Participatory deliberation as policy input on AI

Convene structured, facilitated deliberations among ordinary citizens — random-sampled, regionally representative, sectorally representative — about specific AI policy questions, and feed the deliberated output into regulator, legislative, or platform-governance processes that would otherwise be decided by experts and lobbyists alone. The form is a citizens' assembly, a vTaiwan-style online consensus platform, or a community-rooted alignment assembly; the discipline is procedural rigour that the receiving institution will accept as legitimate.

An actor chooses this strategy because legitimacy is itself a scarce resource in AI governance: regulators and platforms openly say they do not know whom they speak for, and a deliberated public input fills that void in a way no NGO position paper can. Done well it produces a constituency the institution did not have to invent and a record of public reasoning that constrains a future U-turn.

It trades off speed and reproducibility for legitimacy. A good deliberation is expensive and slow, and its findings tend to be moderate where the underlying harms are not; institutions can also use the form as cover — commission the assembly, accept its findings selectively, and proceed as planned. Without a binding link to the downstream decision the strategy reduces to civic theatre.

Verdict — wins legitimacy, lacks teeth

Participatory deliberation as policy input is the corpus's most distinctive legitimacy generator. No other strategy in the corpus is designed, end-to-end, to produce a record of considered public preference at sufficient scale and demographic rigour to be credible to an institutional partner — that is the form's specific output, and it is the output other strategies cannot generate on their own. The adopter set is correspondingly distinctive: the Collective Intelligence Project (the Alignment Assemblies programme, the Anthropic Collective Constitutional AI assembly, the Global Dialogues Index), g0v (vTaiwan, the longest-running national-scale Pol.is deliberation), the Ada Lovelace Institute (Citizens' Biometrics Council, citizen juries on NHS data sharing, Participatory Data Stewardship), Involve (Climate Assembly UK; the UK's facilitation-design infrastructure layer), the Sortition Foundation (the participant-selection infrastructure layer), and Te Hiku Media (the Kaitiakitanga License as the community-stewardship variant). The four effects: above are typical of what the strategy returns to the movement at its working end — a national regulatory regime issued from a citizen-deliberated consensus, a frontier-AI-lab model actually trained on a publicly-written constitution, a national AI statute negotiated against the public-affairs ground a participatory platform had populated, and a recurring multinational dataset on AI governance preferences released as open-source infrastructure.

The strategy is good as the movement's upstream normative-content engine, bad as a binding instrument on its own.

Strong on its primary claim — that a deliberated public output can become the substantive input a receiving institution reasons with. The 2015 vTaiwan Uber deliberation is the corpus's cleanest case: 4,500 participants converged on five concrete positions and Taiwan's Ministry of Transportation and Communications issued regulations along those lines. By 2018 the platform had run 26 deliberations at a documented ~80 percent rate of producing decisive government action — a rate no other participatory infrastructure in the corpus matches. The Anthropic Collective Constitutional AI experiment is the corpus's first documented case of a participatory output reaching all the way into the model-training layer of a frontier-AI lab: ~1,000 representative Americans produced a 275-statement constitution that Anthropic actually used to train a model and published as joint research. The CIP Global Dialogues Index documents at scale — 6,000 people, 70 countries, 8 languages — that the form can produce a standing public-opinion infrastructure rather than a one-off event, the data released open-source for any actor to build on.

Weaker on the secondary claim — that the deliberation's output produces a binding constraint on the target rather than a reputational one. The Anthropic experiment trained one experimental model on the public constitution; the production Claude continues to ship on its in-house 58-principle constitution, and the public constitution sits in the research record alongside the in-house version rather than replacing it. The OpenAI assembly produced a six-recommendation set and a commitment to "use findings to inform" model evaluations and release criteria — language that contains a verb but commits to no specific behaviour, and no published evidence yet attributes a specific OpenAI release decision to the assembly's findings. The Taiwan AI Basic Act passed at third reading on 23 December 2025 with vTaiwan input documented in the legislative trail, but the seven-principle statutory frame the Act establishes — transparency, privacy, autonomy, fairness, cybersecurity, sustainable development, accountability — is a baseline declaration of principles, not a litigated remedy structure that gives the framing its enforcement teeth. The deliberation shaped the public-affairs ground on which the Act was negotiated; it did not shape the Act's downstream enforcement apparatus, which lives in implementing regulations and judicial proceedings the strategy does not author.

The deepest structural problem is the advisory-output ceiling. Every other strategy in the corpus has a moment of taking — strategic litigation takes a writ; mass protest takes the public square; collective bargaining takes a strike; coalition lobbying takes a legislator's vote; empirical audit takes a data point and an institutional citation. Participatory deliberation has only the moment of giving — a report, a constitution, a 275-statement list, a Pol.is opinion map — that the receiving institution can accept, accept selectively, or commission and ignore. The CIP "committed audience" architecture (each assembly tied to a publicly-committed institutional partner) is the corpus's most sophisticated workaround for the ceiling, and the workaround had to be invented because the form does not generate its own binding force from inside itself. The result is that a 1,000-person citizens' assembly's recommendations are not enforceable on a model the public did not pay to build, does not hold equity in, cannot subpoena, and cannot strike against. The legitimacy the form generates is real and is, on the evidence of the four effects above, often consequential; the form's non-recursive output — the gap between what the public concluded and what the institution then did — is the strategy's distinctive structural exposure, and the gap is where the framing-target's communications budget reaches in to convert the deliberated output into the surface vocabulary of a "we listened" press release without any corresponding model-development or release-criteria change.

The strategy is bad as the lone arm. A deliberated output without a binding instrument carrying it forward is a record in the public-policy archive; a deliberated output paired with coalition lobbying carrying it into a legislative process becomes a regulation; a deliberated output paired with strategic litigation anchoring its remedies becomes a court-enforced constraint; a deliberated output paired with collective bargaining becomes a contract. The deliberation arm produces the normative content the rest of the movement uses; whether the downstream arms convert that content into a binding constraint is the open question this strategy cannot answer alone.

The strategy is most powerful at the corpus's edge cases — the deployment decisions for which no other instrument is yet in place. The pre-AI-Act ground in Taiwan, the pre-frontier-model-regulation ground in the United States, the indigenous-language data-stewardship ground for which Te Hiku Media's Kaitiakitanga License became the working precedent — these are the loci at which the deliberation's normative content is what gets used because there is not yet any other instrument to use. As the binding-instrument layer matures around AI (the EU AI Act, the California SB 1047 / SB 53 trajectory, the post-2025 national AI laws), the strategy's incremental value shifts from generating the normative ground to legitimising the binding instruments the lobbying and litigation arms are carrying — a role transition the strategy's adopters have not yet collectively named.

Ecology

The strategy sits structurally upstream of the binding-instrument arms and feeds them in turn, and it is paired with counter-narrative framing as a complementary upstream-normative-content arm. Framing coins the public-affairs vocabulary; deliberation surfaces what publics actually believe inside that vocabulary; together they populate the public-affairs ground the downstream arms operate on. The vTaiwan AI-governance deliberations have produced three consensus positions — higher cultural sensitivity, encouragement of open-source AI models, training-data and source-code transparency — that are now circulating as framings inside the AI-policy field, the deliberation having produced the substantive content the framing arm then carries.

It is fed by empirical audit and expose at the substrate: a deliberation on biometric AI is asking citizens to reason about systems whose actual error distributions an Algorithmic Justice League audit established; a deliberation on algorithmic management is reasoning over the Worker Info Exchange findings; the Ada Lovelace Institute's Citizens' Biometrics Council reasoned over the empirical base of years of facial-recognition audit work. A deliberation without an audit substrate is asking citizens to reason in the dark; the audit-deliberation pair is what produces both the number and the considered preference about the number that the binding-instrument arms then carry.

It feeds coalition lobbying of binding regional regulation as the legitimacy register the lobbying coalition operates in. The vTaiwan-to-NHRC-to-AI-Basic-Act arc is the corpus's cleanest documented case of a participatory output carried into a national legislative process; the Ada Lovelace Institute's deliberative-dialogue work on biometrics has fed UK regulatory consultation; CIP's UK Frontier AI Taskforce engagement has run as a deliberation-into-policy pipeline of the same shape. A lobbying coalition without a deliberated public-input record has to claim a constituency; with one, the constituency claims itself.

It is complementary to parallel community research institution — most of the corpus's deliberation infrastructure is built and run by parallel community-research institutions. The Ada Lovelace Institute, the Collective Intelligence Project, Involve, and Te Hiku Media are all simultaneously instances of the parallel-research strategy and operators of the deliberation strategy; the two strategies overlap in operator but not in output (the parallel-research arm produces the institutional credibility and the methodological infrastructure; the deliberation arm produces the deliberated content the institutional credibility makes consumable to its target).

It is complementary to indigenous and community data sovereignty at the community-stewardship edge. Te Hiku Media's Kaitiakitanga License is itself a community-deliberated stewardship framework — six derived Māori Data Sovereignty Licences (Māori Data, Iwi Data, Hapū Data, Marae/Rūnanga Data, Rōpū Māori Data, Whānau Māori Data Sovereignty) extended the framework across iwi-to-whānau governance levels through tikanga-grounded community process. The deliberation here is not a national stratified sample but a community-internal one, and the binding force comes not from a receiving institution's reputational commitment but from the community's standing authority over its own data — the strategy's distinctive form when the deliberating community is also the data-governing one. CIP's Community Models track is the corpus's most explicit attempt to generalise that inversion (Grandmothers' Collective, ITS Rio, Equiano Institute, Civis) outside the indigenous-sovereignty case.

It is competing with strategic litigation against algorithmic state decisions for movement attention and funding inside the same harm class, and the two strategies have inverted strengths: litigation produces the binding remedy but only against the specific algorithmic state decision its writ reaches; deliberation produces the normative ground that scales across deployments but generates no remedy against any of them. The strategies are not substitutes — a coalition that runs only deliberation produces no remedies; a coalition that runs only litigation produces remedies against the cases that reach court and no normative input to the broader development decisions that produce the cases — but they compete for the same constrained civil-society budget, and the question of how a coalition weights the two is itself one of the movement's underspecified strategic decisions.

The strongest competing strategy from outside the movement is the AI industry's in-house user-preference and feedback infrastructure — the RLHF pipelines, the model-card public-comment processes, the developer feedback loops the frontier-AI labs already run, treated by the labs as their already-collecting-public-preferences apparatus. The competitive pressure is structural: the labs argue that a 1,000-person CIP assembly is a low-resolution version of what their in-house feedback loops already aggregate from hundreds of millions of users, and the deliberation strategy's structural answer — that an in-house feedback loop is not representative of the affected population, is not deliberative, is not adversarial, and is owned by the target the deliberation is attempting to shape — is the movement's core argumentative ground but is not yet operationalised as a binding constraint the labs are required to honour. The strategy's strongest hedge against this competing arm is not louder deliberation but the downstream arms — a court ruling, a regulation, a contractual binding — that hold the deliberated content in place as legal text when the in-house feedback-loop register has been published as the lab's own "we listened" narrative. The deliberation arm and the binding-instrument arms are not substitutes; they are the same democratic chain held against the same target by different ends, and the chain breaks at the end the receiving institution can reach for selectively.