CIP Alignment Assemblies (2023–ongoing)

The Collective Intelligence Project (CIP) launched Alignment Assemblies in February 2023 — a structured programme of representative public deliberations, each paired with a committed institutional audience (an AI company, government agency, or civil society organisation) that agrees in advance to take the results into its decision-making. By mid-2025 the programme had run assemblies with OpenAI, Anthropic, Taiwan's Ministry of Digital Affairs, the UK Frontier AI Taskforce, and the Creative Commons Foundation, and had extended into sustained global public-opinion infrastructure through the Global Dialogues platform and index. The corpus's gap the programme fills is the participatory AI governance sector: vTaiwan's g0v–led deliberative platform established the methodology and the East Asia precedent; CIP's Alignment Assemblies are the US practitioner organisation that carried that methodology into direct engagement with the frontier-AI companies doing the training, evaluation, and deployment decisions the methodology is meant to shape.

Origins: inserting public voice before the fact

CIP's founding whitepaper frames the organisation's core argument: that by the time AI policy surfaces as a regulatory or legislative question, the design decisions most consequential for the public — what values a model should embody, which risks to prioritise in evaluations, how to weight competing societal interests in deployment — have already been made by a small technical and corporate elite. Alignment Assemblies are the campaign mechanism for pushing democratic input upstream — into the development and evaluation phase rather than the after-the-fact policy phase. The "committed audience" structure is load-bearing: each assembly is tied to a named institutional partner who publicly agrees to consider the results, making the public input binding as a reputational commitment even where it is not legally enforceable. Divya Siddarth framed the empirical bet underlying the whole programme: "Every time we run something, we get the most interesting, engaged, thoughtful responses, and we show it can be done."

v0 and v1: Summit for Democracy and the OpenAI risk assembly (March–June 2023)

The first assembly (v0), in March 2023, addressed global policymakers at the US Summit for Democracy — using a deliberative format to surface how democracy's custodians understood generative AI's risks to democratic institutions — and was covered in the New York Times. The second assembly (v1), conducted over two weeks in June 2023 in partnership with OpenAI, is the programme's first direct corporate engagement. Approximately 1,000 demographically representative Americans participated through the AllOurIdeas wikisurvey platform, submitting and ranking statements that completed "When it comes to making AI safe for the public, I want to make sure…". The findings inverted what AI-industry conversations might have predicted: the categories of oversight, understanding, and governance ranked highest; concern about overbearing regulation ranked lowest; and the highest-ranked single statement emphasised avoiding overreliance on AI systems that neither the public nor researchers fully understand. OpenAI committed to use the findings to inform model evaluations, release criteria, and broader standards-setting processes. Six participants from the assembly joined a follow-up roundtable directly with OpenAI representatives — the closest the programme came in this phase to face-to-face deliberation with the committed audience.

v2 and v3: Taiwan and the Anthropic constitutional assembly (May–October 2023)

In May 2023 CIP announced a partnership with Taiwan's Ministry of Digital Affairs — the ministry then led by Audrey Tang — making Taiwan the programme's first government partner outside the United States and its first non-anglophone context. Assembly v2 ran as a series of Ideathons titled "Democratizing AI Futures" from July 2023, followed by deliberative workshops in August examining copyright, bias, fair use, and societal impacts of generative AI, with findings integrated into Taiwan's AI policy recommendations. The partnership connected CIP's programme to the vTaiwan and g0v methodological tradition from which the Polis tool — the consensus-mapping platform CIP uses in its most complex assemblies — descends.

Assembly v3, conducted in October 2023 with Anthropic, is the programme's most significant early output. Approximately 1,000 representative Americans deliberated via the Polis platform to produce a 275-statement constitution specifying how an AI language model should behave — a set of principles that Anthropic then used to train a model, and published as joint CIP–Anthropic research. The published paper describes the project as "one of the first instances in which members of the public have collectively directed the behavior of a language model via an online deliberation process." The public constitution differed substantively from Anthropic's own in-house 58-principle document: the collectively written version placed notably greater weight on avoiding paternalism, protecting privacy, and providing balanced perspectives on contested political and social questions — differences that are themselves a data point about the gap between expert-corporate intuitions and democratically surfaced preferences. Also in October 2023, v4 with the Creative Commons Foundation addressed the contested question of AI training on CC-licensed materials, applying the same assembly methodology to a civil society organisation rather than a corporation.

Expanding infrastructure: 2024

CIP's 2024 Annual Report documents the programme extending along two new tracks. A nationwide US assembly on information integrity ran in March 2024, applying the Alignment Assembly methodology to AI-generated misinformation and examining how democratic processes could address that question at policy scale while preserving cultural context and local values across a nationally diverse participant sample.

The Global Dialogues initiative moved the programme from episodic assemblies to recurring international infrastructure: a 2024 pilot engaged 1,200 participants across 68 countries in 8 languages on AI governance and values questions, producing the first systematic multinational public-opinion dataset specifically on AI governance. The Community Models track inverted the assembly architecture: rather than sampling from the general public to advise corporations, it enabled specific communities — the Grandmothers' Collective, ITS Rio, Equiano Institute, and Civis among the 2024 cohort — to build AI systems shaped by their own deliberated values, directly addressing the historic exclusion of non-Western and non-technical communities from AI development decisions.

Global Dialogues Index (2025)

The 2025 Global Dialogues Index, compiled from seven rounds of deliberation over the preceding year with over 6,000 people across 70 countries, was released as an open-source dataset structured around three indices — Usage, Trust, and Perception — designed as "actionable signals for policymakers and developers." Headline findings: 58 percent of respondents trust AI chatbots more than their elected representatives; a significant gap separates trust in AI chatbots (55 percent) from trust in AI companies (34 percent); 75 percent of employees expect weekly AI use at work; and AI reinforces certainty about existing beliefs more powerfully than social media, with 44.5 percent reporting increased certainty after AI interaction. The dataset was released open-source for any party to build on.

Significance for the broader AI-good movement

For the make-AI-good corpus, the Alignment Assemblies programme matters on three connected counts. First, it is the corpus's principal US anchor for the participatory-AI-governance strategy — the organisation that moved the strategy from theoretical advocacy into concrete deliverables: a trained language model on a publicly written constitution; a seventy-country multinational opinion dataset on AI governance; and a working roster of institutional partnerships with the frontier-AI companies whose training and deployment decisions the strategy is designed to shape. Second, by tying each assembly to a named, publicly committed audience, CIP operationalised the accountability mechanism that generic "public consultation" processes notoriously lack — the results don't only go into a report, they go to a named party with a stated commitment, making the gap between published findings and observable company behaviour legible to subsequent scrutiny. Third, the Global Dialogues infrastructure — now producing a recurring open-source multinational dataset — represents the movement's most mature attempt to institutionalise global public voice as a standing input to AI governance rather than a one-off event, and to make that voice available not just to AI companies but to any actor who can build on the data.