The Collective Intelligence Project (CIP) is a US-based non-profit research and deployment organization, with global partnerships, co-founded by Divya Siddarth and Saffron Huang. It treats AI governance as a collective-decision-making problem and builds deliberative-democracy tools intended to channel broad public input into concrete AI decisions, both inside companies and inside governments. Its core method is the Alignment Assembly.
Overview
CIP frames AI governance as a collective-decision-making problem and develops deliberative-democracy tools to allow broad public input into concrete AI decisions, both inside companies and inside governments. Its flagship method is the Alignment Assembly.
Activities
Alignment Assemblies (2023+)
Alignment Assemblies, run from 2023 onward, are structured deliberative processes that convene representative publics, often selected by sortition, to develop input into AI governance decisions. CIP applies four design criteria: (1) a defined outcome, (2) a relevant polity, (3) a defined discussion scope, and (4) a set of tools and process. The method is detailed on the Alignment Assemblies and Collective Constitutional AI concept page.
Collective Constitutional AI (with Anthropic, 2023)
In a 2023 project with Anthropic, representative Americans drafted a constitution for Anthropic's LLM. Compared with the researcher-drafted constitution baseline, the publicly-drafted model was less biased but equally capable. The project was one of the first instances of collective direction of LLM behavior. See Constitutional AI.
Participatory risk prioritization (with OpenAI, 2023)
In a 2023 project with OpenAI, 1,000 demographically-representative Americans participated via the AllOurIdeas platform on AI risks. The process surfaced "over-reliance" as the top public concern, distinct from the AI industry's typical focus on existential risk.
Taiwan moda partnership (2023+)
In a partnership with Taiwan's Ministry of Digital Affairs beginning in 2023, CIP ran ideathons, a Recursive Public AI governance deliberation, and an information-integrity assembly for the 2024 election cycle. Audrey Tang had a role in this work. Taiwan's open-source LLM (TAIDE) uses a constitution drafted in CIP's 2023 Taiwanese Alignment Assembly for its alignment-tuning phase.
Positions
CIP has advanced the framing that democratic governance can apply to corporate AI decisions, not only to state decisions, arguing that "democracy can form and sustain a system of checks and balances in AI governance." The Siddarth/Huang/Tang Digitalist Papers essay ("A Vision of Democratic AI") is the canonical articulation of this view.
Funding and governance
CIP's academic and foundation partners include Stanford HAI, Harvard's Edmond J. Safra Center for Ethics, and the Oxford Internet Institute. Its corporate partners include Anthropic, OpenAI, and Meta. Its government partners include Taiwan moda and the UK Deliberation Lab.
Relationships
- depends-on: Divya Siddarth / Saffron Huang — founders.
- supports: Alignment Assemblies and Collective Constitutional AI — flagship method.
- supports: Constitutional AI — Collective Constitutional AI extension.
- supports: The Digitalist Papers (Stanford, Volumes 1–2) — Siddarth/Huang/Tang essay source.
- related: Audrey Tang — frequent collaborator (Taiwan moda).
- related: Anthropic / OpenAI — lab partners.
- related: Taiwan Ministry of Digital Affairs (moda) — government partner.