules written by LLMs can be run very quickly by moving the rule processor into Rust. In this case, LLMs are also smoothing the process of writing in Rust by transposing a function initially written in Python (as this is compiled to a Python package, the test suite can itself be ported over from the Python version).

This introduces new complexities for future developers, and needs careful consideration weighing up the benefits of a speed increase against obstacles to maintenance.

Low-resource use of LLMs

In other chapters we have discussed general uses of LLMs in different parts of pro-democratic tech. These new technologies can create new infrastructure requirements and dependencies on external services with new costs.

Here we want to explore options for either lower ongoing infrastructure costs, or moving towards a model where more expensive up-front work (which might be done through temporary support from volunteer or grant-funded data scientists) results in cheaper and more maintainable work products.

Smaller models

One way of reducing either financial or infrastructure requirements is to use smaller LLM models. Many models produce variants with smaller numbers of parameters that are either more straightforward to run on consumer-level hardware, or cheaper to run on remote infrastructure. Smaller models are less capable, but still objectively open up a new set of capabilities compared to what was possible five years ago.

Performing evaluations on larger and smaller models helps you to understand whether the difference in capacity is a practical problem. Especially in non-English speaking contexts, smaller models perform less well multilingually, with the 70 billion parameter version of Meta's Llama3 LLM scoring 78.2% on a multilingual evaluation, compared to 58.6% for the 8b equivalent. Depending on the purpose, this might be sufficient to be useful, or might mean the approach is not viable.

To make the most of smaller models, fine tuning on a dataset based around the task can make the most of the reduced "general-purpose" nature of a smaller model. This is the approach mySociety has taken to fine-tune Llama to help detect personal immigration data amongst FOI requests, while having a model small enough to run on our normal infrastructure. This is a bigger initial investment, but reduces long term complexity. If a project is in a position where someone can provide an injection of expertise and time, this might be a rewarding approach.

Products of LLMs

Another approach is to use LLMs to initially generate code or content that can be used in a cheaper ongoing way. As part of mySociety's parliamentary monitoring, we've experimented with three approaches: generating data, rules, and code as ways of using LLMs at the most cost-effective point in the process.

• Generating data: To add search suggestions without fully integrating a vector database, we used an offline vector database to create lookups for common searches.
• Generating rules: Regular expressions are a fast and flexible way of classifying data; we've used LLMs to generate Regex rules that can then be applied quickly to large datasets.
• Generating code: For our APPG scrapers, we've experimented with interacting with these mostly via code agents, specifying basic data structures but letting the LLM create scraper functions that turned the HTML source into the data structure we wanted.

Here, the goal is to do more work/data crunching up front, but in a way that avoids introducing new dependencies into the ongoing infrastructure.

Wait and things will get easier

A very viable approach, if LLM approaches seem at the edge of an affordable or sustainable approach, is to wait a little while and see if things get easier.

General improvements in models (and especially availability of performance of smaller open and open-weights models) mean it is likely that new models that perform better on your use case will be available in the next year. Investment in a process of evaluation for whether current models meet your needs isn't wasted if none currently do; being able to rerun evaluations on new models in six months' time might reveal a useful approach.

At the other end, there is greater and greater integration of LLM-derived approaches into mainstream packages. For instance, while there is a market for ‘vector stores' that store embeddings for semantic search, there is also increasing support for vector indexes in normal database products like Postgres. While the cutting edge requires lots of knowledge of specialist tools and approaches, over time better support becomes available in frameworks that are already being used. Keeping a clear eye on where LLMs have useful applications and can be applied in a sustainable way makes it easier to understand when might be a valid time to move from experiments to greater investment in tools and approaches.

Improved analytics approaches - DuckDB

While LLMs create a need for additional and more heavyweight infrastructure, other developments have reduced the complexity of the stack needed for creating and sharing data analysis.

DuckDB is a library that enables analytics queries to run directly on large data files. The practical result of this is that if you have a CSV file, you can easily run fast and efficient SQL queries on the contents of that CSV without needing to load it into a database first. This makes analysis and data processing pipelines much easier because you do not need to have a dedicated database set up to process even quite large amounts of data.

The DuckDB argument is essentially that the era of "big data" cloud services was a step in the wrong direction, and even very large data analysis jobs can be done on relatively modest hardware. The DuckDB team makes this point by demonstrating that a 2012 Macbook can effectively run the benchmark queries in reasonable time over a 256GB database.

The utility of this for pro democratic tech is that it makes sharing, analysing and transforming large datasets much more straightforward. This is useful both in services themselves, and in making data available for reuse and wider impact.

DuckDB/Parquet API Pattern

DuckDB's support for parquet data files is extremely useful. Internally, parquet files have compressed column data to reduce file size, but these sit alongside a set of summary statistics for each chunk. This means that the file on the server is often quite small, but even when it's not, DuckDB can access sections of the file to test for and retrieve relevant sections. Using a combination of parquet files and DuckDB, bulk data APIs can be made available at very low costs, with less technical setup involved.

A DuckDB + Parquet approach is where the data is cheaply stored and accessed, and the underlying datatables are available to the end user, enabling users to just run their own SQL queries on the data.

So for instance, while an old approach to making a database available might be to gzip a database dump (which requires unzipping the database before loading it into a local database, before querying it), an approach in this system is to make database tables available as parquet files. These can then be queried directly from the URL, where the DuckDB engine might make a series of small queries through HTTP Range request to understand which chunks are likely to contain relevant information. The efficiency gain here is that the smart aspect is in the client: the server itself does not need to run an interface to narrow down the data; it just needs to host the file (which is much cheaper).

Very large datasets (hundreds of billions of records)can be split into multiple parquet files, with a DuckDB database used to create a reference: this is the pattern used by the opentimes.org travel times website.

Application in parliamentary monitoring: TheyWorkForYou Votes

mySociety's TheyWorkForYou Votes is a fairly normal public-facing web app and database. But behind the scenes, most of the calculations are done using DuckDB and are exported in ways that make follow-up calculations even easier.

As DuckDB is much faster than PostgresSQL at analytics queries, when loading data, queries are run on a series of parquet files using DuckDB, and the final results are then loaded into the database. While it used to take 20-30 minutes to calculate voting summary comparisons (where the record of one representative is compared to the votes of other representatives of the same party), we can now run this for each representative, over roughly twice as many votes, for four different time periods, in about two minutes. This gives a huge amount more flexibility in testing when loading new voting information.

This approach of using analytics queries to transform data also makes elements of development and testing more straightforward. Development environments can quickly be created and populated with a full set of data, and our automated test suite tests fully populating the database from scratch.

Analytics applications

The big benefit of DuckDB is reducing the number of steps and hardware required to access data and do useful analysis. This can both be used directly, or wrapped in a library for a more specific dataset:

• Open Ownership have created a notebook to demonstrate how DuckDB can be used to interact with Beneficial Ownership data standards.
• CovidMXis an R library to analyse Covid-19 data in Mexico, built over DuckDB to quickly access and analyse data.
• Geocodebr uses DuckDB and arrow tables to create a fast, free R-based geocoder, powering research tasks looking at Brazilian administrative records.

DuckDB can also be extended. In the Generative AI Newsroom series, Nick Hager explores how DuckDB can be extended with custom Python functions, which includes being able to run text in columns against local LLMs, helping with categorisation, processing and summarisation of data.

Conclusion

Pro-democracy tech is created and used under financial and technical constraints. The approaches we take to build sustainable approaches will vary a lot depending on the situation. Sometimes sustainability means writing in the same language all our developers already know, sometimes it means "I'm going to be supporting this myself, and it needs to be as cheap as possible to run to have as much impact as possible".

But it's important to remember that not having resources (and building for the widest group of users) often means needing to be more innovative: knowing where the beaten track is not sufficient, how to pick and choose from new approaches to get the most benefit with least cost, and where extra work and thought needs to be put in to best achieve the goals of the project.

Acknowledgements

This report only represents a fraction of the work and expertise of those who came and contributed to TICTeC conferences, seminars and conversations through this current programme. 352 organisations from 63 countries have attended TICTeC activities over the last few years - sharing their work, code and inspiration with a global audience. One of the best things about TICTeC is the sense that, even if working in different places doing different things, we are part of a wider community trying to make things better.

We're grateful to the National Endowment for Democracy for funding this programme over the last few years and to Porticus for additional funding for the 2025 conference.

TICTeC Attendees - Mechelen 2025

TICTeC Attendees - London 2024