In August 2025, Policy Exchange published a somewhat controversial report titled Citizens’ Assemblies: Less Accurate Than Tossing a Coin?

This report makes several arguments against Citizens’ Assemblies – some of which have some merit, some of which… not so much. I’ve written this piece as a (somewhat) quick exploration of the claims made in the report, so we can shine some light on what’s really going on. Full transparency – as a deliberative democracy nerd, I am definitely biased in favour of citizens’ assemblies, but I do also want to make sure the thing I think is really cool is actually really cool. So in that spirit, I’ve broken this piece up into three sections: What they totally got wrong, What they sort of had a point about, and some extra bits. I’ll be keeping everything fairly high level, and won’t focus on the publisher’s reputation – but you can look that up for yourself if you’re curious.

What they totally got wrong

“Citizens’ Assemblies aren’t accurate”

The titular argument is that citizens’ assemblies, when compared with subsequent participatory processes (eg. Referenda), are poorly correlated with what the public will decide if asked – rendering the whole process inert. This claim has a lot of issues, but I’ll focus on two – the statistical validity of their proof, and the foundational assumption that citizens’ assemblies should mirror a non-deliberating population.

While on the surface this assumption might seem reasonable, the purpose of a citizens’ assembly is not actually to reflect the results of a referendum on the same topic. If they did that, you wouldn’t really need citizens’ assemblies at all (other than if you wanted some cost savings over a referendum). One of the key ingredients of a citizens’ assembly is deliberation; the process by which citizens can learn about a complex topic, discuss it, and reach a consensus. This process should yield a different outcome to a referendum – and by virtue of greater information, the ability to collaborate and compromise, and the greater tendency to consider the collective needs of the affected population. There are deliberative methods where you can directly compare a poll or referendum and a deliberative process (Fishkin et al.), but claiming a citizens’ assembly isn’t ‘accurate’ because it doesn’t reflect a subsequent poll or referendum isn’t just missing the point; it’s flipping the point on its head and spinning it around like a beyblade. Comparing these two processes, even when they’re held in a short space of time on similar or identical questions, is comparing deliberative apples to participatory oranges.

Let’s assume for now, though, that comparing these two approaches works. There’s another massive (or more accurately, far too small) issue to address that undermines the entire statistical backbone of this argument. Policy Exchange cited the OECD deliberative database (2023), which contains data on ~700 Assemblies. Within this database, they found 17 deliberative processes that were followed by a referendum. This is simply too small a sample size to draw any meaningful conclusion from. Not only that, but OECD has made their own analysis, and found the recommendations made by a well-run Citizens’ Assembly are robust – so much so that they recommend for both their continued use and their institutionalisation. 

It gets worse still – if we combine the above argument on ‘apples and oranges’ with our case on the statistical challenges, there are 2 data points remaining that are worth paying attention to; those involving Citizen Initiative Reviews. These processes are designed to integrate deliberative and participatory methods directly, allowing for the direct comparison that Policy Exchange is looking to do. But we then turn this not into a review of citizens’ assemblies, but of CiR’s – and 2 data points is an even smaller dataset than 17. In Policy Exchange’s own report, they mention that one of these had amazing results, with the CiR intervention seeing a huge impact on undecided voters. 

What they didn’t totally get wrong

“Citizens’ Assemblies lead to overly progressive outcomes”

Policy Exchange claims that of the 17 analysed Assemblies, 15 yielded more progressive results than what the general public voted for when faced with the same choice. Putting aside the subjectivity of such a determination, there is a somewhat legitimate argument to be explored here. We’ve actually seen similar results in our own backyard. Our friends at Koi Tu / Complex Conversations ran a deliberative forum on the future of transport for Auckland in 2023 – in their thoroughness and zeal for data, they ran pre- and post-intervention surveys on matters relating to both the understanding of the topic at hand, and participant’s opinions on possible changes that could solve those challenges.

Participants grew significantly in their support for what might be considered ‘progressive’ actions; providing a safe and connected bike network, building more housing close to the city – and favourability fell for what might be considered ‘conservative’ actions – building more lanes on motorways and main roads. 

There is a definitive trend here – and while there is evidence that the trend is more towards evidence-based conclusions (At least for deliberative polls – Fishkin), more work remains to be done in this space. Policy Exchange hypothesises these outcomes are a result of selection bias, framing effects and other methodological flaws (I’ll get to those in a bit). My pet theory is two-fold: I think there’s some insight to be gleaned in the first survey conducted by Koi Tu; if participants grow in their understanding of certain issues like the impact of transport, they may reconsider what solutions are best-placed to address those issues. I also think it’s worth separating the notions of ‘progressive’ and ‘conservative’, and ‘collective’ and ‘individualist’. These concepts may well correlate strongly with each other, but if we look instead at whether the results of deliberation trend towards ‘collectivist’ outcomes instead of ‘progressive’ outcomes, we both remove the political bias from the equation, and can now examine this in a way that, to me, can have more psychological rigor applied to it. When people get to know each other, work on a problem together, engage in challenging conversations in a facilitated environment – they form a shared group identity. That identity will impact their decision-making – and if that group is reflective of the wider population, it makes sense that their decisions will be more considerate of that broader collective. 

“CA’s are unavoidably vulnerable to bias and methodological flaws”

Selection bias, framing effects, and polarisation effects are all cited by Policy Exchange as inherent flaws – potentially lethal ones – that are baked into citizens’ assemblies. They’re not. But they are valid risks that are worth acknowledging and working to prevent – and deliberative practitioners put a lot of work into ensuring they are prevented.

Let’s take selection bias for example; the potential for an assembly to be warped by picking people who will be biassed to a certain outcome. Policy Exchange attributes this as the reason so many of the examples they looked at were, by their measure, more progressive than the broader population in their recommendations. They theorised that progressives are more involved in civics (this is true), and therefore more likely to volunteer for something like a citizens’ assembly. This might have merit in a poorly run process, but current best practice would massively reduce the risk of this happening. Practitioners use sortition to ensure participants are randomly selected while remaining representative of the wider population across a given range of factors – gender, ethnicity, age, socio-economic status, where they live – and more. Some also ensure representation across a range of political views, but there’s divided opinion on the efficacy of this. 

Practitioners also put a lot of work into ensuring historically marginalised populations are in the room – something the Porirua Climate Assembly recently worked especially hard to do. Participants are usually paid a daily stipend to ensure low-income earners can join in, and childcare is often supported throughout the process. So much work goes into preventing selection bias – and the crazy part of all of that is that referendums and elections have selection bias too – but with referenda and elections, there’s virtually nothing you can do about it – the disenfranchised stay disenfranchised. 

Now on to framing effects – this is also something that can be massively curtailed with effective methodology. Framing effects – where people can make different decisions depending on how information is presented – can show up in many ways throughout an Assembly. The way a question is phrased can have a huge influence on the rest of the process. How facilitators structure a session can subconsciously influence peoples’ thinking. The experts that are selected can leave people totally blind to entire schools of thought. Framing effects are really scary for practitioners – but there are many ways to neutralise the impact. Across the board, sunlight is the best medicine. If the question posed to the assembly is totally public, and the group who decides what the question is has been selected for diversity and has stakeholders from every side of the problem at hand, the risk goes way down. Similarly, some of the most effective Citizens’ Assemblies ever were livestreamed throughout most of their run time. If you want to lower the framing effect in the selection of experts – simply let participants pick their own alongside the initial selection – and make sure they know they can do that at any time. 

The polarisation effect – where individuals can adopt more extreme positions through group discussion – is honestly not something I’ve seen much genuine concern about in the deliberative field. It’s certainly something to be conscious of, but can be almost completely solved by being conscious of how you group participants, and ensuring everyone has their voices heard. There’s far more evidence that Assemblies lower affective polarisation, and increase social cohesion and cross-group understanding.

“CA’s don’t lead to policy impact”

Some extra bits

Citizens’ Assemblies aren’t a Panacea

I think a lot of the criticism levied at assemblies is because people think of them as being a replacement for referenda or even elections – and that the advocates of such methods would have officials use assemblies to solve every policy under the sun. 

This is not the case – assemblies are suitable for a narrow band of challenges; complex challenges that require trade-offs, matters of social or ethical contention, issues that require strong public buy-in, and topics that have proven intractable when left with elected representatives are all great traits for a successful assembly topic. There are instances of assemblies being asked the wrong question, and we see the results in the output – for example, Scotland held an Assembly asking “What kind of country are we seeking to build?”

This is way too high-level a question – the scope is virtually unlimited! Similarly – when a decision needs to be made quickly, deliberation is not the right path. No one wants to be in an empowering and engaging conversation about how to best escape a burning house – at least not as it’s falling down around you.

Citizens’ Assemblies aren’t a single ‘thing’

Policy Exchange makes a few comments in the report that imply a very singular definition of citizens’ assembly – which is funny because the sample of 17 they worked with included a wide range of deliberative approaches. But a citizens’ assembly is not so rigidly defined – you need certain ingredients, but the recipe can be tweaked to suit the chef. Assemblies, and many other deliberative processes, are adaptive – they can and should be tweaked and experimented with to fit the environment they find themselves in.

Closing Thoughts

Wrapping this whole thing up – there are valid concerns when it comes to implementing citizens’ assemblies at scale. We should be looking at them critically, especially with regards to ensuring they deliver good policy outcomes over the long term. We need more literature in this space to have any reasonable sense of the true potential (or lack thereof) of citizens’ assemblies – and like all things in Political Science, confounding variables are abundant. But it pays to be wary of any reports that claim they aren’t accurate, or that they don’t work – because there’s a lot more data in favour of them than not. If you’d like to explore this topic further, you can become a member – and feel free to reach out to me directly on whatever platform works best for you.