Anchit Som

Government is facing a problem: AI is making it easier and easier for citizens to get stuff done...and it turns out government was never really designed for it.

It was only a decade ago(ISH) that we all hoped government could be made transparent and responsive through civic technology, hence freedom of information act and sites like What Do They Know.

But today, when anybody can ask ChatGpt to draft and comprehensive request, government departments find themselves sinking under demand, in an AI arms war.

It turns out friction is kind of important! Richard Thaler describes this as "sludge": friction on the system that's seemingly tiny, but enough to affect behaviour.

A world where we all have an AI agent might be the death of friction, and the trend of the last few decades which sees democratic governments being ever more transparent and accessible... But somehow less trusted than ever.

Let's have a bit of a discussion around efforts to make government more transparent, how that didn't seem to be working, and how just maybe we can design some intentional sludge to fix it?

Asymmetric information arises when one party knows more than another. Understanding the implications of this will help you understand everything from why private health insurance is difficult to why socialising with new people is hard.

Asymmetric information occurs in a range of contexts:

• the buyer doesn't know if your car has a hidden fault
• the health insurer doesn't know if you have a hidden condition
• your car insurer doesn't know if you bother locking your car
• your boss doesn't know if you're working or scrolling Tiktok
• you don't know whether or not your new acquaintance is a weirdo

The first three are hidden facts. The second two are hidden choices. Economists study these situations because they can lead to markets unraveling and collapsing. There are also a range of ways to mitigate them including: warranties, references, qualifications, reviews, medical questionnaires, insurance excesses.

For a seminal paper on this issue, see: The Market for "Lemons": Quality Uncertainty and the Market Mechanism (1970)

Some time ago, I started building an AI negotiation coach. A negotiation coach is someone who helps a person prep for a negotiation via the sokratic method and, on a limited basis, expert advice. It follows the principle "I am an expert in negotiation, but you are an expertin YOUR negotiation."

I never finished this, as detailed below. I offer this project up to anyone who is interested in building it, or some version of it, or building a completely different thing inspired by it. I can tell you how to do negotiation coaching, and advise you on the lessons I learned so far about building with LLMs.

I abandoned this project after realizing two things:

1. Building things is not a priority for me (The hard part isn't building, it's testing and quality control!)
2. I was not entirely sure how to do negotiation coaching

Since then, two things have happened

1. I have begun doing negotiation coaching, and developed a practice
2. AI has gotten a lot better

The URL below shows the half-baked app. It shows the basic logic and flow, but it uses old models, has no quality control and I honestly don't even remember what process it's attempting to follow.

https://planning.negotiation.solutions/

Software development is necessarily social.

Developers need to communicate well for plenty of reasons:

• to share an understanding of common goals and desirable outcomes

• to share standards for quality

• to share contracts and interfaces for interoperability

• to share patterns for efficiency and readability

• to share learnings to avoid common pitfalls

• to write clear code that’s easy to review

Over the past few decades a variety of ways of working and digital tools have been developed to support this. The agile manifesto and agile software development offers principles and ways of working for teams, and  lean software development goes further, applying lessons from manufacturing to support larger organisations and optimise cross-team communication. Tools like Jira, Trello, Asana, GitHub Projects (and plenty of others) offer easily recognised interfaces to manage this communication.

Open source software is an ecosystem of interoperable tools including operating systems, command line tools, browsers, coding languages and libraries. Many interconnected systems are linked together by APIs running common protocols, with common data standards.

AI assisted coding is becoming prevalent, and may alter the way we work together. Developers can create far more complex projects in much smaller teams, or in isolation. Are shared contracts, open APIs, common data types, and interoperability at risk? What are good outcomes, and how can we incentivise them?

For this topic, we’ll discuss and test principles that could contribute to a manifesto for assisted coding. We’ll be thinking about this, and exploring a variety of ways of working. All ideas welcome!

No, I don't have the answers of how to completely successfully go against the enormous power of Big Tech companies. But I can talk about the various types of efforts being deployed, some case studies (I list a few below, largely as examples), and where they might work together - including internationally, and by bringing in parties who do not normally focus on political technology.

• Investigative Journalism: From the Facebook Papers to 404 Media
• Political stands: From social media age bans to Palantir.
• Public pressure: From Cambridge Analytica to the Anti Data Centre Movement
• Regulation: From antitrust to tech regulation
• Internal efforts: From Project Maven to Trust & Safety
• Lawsuits: From chatbot harms to social media addiction.
• External Research: From Crowdtangle to the Coalition for Independent Technology Research

I can also talk about some of the disadvantages and discomforts that may be involved, for instance where there is division of what "success" looks like, or trade-offs between (e.g.) completely accurate and truly engaging messaging.

There are numerous opinions and camps around the ethics of various technologies (most strongly and recently, AI), which can create frictions even amongst political technologists who are aligned on other values - let alone people who do not consider themselves political technologists. I can provide advice for Fellows who wish to discuss or navigate polarised views around tech. I will not attempt to convert Fellows to any particular perspective, nor help them to do so with others. I will instead help them consider how they can locate their own work, projects, and views within the existing space; and engage with & prepare for potential criticisms and opposition.

My experience in this topic: Besides my broader work in tech accountability and safety, also at the NGO AlgorithmWatch I was responsible for a lot of work at the intersection of differering views on tech, and a spectrum of organisations who took different approaches and stances. I developed AlgorithmWatch's principles for use of Generative AI, building on surveys which showed diverging views of AI in our organisation. I was also responsible for our research strategy, much of which challenged technologies while also often using (and sometimes encouraged use of) the same technologies.

Machine learning and AI algorithms are used for decisions that impact people's lives, such as assessing insurance risk, credit approval, and screening job applications. And these algorithms can sometimes discriminate against protected groups.

I am particularly concerned about proxy discrimination, where seemingly innocuous characteristics (e.g. postcode, size of car engine, etc) act as stand-ins for protected characteristics (e.g. race, sex). Even when protected characteristics are excluded from models, they can still discriminate via the proxies.

As a professional in economics and causal inference, I break this problem down through the lens of causal analysis. By understanding how inputs causally impact a model's outputs, we can assess if and when discrimination arises.

Further reading:

• Implementing Anti-Discrimination Policies in Statistical Profiling Models (2011)
• Reuters: Insight - Amazon scraps secret AI recruiting tool that showed bias against women (2018)
• Discrimination-Free Insurance Pricing (2021)

Economic theory often seeks to identify a model's "equilibrium" - a stable state towards which the systems trends and then rests. However, many economic, political and social systems have multiple equilibria. This provokes questions: which equilibrium will emerge, what determines that outcome, and how stable is it?

Understanding this illuminates the robustness and fragility of the systems on which we rely. It also explains why seemingly arbitrary features can be highly influential.

For example...

• Why a bank can be healthy one day and collapse with a bank run next.
• How and why power within an organisation can be concentrated on a single individual.
• Why political norms can remain entrenched for decades until suddenly they are not.
• How your life could change from going great to collapsing quicker than you hoped (and, more optimistically, vice versa).

What are the different forms and ways in which the collectives and political activities you work with are imagined? Getting clear about what metaphor is in play can help us stay honest when it comes to generalizations.

Metaphor often guides processes, obscures assumptions, and impregnates practical processes.
Revealing what the imagined pictures being used in politics are can be an excellent first step in figuring out what on earth is going on out there.

Most good work dies in an inbox. People spend weeks on a project, a proposal or an idea, and then thirty seconds on the message that's supposed to get someone to engage with it. Getting things in front of people is a genuine skill and one you can learn.

I've spent years operating in ecosystems where nobody has to reply to you: MPs, funders, journalists, party staff, senior volunteers, busy organisers. I can help fellows with the practical craft of cut-through: writing emails people actually answer, WhatsApps that get a response rather than a "seen", subject lines, timing, length, asks that are easy to say yes to, and knowing when a message should be a call instead. This also includes the wider game of visibility, how to warm someone up before you ask, when to use an intermediary, how to follow up without being annoying, and how to make your work legible to people with no time.

I can help fellows think about questions like: Why is nobody replying to me? What's the actual ask here? Who should this come from? How do I get my project in front of the person who can act on it?

Political communication increasingly isn’t just about having the strongest message, it’s about understanding how information moves. I’m interested in the relationship between message, messenger, medium and distribution: why some political ideas spread while others disappear, why the same message lands differently depending on who delivers it, and how platforms shape what people see, trust and act on. Most political actors obsess over the first and neglect the other three which is usually why their ideas disappear.

Some things I've learned from testing this in practice:

• Messenger often beats message. In our crowdsourced message testing, the most persuasive messages came from people who shared demographics and political history with the audience, not from professional comms people. In our pilot work distributing broadcast clips into Reform-adjacent spaces, a named, working-class trade unionist's account outperformed an anonymous "in-group" account on every metric — roughly 5x the views — while avoiding the platform bans that killed the anonymous ones. Authenticity is a distribution strategy, not just a virtue.
• Distribution Mechanics that grab attention are different depending on what platform you are on. If you want something to grab attention on TikTok you might make a very short video with a shocking hook. If you want to grab attention in a whatsapp group locally, asking a question and telling a "gossip like" story is way more effective.
• Proximity drives salience. Ads featuring local candidates consistently beat national politicians; in one pilot, everyone who scanned a QR code on a leaflet lived within 1km of the local issue it discussed. People pay attention to politics that is recognisably theirs.

Through my work across campaigns, organising, experimentation and political technology, I’ve explored how political attention forms in practice , from local Facebook groups and messaging apps to campaign content, community networks and conversational spaces. I can help fellows think about questions like: How do people actually encounter politics today? When does messenger matter more than message? How do platform incentives shape behaviour? Why do some issues suddenly become politically salient? How do trust and legitimacy get built online? And how do you design political interventions that create participation rather than simply visibility?

I’m particularly interested in moving beyond broadcast models of politics towards understanding politics as something relational and distributed, built through networks, conversations, repeated interactions and trusted messengers rather than centralised communication alone.

A lot of people understand UK politics formally but not operationally. I can help fellows understand where decisions actually get made: the relationship between parties, MPs, special advisers, local government, campaigns, funders, members, parliamentary structures and informal networks.

My background includes working in parliament, running Labour Together, convening election review processes and spending years operating between campaigns, think tanks and politicians. That means I can help translate abstract questions (“How do you influence Labour?” “Who matters?” “What does Parliament actually do?”) into practical strategy. I’ve run sessions on navigating UK political systems, identifying pressure points and understanding the difference between formal constitutional power and real-world political leverage.

Politics rarely works the way the diagram says it does. A lot of strategy is understanding timing, relationships, legitimacy, incentives and institutional rhythms.

One of the things I spend a lot of time helping people with is moving from “this is unfair” to “this is changeable”. Political systems often look impossible until you identify where decisions, incentives and coordination failures actually sit.

I can help fellows map systems and identify leverage: who controls implementation, who shapes narratives, where informal veto points exist, where legitimacy matters, and where small interventions create outsized effects. This includes practical tools for actor mapping, coalition design, identifying hidden decision-makers, sequencing campaigns and understanding institutional incentives. I’ve run sessions that focus less on ideology and more on operational politics: how things move in practice.

I can help fellows think through different routes to change: insider versus outsider strategies, coalition building, narrative pressure, institutional entrepreneurship, campaigning around moments of opportunity and understanding how influence accumulates. This includes practical questions like: Who needs persuading? What kind of evidence matters? When should you campaign publicly versus privately? How do you build influence without formal authority?

I’m particularly interested in helping people understand that lobbying is rarely just meetings, it’s often about creating conditions where a decision becomes easier to take.

One of the more unusual areas I can support on is understanding MPs’ offices as organisations. Through our research and practical work we’ve looked at how MPs’ offices actually function — workflows, staffing, onboarding, technology, leadership, casework, decision making and institutional constraints. They are the hidden infrastructure of UK democracy: 650 effectively independent micro-organisations, employing over 3,500 staff at a public cost of around £200m a year, and for most citizens the primary point of contact with the state when something goes wrong. Yet they are barely researched, and nobody — not MPs, not Parliament, not the parties — has a clear picture of what a high-performing office would even look like.

Through the Parliament Workflow Study (convened by Campaign Lab and Newspeak House), I co-authored Democracy on Default Settings with Edward Saperia, published by Demos in 2026 and based on fieldwork, interviews and surveys with staffers across all parties. Some of what we found: 73% of staff had no written guidance or handbook for how their office runs; offices recruited for 89 different job titles in a single year; MPs collectively purchased over 275 different software tools since 2020 (including 34 different video editors) while barely 1% claimed for any AI tool; and staff described the job as a "fourth emergency service", absorbing casework that should sit with councils and public services, often without training or mental health support. The pattern throughout is offices running on default settings — personalised, reactive, and dependent on individual heroics rather than institutional design. Ahead of the historic 2024 intake of 335 new MPs, we also produced a comprehensive MP office manual, now hosted on W4MP.

What makes this interesting beyond Parliament is that MPs' offices are a perfect case study in why good ideas fail to get adopted: fragmented authority, restrictive central IT, no shared standards, high turnover destroying institutional memory, and a monopoly casework tool with no competitive pressure to improve. If you can understand how to change an MP's office, you understand something general about institutional reform.

I can help fellows understand questions like: What is an MP's office actually optimised for? How does constituency representation really work, and where are the bottlenecks? Why does £200m of democratic infrastructure run on spreadsheets and WhatsApp? What would better infrastructure look like, and who would have to act to build it? I've run practical sessions on office operations, and this is especially useful for anyone interested in democracy, public institutions, civic products or state capacity.

My starting point is that conflict is not a failure state: conflicts are necessary to politics, inevitable, and in fact creative. Politics is the reconciliation of paradox — it doesn't deny differences, it holds them. And conflict isn't always anger; often the worst kind is stuckness, a cycle of stasis and indecision. Most governance systems were built precisely to get unstuck: the Labour rule book exists because trade unionists and middle-class intellectuals with limited communications had to coordinate through enormous disagreement.

Over the years I've collected a toolkit of approaches beyond voting for resolving conflict and stuckness — getting concrete about interests rather than fighting at the values level; separating the minimal from the substantial; going process by process; thinking creatively about legitimacy beyond formal votes; constructing shared purpose or shared identity; building coalitions of the willing; repairing the relational; spotting when a fight is a proxy for something else; using stories; and, occasionally, harnessing constructive ambiguity. Each has different moral implications and costs, and knowing which to reach for is most of the craft.

I learned much of this running Labour Together through Labour's most divided years. We found format shapes what's sayable: roundtables built around a shared meal let members from opposing factions have conversations a CLP meeting's for-or-against structure makes impossible — the format did the bridging. And shared analysis has to come before shared strategy: the 2019 election review worked because a cross-factional commission and a survey of 11,000+ members meant everyone had co-produced the analysis and no one could dismiss it as spin. Campaign Lab applies the same craft to a different boundary, creating environments where technologists and campaigners can experiment together.

I can help fellows think through questions like: When is disagreement useful? How do you stop groups fragmenting? How do you build consensus without flattening differences? When do you compromise versus hold the line? That includes practical approaches to facilitation, coalition governance, designing decision-making processes, and creating enough trust for experimentation. I've spent a lot of time in spaces where people care deeply, disagree strongly and still need to move together.

I've spent a lot of time trying to understand what politics looks like when you stop starting from Westminster and start from people's everyday lives. Between 2014 and 2024 the proportion of people with low or no trust in MPs rose from 54% to 76% and my working theory is that trust isn't rebuilt through one big act but thousands of small ones, most of them local. People engage with issues they can see from their window: in our QR code leaflet tests, 92% of school-petition scans came from households within 400m of the school, and 72% of shop-consultation scans from within 100m of the proposed site. Campaign Lab research also found links to local newspapers are shared five times more in local Facebook groups than links to national news. The local level is where abstract politics becomes tangible.

Through Campaign Lab we've built infrastructure for working at this level: a register of local Facebook groups by constituency, social listening on community conversations, and analysis of how issues emerge and travel through local digital spaces, from Facebook groups and Nextdoor to community WhatsApps.

The framework we use for turning all this into strategy is Listen → Act → Communicate.

Listen through casework clustering, local groups, local news and petition data to find what's actually salient (distinguishing quick reactive wins (flytipping on George Street) from long-term proactive work (the stalled hospital)).

Act by choosing concrete actions over abstract campaigns, scored against target relevance, resource, achievability and salience.

Communicate by closing the loop with the specific people the action mattered to, through the channels they already use, ideally via endorsers because it's always better when someone else says you're good than when you say it yourself.

The step everyone forgets is the third one: plenty of good local work happens and nobody ever hears about it.

I can help fellows think about questions like: How do people experience politics outside election periods? How do local identities shape issue salience? How do you design interventions that feel genuinely local rather than nationally templated? And how do you build the listening infrastructure (human and technical) that makes any of this systematic rather than accidental?

The central thesis of my book The Majority Myth is that voting is usually a social ritual. If that is at all true, then those of us who wish to foster or impulse collective decision of any kind ought to be in the business of ritual design. Some practical versions of Ritual Aware Design voting systems, such as 3dV Voting, can be seen at www.codynamics.work.

A decision ritual designer takes the view that the process taking place is one of changing how the participants think.  The idea that voters combine their opinions for the purpose of choosing between options is replaced with the idea that the voting acts upon the participants to change their minds and produce acceptance of the result.

This is the main work of our fledgling Collective Dynamics Lab. I can work with you to review the design of your collective participation systems through this lens.

This topic builds directly on Building Evidence: Data Science in the Age of AI and assumes familiarity with the foundations introduced there: how data represent the phenomenon being studied, how samples are constructed, and how patterns can be measured without claiming more than the evidence supports. Social media listening puts those methods to work on a particularly difficult form of political data. It begins with a deceptively simple question: what are people saying?

Before we can answer it, we have to decide which people, which platforms, which posts and which kinds of speech count as data. APIs, commercial datasets, scraping and platform search each produce different windows onto the conversation, shaped by technical constraints and design choices. None of them is the public itself, but each is a partial representation that must be made explicit if we are to treat it as evidence.

Once collected, the data can be filtered, classified and organised into issues, positions, communities, places and changes over time. Network analysis can show how conversations cluster and travel. New language models can process volumes that would previously have been impossible, but sophisticated analysis does not make a biased sample representative.

I have spent years analysing online political behaviour and now build pipelines that turn large collections of posts and media into structured, reviewable evidence. I can help you construct a small listening workflow—from defining a question and collecting data to categorisation, network analysis and a short political brief. The aim is to learn how to turn online conversation into useful intelligence while making its blind spots and uncertainties explicit.

Core reading

• Matthew R. DeVerna et al., “CoVaxxy: A Collection of English-Language Twitter Posts About COVID-19 Vaccines” (2021).
• Deen Freelon et al., “The Post-API Age of Social Media Data Access: Past, Present, and Future” (2025).

You can build AI to support decision-making in a variety of ways. I have tinkered with some of these, and others are untested ideas.

For example..

• AI as negotiation or conflict management coach: one-on-one conversations to help someone plan how to navigate a situation (basically an AI version of me)
• AI as a live conversation mediator (basically an AI version of me doing a different thing)
• AI as asynchronous group decision support. This could mean a lot of things, depending on where the group communication happens and how large the group is.

I am an expert in mediation and negotiation, you are an expert in building things. I would be thrilled to guide you in building cool stuff.

This topic builds directly on Systems Thinking: Understanding Political Dynamics. Once we have a shared language for feedback, stability, adaptation and failure, we can ask what cybernetics adds: how systems use information to regulate themselves when no single actor can understand the whole.

Norbert Wiener developed cybernetics across machines, organisms and societies; Stafford Beer applied it to organisations and government. Their work shares Friedrich Hayek’s starting point that knowledge is dispersed, local and often tacit, but asks whether autonomous parts can still be coordinated through carefully designed feedback, information flows and escalation. I know this territory less well than mathematical modelling and network analysis, so this is genuinely exploratory: we can investigate the tradition together rather than treat it as a settled doctrine.

AI makes the attempt to reinvent cybernetics especially urgent. We can examine or redesign a real institution as an information-processing system, asking what should remain local, central, automated or contestable. The aim is to use cybernetic ideas to evaluate—and begin designing—forms of coordination that enlarge collective intelligence without simply strengthening surveillance and control.

Core reading

• Norbert Wiener, The Human Use of Human Beings: Cybernetics and Society (Houghton Mifflin, 1950).
• Stafford Beer, Designing Freedom (J. Wiley, 1974).

Political systems can be understood through a set of mathematical and mechanistic ideas that help make sense of how they behave over time. Concepts such as homeostasis, autopoiesis, growth, equilibrium, feedback, and phase change give us ways of describing how political processes stabilise, amplify, or transform. Rather than treating politics as purely contingent or narrative-driven, these tools allow us to see recurring structural patterns: how actions generate responses, how those responses feed back into the system, and how certain configurations persist while others break down or reorganise.

Institutions can be understood in similar terms. Some behave in a homeostatic way, actively counteracting disturbances in order to maintain a stable state. Others exhibit growth dynamics, expanding their reach, resources, or influence until they encounter constraints or diminishing returns. In some cases, organisations take on autopoietic characteristics, reproducing the internal rules, roles, and cultures that sustain them, even when their original purpose has weakened or disappeared. These are not just abstract or playful analogies: they point to the fact that political systems have strong quantitative foundations, where votes, resources, and participation are all measurable. Thinking in these terms helps build a more precise understanding of why political systems sometimes appear stable, sometimes become unstable, and sometimes shift abruptly into new forms. In this topic, we will use guided reading and writing to explore how these mechanistic and systems-based ideas can be applied to better understand real political systems.

I have spent much of my career building mathematical and computational models of biological and social systems. We can learn the fundamentals of systems thinking by mapping political systems and experimenting with simple models of their behaviour. The aim is to develop a practical vocabulary for recognising feedback, stability and structural change—and a set of tools that can be applied to political questions.

Core reading

• Donella H. Meadows, Thinking in Systems: A Primer (Chelsea Green, 2008).
• Mitchel Resnick, Turtles, Termites, and Traffic Jams: Explorations in Massively Parallel Microworlds (MIT Press, 1994).