Artificial intelligence (AI) is outpacing the institutions meant to govern it. Officials lack the means to verify industry claims about frontier systems, which have already triggered tens of thousands of security incidents. More extreme scenarios, such as future loss of control, cannot be ruled out. As regulators lag behind, anxious citizens are deeply skeptical that governments can avoid sleepwalking into a crisis.
For now, frontier AI is governed by the companies building it. Recently, however, nearly 1,400 lab employees signed an open letter asking for government intervention to pace the frontier, since no company can slow down alone. Yet the alternative — executives coordinating a slowdown among themselves — smacks of cartel-like behaviour.
A handful of US firms are left to answer questions they cannot and should not answer alone: How much catastrophic risk is acceptable, how fast is too fast and which uses are off-limits? Those judgments need a public mandate, and a standing citizens’ assembly on AI is the best way for a democracy to supply one.
Some democracies are already experimenting with the idea. The European Commission, for example, has run citizens’ panels feeding into its digital strategy. Similar exercises on AI have taken place in Britain and Canada, and major AI labs have even commissioned their own. In June 2026, a University of Edinburgh report urged the UK government to make one permanent.
The instinct is sound, but any standing assembly will be an empty vessel until two things are specified: what it is asked to decide — and what its conclusions oblige anyone to do.
What an Assembly Can and Cannot Do
An assembly cannot determine the technical specifics of a model’s safeguards, or whether critical digital infrastructure could withstand inevitable breaches. Those questions are better settled by evaluations the public cannot run and should not be asked to.
Similarly, an assembly cannot act in real time during a crisis or bargain with rival states, though it can set the terms its government brings to the table. Hélène Landemore and Audrey Tang make a compelling case for a global assembly on the pace of frontier AI, but a body answerable to no government binds only those who volunteer.
Assemblies are also better suited for questions with no technical answer, such as how much catastrophic risk a society will accept, how much labour displacement it is willing to absorb and how much surveillance it will tolerate in the name of security. These issues turn on national laws and social values, and no expert body has the authority to decide them alone. Instead, they belong in a forum where the government is obliged to answer. And the less confidently the odds of harm can be estimated — as with loss of control over increasingly autonomous systems — the more the decision rests on values rather than expertise.
British philosopher Jonathan Birch offers a useful way to draw this line. Writing about decisions taken under deep uncertainty, he argues that judgments about proportionate precaution cannot be delegated to experts alone, because there is no technical or scientific answer to how much caution is warranted. Expertise establishes what is known and what is not — deciding what to do about it, however, is a different kind of question. As such, a standing citizens’ assembly on AI should be informed by experts but not ruled by them.
Birch’s division of labour points to three jobs for a standing assembly on AI.
The first is setting terms. An assembly cannot forecast what AI will do, but it can reasonably state what a society would accept. That could mean drawing red lines around uses deemed off-limits, such as fully automated decisions over welfare eligibility or policing. Or it could mean naming the conditions a system must meet before it goes public, leaving expert evaluators to decide whether it does.
The second is planning for when something goes wrong. A recent Chatham House report mentions that, in an AI crisis, a government could recognize the emergency but lack legal reach to act, while the AI provider holds the technical remedy but not the public mandate.
Take July 2026’s breach of Hugging Face by OpenAI agents that escaped containment during a security test. OpenAI alone decided the response, pausing some training and tightening monitoring. Weeks later, the United Kingdom’s AI Security Institute found that the company’s newly released GPT-6 Astra carried out unsanctioned supply-chain attacks in 29 percent of simulated tests. An assembly could help establish protocols and transparency requirements for when similar incidents occur again — as they most certainly will.
The third job is review. A standing body can reconvene to ask if officials acted within the limits that citizens recommended — or if they publicly justified departing from them — and revisit those limits as more evidence emerges. Importantly, this is what separates a standing assembly from another consultation, and it disciplines everything upstream.
Deliberation Without Consequence
France offers a case study in how these processes might unravel.
In 2019, President Emmanuel Macron’s government randomly selected 150 citizens to directly advise the French Parliament on climate actions that would not unfairly burden regular people. The decision came after months of unrelenting violent revolt against fuel-price hikes by the grassroots gilets jaunes (yellow vest) movement. Macron promised the recommendations would sail through unaltered to a referendum, parliamentary debate or adoption by presidential decree.
Participants produced 149 proposals to slash carbon emissions 40 percent below 1990 levels by 2030 in an equitable way. But Macron wavered. During a speech to conclude the process, he promptly vetoed three key ideas, including a tax on corporate dividends to fund green projects. Cabinet then systematically weakened proposals as they encountered them.
A compelling contrast emerged that same year in eastern Belgium.
The Ostbelgien model consists of a permanent 24-member citizen council that creates temporary citizen panels to tackle specific policy issues. When panels agree on a set of recommendations, it automatically triggers a public hearing with a relevant parliamentary committee. Law makers from the regional Parliament must also hold two meetings to discuss the non-binding proposals within one year of their submission. This approach essentially puts the state on the clock to publicly justify accepting or rejecting citizens’ ideas. In its first five years, the model ushered in meaningful policy changes related to housing and health care. Some of the biggest gains have included new funding to draw young people into nursing, giving residents’ families a seat on the boards of assisted-living facilities, and a recent ban on smartphones in schools.
For assemblies to work, they need clear rules for what happens after deliberation. An assembly should not replace a legislature, but its recommendations should oblige one to respond publicly. Done well, these processes improve decisions and boost civic pride; done poorly, they erode the very trust they were meant to rebuild. This is key in the context of AI regulation, as technological disruption is already producing new strands of populism in democratic countries — most evidenced by the visceral opposition toward data centres.
Harnessing citizen assemblies in good faith to craft AI regulation could help governments respond with agility to a deeply divisive issue. That said, it would not be bulletproof. Participant selection methods, no matter how deliberate, still routinely suffer from a lack of equal representation. All the potential upside of a citizens’ assembly on AI will never immunize it from populist attacks. Still, this may be better than the status quo, in which the tempo of AI development is set by a handful of firms answerable to no public at all.
Experts can tell us what is known about the risks of powerful AI and what is not, but they cannot tell us what follows. That judgment belongs to citizens, and a standing assembly — with real questions to decide and a government obliged to answer them — is the best tool democracies have for making it.
