COLLECTIVE INTELLIGENCE PROJECT

OLYMPUS opened an institutional file on the Collective Intelligence Project because it is the one member of the AI Evaluator Forum that does not test a model against a capability. It asks the public what a model should do and converts the answers into rubrics. The labs are its partners on both halves: Anthropic trained a model against a constitution CIP helped draft, OpenAI sat as a partner in a 2023 risk-ranking exercise, and both appear as organizational partners in CIP’s 2025 annual report. The Forum’s standard asks evaluators to disclose conflicts and says “The system provider does not control any of the evaluator’s board seats.” This is not a psychometric profile. It is a funder list, a board list and a partner list. The finding is the shape: an evaluator whose work product is legitimacy, paid by philanthropy, partnered with the developers whose models it grades, and seated at the table that writes the rules for independence. Not a capture. A fit.

Institutional Archetype

THE DELIBERATION SHOP — Subject says its mission is “to direct technological development towards the collective good” (CIP, Team), and its 2025 annual report describes the method: “We combine deliberative methods, large-scale human evaluation, and institutional partnerships in a way no lab, regulator, or civil society organization can do alone” (CIP, 2025 Annual Report). It runs three instruments: Global Dialogues, recurring AI-facilitated deliberations with participants recruited through Prolific; Weval, “An open platform for building evaluations that test what matters” (Weval); and alignment assemblies, one-off public-input processes with a named partner (CIP; CIP, Global Dialogues; CIP, Alignment Assemblies). The throughline is that the unit of evaluation is the public’s stated preference, not a model’s measured capability, and the customer for that unit is the lab or government that wants a defensible reason for a setting. This profile scores the shop, not its staff.

Mandate & Origin

• Founders and current leadership. The team page names “Divya Siddarth” as “Co-Founder & Executive Director” and “Saffron Huang” as “Co-Founder & Board Director (Secretary).” It also lists “Joal Stein” under the title “Executive Director,” beside Siddarth, and “Austin Wu” as “Board Director (Interim President of Board)” (CIP, Team). The IRS return for fiscal 2025, filed 28 July 2026, lists Siddarth as Executive Director, “Joel Stein” as Deputy Director, Huang as Board Member and Wu as Treasurer (ProPublica Nonprofit Explorer). The site and the return agree on Siddarth and differ on Stein’s title.

• Siddarth’s record. The team page says she was “a political economist and social technologist in Microsoft’s Office of the CTO” and “the AI and Democracy lead at the U.K.’s AI Safety Institute” (CIP, Team). The Forum’s AEF-1 paper spells her name “Divya Siddharth” in the author list; the organization’s own pages spell it Siddarth.
• Legal form and sponsor. “Collective Intelligence Project is a nonprofit, tax-exempt 501(c)(3) organization” (CIP, Funders and Partners). ProPublica lists “Collective Intelligence Project Inc,” San Francisco, “Tax exemption issued: Oct. 2024,” EIN 92-0327339. The 2024 annual report adds “Special thanks to RadicalxChange Foundation for serving as our fiscal sponsor” and “2024 sees CIP’s inauguration into the halls of official 501(c)3 nonprofits” (CIP, 2024 Annual Report). The Survival and Flourishing Fund’s list names “Radicalxchange Ltd.” as the legal recipient of a 2024 grant for CIP’s “Global Representation/Constitutions Project” (SFF, Recommendations).
• The instruments, in its own words. Weval: “We empower a global community to create qualitative benchmarks for any domain.” Global Dialogues: “We engage thousands of people across the globe through structured, AI-enabled deliberation to understand how they’re interacting with and impacted by AI systems.” Jhilmil: “We are conducting cultural and social evaluations to better measure and improve the contextually-specific capabilities of frontier AI” (CIP). Weval’s own about page says evaluations “provide a rigorous, credible way for AI labs and governments to make decisions on model development or crafting policy” (Weval, About).
• Collective Constitutional AI, October 2023. “Anthropic partnered with the Collective Intelligence Project to run a public input process using the Polis platform”; about 1,000 Americans voted on rules for a chatbot. Anthropic’s stated reason: “it also highlights the outsized role we as developers play in selecting these values—after all, we wrote the constitution ourselves” (Anthropic, Collective Constitutional AI). CIP lists it as “v3: Collective Constitutional AI, October 2023” with “Partner: Anthropic,” and lists “v1: Democratizing Risk Assessments, June 2023” with “Partner: OpenAI” (CIP, Alignment Assemblies). The Global Dialogues page says “CIP ran a series of alignment assemblies in 2023-24 alongside committed audiences like OpenAI, Anthropic, and the UK AISI, where partners committed to taking into account public voice in their decision-making” (CIP, Global Dialogues).
• The Forum seat. The AI Evaluator Forum’s members page describes CIP as “Embedding democratic global guidance into the development and evaluation of frontier AI” (AI Evaluator Forum, Members). Two CIP staff, Siddarth and Faisal Lalani, are among the 26 listed authors of AEF-1 (AEF-1 PDF). The Forum’s letter of 18 September 2026 lists “Joal Stein” as “Executive Director, Collective Intelligence Project” among its signatories, who “sign in their personal capacity” (AI Evaluator Forum, letter); see the AI Evaluator Forum.

Funding & Backers

• Named funders, from CIP. The 2025 annual report thanks, as “funding partners,” Fast Forward, the Future of Life Institute, Google.org, Omidyar Network, the Robert Wood Johnson Foundation and the Survival and Flourishing Fund (CIP, 2025 Annual Report). The funders page also carries logos linked to the Ford Foundation, OneProject and the Amaranth Foundation (CIP, Funders and Partners); the 2024 report names the Omidyar Foundation, the Robert Wood Johnson Foundation “and the Ideas for an Equitable Future team,” “Google.Org and the Jigsaw team,” the Survival and Flourishing Fund, the Future of Life Institute, the Ford Foundation and OneProject (CIP, 2024 Annual Report). No amounts are published on CIP’s pages. Its home page says “Funded by leaders in AI.”
• The Survival and Flourishing Fund, amounts from the Fund’s own table. Three rows name CIP: “SFF-2023-H1,” source Jaan Tallinn, $300,000 to “The Collective Intelligence Project”; and two rows for “Initiative Committee 2024,” source Jaan Tallinn, $500,000 each, one to “Collective Intelligence Project Inc.” for “General support” and one through “Radicalxchange Ltd.” for “General support of CIP’s Global Representation/Constitutions Project,” a total of $1,300,000 (SFF, Recommendations). The 2024, 2025 and 2026 main-round tables do not name CIP. See Survival and Flourishing Fund and Jaan Tallinn.
• The conflict-of-interest policy. “The Collective Intelligence Project operates exclusively through philanthropic charitable donations and currently does not accept funding from commercial partnerships or contractual relationships.” The same page: “Our work is made possible through the support of committed philanthropic partners who believe in our mission and sustain our independence as a nonprofit organization” (CIP, Funders and Partners).
• The returns. Fiscal 2024: revenue $1,908,186, of which contributions $1,907,953; expenses $268,349; net assets $2,463,506. Fiscal 2025: revenue $110,693 (contributions $98,460, program services $9,000), expenses $1,125,633, executive compensation $261,709, net assets $1,448,566 (ProPublica Nonprofit Explorer). The 2025 return lists Schedules A, B, D, I, J and O; their contents are not in the summary.

• Partners that are also developers. The 2025 report lists as “organizational partners” Anthropic, the Aspen Institute, Common Knowledge, Microsoft Research, Fathom, Jigsaw, Karya, Mistral AI, Mozilla, OpenAI, Prolific, Public AI, Remesh, the Stanford Deliberative Democracy Lab, the Industry-Wide Deliberative Forum and Transluce (CIP, 2025 Annual Report). Anthropic and Mistral AI also appear on the funders page’s partner row (CIP, Funders and Partners). The report lists no lab among its “funding partners”; Google.org is the philanthropy of a lab’s parent company.
• Board seats held by lab employees, on their own pages. Austin Wu’s team-page bio says “his day job is working on language models at Google.” Saffron Huang’s own profile says “Currently a research scientist on the Societal Impacts team at Anthropic” (Saffron Huang, Substack); the Form 990 lists her as a board member who drew $41,863 from CIP in fiscal 2024 and nothing in fiscal 2025. AEF-1’s requirement is that “The system provider does not control any of the evaluator’s board seats” (AEF-1 PDF); the team page lists two board directors, Huang and Wu, and no CIP page states how many seats exist or who votes.

Institutional Voice & Intent

The register is civic and procedural, with the verbs of an institution-builder: “infrastructure,” “pipeline,” “standing,” “legitimacy.” The 2025 annual report states the claim in one line: “Global Dialogues captures values. Weval turns those values into leverage.” The aim is dated: “By 2027, we aim for Global Dialogues and Weval to function as a standing global pipeline used by multiple governments and AI labs: capturing representative public input, translating it into operational evaluation criteria, and embedding those evaluations into model release, procurement, and oversight decisions” (CIP, 2025 Annual Report).

Stated intent: “Steering transformative AI towards better outcomes. For everyone” (CIP); to ensure decisions on “how and when to release models, determining underlying principles for AI behavior, building for cultural pluralism and language diversity” do not “fall to a small percentage of those likely to be affected” (CIP, Global Dialogues).

Observed intent: The record shows project-by-project lab work. In 2023, a risk-ranking assembly with OpenAI, a constitution with Anthropic and a Taiwanese ministry’s process; in 2025, Weval evaluations for India, Sri Lanka and the United States. The 2025 report says Weval evaluations “influenced model tuning, safety decisions, and release planning,” names “Anthropic + Karya” on two India evaluations (legal assistance and agriculture), and says the Anthropic agriculture evaluation “will launch” (CIP, 2025 Annual Report). It also names, as users of Global Dialogues insights, Meta, Cohere, Taiwan’s Ministry of Digital Affairs and the UK and US AI Safety Institutes.

Gap: The Forum’s standard asks an evaluator to disclose funding that comes from “a system provider, its employees, or its direct competitors” (AEF-1 PDF). CIP’s disclosed money is philanthropic, and it says it takes no commercial or contractual funding. Its disclosed relationships with developers are partnerships, a Google philanthropy as funder, and two directors’ employers, not payments for evaluations. The record can settle who is listed where. It cannot settle whether a partner shapes a rubric or a finding, and the report names “Anthropic + Karya” as the pairing on evaluations of Anthropic-relevant domains without saying who pays for them.

Interest: What each position would pay CIP, documented, with its own best case. The positions are CIP’s.

1. Position: democratic public input should be a standard part of how frontier AI is evaluated and authorized. The annual report: “Success looks like democratic legitimacy becoming a routine requirement of deployment, not an afterthought, especially in high-stakes domains such as health, elections, public benefits, and personal advocate agents” (CIP, 2025 Annual Report).
   • What adoption would pay CIP. A requirement of that kind runs through the two products CIP owns and operates: Global Dialogues, which supplies the input, and Weval, which runs the resulting evaluations on a platform CIP built. The 2025 return shows program-service revenue of $9,000, total revenue of $110,693 against $1,125,633 of expenses, and net assets down from $2,463,506 to $1,448,566.
   • CIP’s best case, steelmanned. Developers write the rules by default: “By default, these decisions fall to a small percentage of those likely to be affected: a recipe for blind spots, overlooked points of failure, and monoculture.” Anthropic’s own post concedes the premise about its constitution.
   • Who else gains from the same rule. Developers that want an external, representative basis for a behavior setting (Anthropic called the 2023 exercise a way to curate a constitution with “a large number of people who do not work at Anthropic”); governments buying AI, which the report says used CIP evaluations “for oversight and procurement”; and the philanthropic funders of the field.
2. Position: pluralistic, community-built evaluations alongside lab and government benchmarks. Weval: “Just as Wikipedia democratized knowledge, Weval aims to democratize evaluation, ensuring that AI works for, and represents, everyone” (CIP).
   • What adoption would pay CIP. Usage of its platform and its cited role in the evaluations labs and governments publish; the report’s 2026 plan says CIP will “publish reference evaluation suites” and launch an “Epistemic Evaluation Suite” whose benchmarks “will run on Weval as continuously updated leaderboards.”
   • CIP’s best case, steelmanned. “Evaluations as they are now measure what’s easy, not what’s important to everyday people” (Weval, About).
   • Who else gains from the same rule. Civil society partners the report names as evaluation partners (Karya, Factum, the Human Line Project and others); labs that want evaluations of cultural fit in markets they are entering; and the other benchmark platforms, which gain a rubric source.
3. Position: independent evaluators should disclose and manage conflicts (AEF-1). CIP is a founding member and two staff are authors.
   • What adoption would pay CIP. A standard that credits philanthropic, non-contractual funding and a published conflict policy describes CIP’s own arrangement; the evaluators with contracts or lab-linked donors are the ones with more to disclose.
   • CIP’s best case, steelmanned. “The operating conditions of third-party evaluations can be opaque, even though they can significantly impact whether an evaluation is trustworthy and impartial” (AEF-1 PDF).
   • Who else gains from the same rule. The Forum’s other seven members, whose reports can carry the checklist, and regulators who want a reference text.

Position in the Apparatus

• The public-input evaluator, between the labs and the legitimacy question. METR, Transluce and SecureBio test what a model can do. CIP supplies the answer to what a model should do, and it does so as the partner of the developers. Its 2023 projects are the template: a lab runs the model, CIP runs the public.
• Beside its Forum peers. Like AVERI, a Forum member whose staff co-wrote AEF-1 and whose funders overlap, and like Apart Research and Artificial Analysis, a source of evaluation inputs, CIP is a supplier to the evaluation field more than a grader of it. It differs from all three on the money: no cash from commercial sources by its policy, and philanthropy that includes two of the same funders AVERI and the Forum members draw on.
• Where it sits on the funder map. The Survival and Flourishing Fund funds METR, SecureBio and CIP among Forum members, per the Forum’s account and SFF’s own tables; the Future of Life Institute funds CIP; Fathom, a named AVERI funder, appears among CIP’s organizational partners; Transluce, the Forum Chair’s employer, is also a CIP partner. Schmidt Sciences, Open Philanthropy and Halcyon Futures, the funders behind other members, are not named on CIP’s pages.
• Government work. CIP’s partner list includes Taiwan’s Ministry of Digital Affairs, which its 2023 pilot “v2” names as partner; its 2025 report says the UK and US AI Safety Institutes used Global Dialogues insights and that “India, Taiwan, Sri Lanka” received evaluations “for oversight and procurement” (CIP, Alignment Assemblies; CIP, 2025 Annual Report). Its team page lists Audrey Tang, “Taiwan’s 1st Digital Minister (2016-2024),” as Senior Research Fellow.

Actions & Leadership Choices

Read by deeds, CIP is a small organization that has convened the public for the labs, built an evaluation platform around the result and joined the table that writes the rules for evaluators. Its leadership choices are two co-founders in the organization and on its board, a board that includes employees of two developers, and a fiscal sponsor until its own exemption arrived.

Actual founding purpose. To put the public inside the decisions developers make about model behavior. The 2023 whitepaper-era pilots, the White House Summit for Democracy exercise, the OpenAI risk-ranking and the Anthropic constitution, were each run with a partner that agreed to consider the result (CIP, Alignment Assemblies).

• Global Dialogues. Launched “at the Paris AI Action Summit in February” 2025; “In 2025, we ran six deliberative dialogues across more than 70 countries, engaging over 10,000 participants” (CIP, 2025 Annual Report). The 2025 Index page says “seven rounds of deliberation with more than 6,000 people across 70 countries” (CIP, 2025 Global Dialogues Index). The two pages give different counts for the same year.
• Weval and the India work. Weval’s about page says “Built by The Collective Intelligence Project.” Its Jhilmil pilot: “The Collective Intelligence Project, in partnership with Microsoft Research India and Karya, is launching a pilot to test a community‐driven approach for evaluating frontier AI models on real‐world, social impact criteria in India” (CIP, Jhilmil). The 2025 report’s “Unbiased AI in the United States” process had “approximately 1,000 liberals, moderates, and conservatives” and produced “400 prompts and 107 evaluation criteria.”
• The Forum. A founding member; two staff on AEF-1; no CIP evaluation is listed among the three that carry the AEF-1 checklist (AI Evaluator Forum, AEF-1).

Leadership choices. The team page lists staff, two board directors, research fellows, an advisory council of twelve names and a community council of twenty, the council lists carrying names without titles, among them “Jade Leung.” The executive-director title appears twice on the site and once on the IRS return. The board’s size is not published.

CONDUCT: PARTICIPANT-SOURCING EVALUATOR — THE PUBLIC’S INPUT IS THE PRODUCT, AND THE LABS ARE THE PARTNERS. CIP is not hostile to the developers and its stated purpose is the opposite of capture. What it has is a business model that depends on developers and governments adopting its input as a legitimating step, board seats that two developers’ employees hold, and a Forum seat at the table whose standard asks every evaluator to disclose exactly that kind of tie. The structure is the finding; the intent is not asserted.

Alliance Map

These run at the same time. One San Francisco nonprofit is paid by the philanthropic funders behind several other Forum members, lists two developers as partners and a third’s parent as a funder, seats employees of two developers on its board, hosts evaluations for one of them in India and sits at the table that writes the standard asking evaluators to disclose exactly this. Each tie is stated by CIP, the Fund, the IRS summary or a developer. The overlap is the finding.

Reach Assessment

Institutional: Moderate-low. Its reach is the reach of a supplier of inputs: developers and governments have used its dialogues and evaluations, on its account, and none is required to. It cannot gate a release or a purchase.

Memetic: Moderate. Collective Constitutional AI is a public-input exercise with a frontier lab that the lab published itself, and the 2025 Index publishes headline numbers, among them “58% of people trust AI chatbots more than their elected representatives” (CIP, 2025 Global Dialogues Index).

Civilizational: Low to moderate, with low durability. The shop builds no model. It shapes the claim that a model setting is publicly legitimate, and a setting that carries that claim is harder to contest. The hazard is a legitimating step supplied to the developers it checks, by an organization partnered with them and paid by philanthropy they share with its peers. The reach is the finding; the malice is not asserted.

Score move: none (new profile). Calibrated against AVERI (5/6/4, hazard 50), the AI Evaluator Forum (5/6/4, hazard 50), Apart Research (5/5/5, hazard 50) and Transluce (6/7/6, hazard 52), with Artificial Analysis (8/4/6, hazard 60) as the reach contrast.

Score basis. Atk 4 is reach: CIP’s own report says Global Dialogues insights “were used by partners including Meta, Cohere, Taiwan’s Ministry of Digital Affairs, and the UK and US AI Safety Institutes” and that Weval evaluations “influenced model tuning, safety decisions, and release planning” (CIP, 2025 Annual Report), and Anthropic’s post confirms the 2023 constitution exercise (Anthropic). It is one under AVERI, Apart Research and the AI Evaluator Forum at 5, because AVERI’s coauthored standard and an endorsed bill that became law, and the Forum’s reported EU endorsement, reach a regulator, while CIP’s outputs are inputs a partner may use; it is far under Artificial Analysis at 8, whose index buyers use as a procurement reference. Def 5 is funding insulation: CIP says it “currently does not accept funding from commercial partnerships or contractual relationships,” names its philanthropic funders and publishes a conflict policy (CIP, Funders and Partners). It is not higher because Google.org is a funder, Anthropic and OpenAI are partners, two directors’ employers are a developer and a developer’s parent, the Survival and Flourishing Fund’s $1,300,000 is large against FY2024 contributions of $1,907,953, and FY2025 revenue was $110,693 against $1,125,633 of expenses (ProPublica Nonprofit Explorer). It is one under AVERI’s 6, whose own page says it takes no cash from frontier AI companies, and level with Apart Research at 5, whose funders are also philanthropic with disclosed amounts. Hp 4 is entrenchment: an IRS exemption since October 2024, net assets of $1,448,566, two annual reports and a platform with, on its site, “1,000+ contributors” (Weval). It is not higher because no statute, contract or standard installs it, and the 2025 return spent ten times what it took in. It matches AVERI and the Forum at 4. The values hold at 4/5/4, hazard 43.

Sources: CIP; CIP — Team; CIP — Funders and Partners; CIP — Global Dialogues; CIP — 2025 Global Dialogues Index; CIP — Alignment Assemblies; CIP — Jhilmil; CIP — 2025 Annual Report; CIP — 2024 Annual Report; Weval; Weval — About; Anthropic — Collective Constitutional AI; Survival and Flourishing Fund — Recommendations; ProPublica Nonprofit Explorer — Collective Intelligence Project Inc; Saffron Huang — Substack; AI Evaluator Forum — Members; AI Evaluator Forum — AEF-1; AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations (PDF).