Collective Intelligence Network Get involved

About · Governance

How decisions get made, and how we fix mistakes.

A foundation that asks AI to be accountable should be accountable itself. Here's how we work.

If a town hires an inspector to check that bridges are safe and scales are honest, the first thing residents reasonably ask is: who checks the inspector's own books?

We think any group proposing public guardrails for AI should live by those same guardrails in its own daily work. That means being plain about who makes decisions, how we use AI assistants internally, how we pay the bills, and how anyone can challenge an error on our pages.

Who decides today, and how that grows

Today, the Collective Intelligence Network is in its founding stewardship phase. One human steward, the founder, reviews and approves everything that is published on this site, every change to our nineteen floor principles, and every public specification.

Think of this phase like the scaffolding around a new community library while the roof is still being raised. At the very beginning, one builder signs off on the blueprints so the walls don't pull in five directions at once. But scaffolding is meant to come down as soon as the stone arches can stand on their own.

We are writing a formal charter and succession plan so the work never depends on a single person's health, judgment, or goodwill. As contributors, independent researchers, and partner communities join, governance is designed to move through three clear steps:

  • Open public review of all specifications. Every technical draft and policy report is published in plain view so outside critics can test our assumptions without needing our permission.
  • Multi-steward sign-off on high-impact changes. Just as we argue that no one should command a swarm with a single key, changes to core rules and shared registries move to two-person and multi-party sign-off across independent stewards.
  • One person, one voice on questions of value. In line with our seventh principle, no financial donor, hardware provider, or institutional partner will ever buy extra voting weight over our floor commitments.
The Three-Stage Transition from Scaffolding to Democratic Standing A progressive diagram showing governance moving from founding stewardship scaffolding, to federated multi-steward sign-off, to broad democratic public standing, all permanently supported by nineteen non-derogable floor principles. GOVERNANCE EVOLUTION · FROM SCAFFOLDING TO PUBLIC STANDING STAGE 1 · FOUNDING STEWARD Temporary Scaffolding • Human founder gate • Automated tests catch errors • Open specification drafts • Keeps early walls coherent STAGE 2 · FEDERATED STEWARDS Multi-Party Sign-Off • Two-person rule changes • Independent lab oversight • Cryptographic commit receipts • Shared registry governance STAGE 3 · DEMOCRATIC STANDING Public Constitutional AI • One person, one voice on values • Broad civic deliberation • Community local self-rule • Zero commercial voting power PERMANENT BEDROCK · NINETEEN NON-DEROGABLE FLOOR INVARIANTS No stage of governance, emergency, funding source, or majority vote may weaken the floor.
The three-stage governance path: temporary founding scaffolding gives way to federated stewards and broad democratic standing, while the nineteen constitutional floor principles remain permanently non-derogable.

How we use AI in our own work

We use AI agents every day to help draft research syntheses, check citations against primary sources, build interactive simulations, and test our own arguments from multiple viewpoints. We think it would be strange to study how humans and AI agents coordinate without practising it under our own roof.

We follow three strict house rules whenever software helps us write or build, the same rules we recommend to others:

  • Agents propose; a person decides. Think of an architectural workshop where apprentices draw initial floor plans and check the structural maths, while the licensed architect inspects every joint and signs the final drawing in ink. No agent can publish to our live site without the human steward's gate check.
  • Automated checks catch what tired eyes miss. Every build of this site runs automated checks that verify every primary-source link, flag undefined jargon in main text, reject unsupported financial promises, and enforce clear sentence length.1
  • Every agent run leaves a receipt. In our own repository, every commit made with agent assistance records the agent harness, the model used, and the verification checks that passed.

Funding and independence

Independence is easiest to lose through the wallet. A consumer watchdog that takes undisclosed sponsorship from the companies it rates soon softens its words, even when nobody in the building intends to sell out.

Today, our funding position is simple: we take no outside donations, no venture investment, and no corporate sponsorships, and we sell no products, tokens, or paid subscriptions. The founding work is self-supported by the steward.

If and when we accept public-interest grants or philanthropic support in the future, we commit to three conditions before a single payment is accepted:

  • Full public ledger of funders. We will list every funder, the exact amount given, and the purpose of the grant on this page.
  • Zero editorial or research veto. No donor or grantmaker may review, delay, or alter our research conclusions, institutional evaluations, or Observatory scores prior to publication.
  • No commercial lock-in. Everything we publish (our essays, standards drafts, simulation code, and institutional datasets) remains open for anyone to read, audit, and fork.

How we handle mistakes and corrections

Any group that writes about fast-moving technology and public policy will get things wrong. A number will be superseded by a newer study, an analogy will oversimplify an edge case, or an evaluation of an organisation in Shared Horizon will miss a newly published safety policy.

In many institutions, admitting a mistake is treated like a public-relations crisis, so errors are edited over without a note. We take the opposite view, borrowed from scientific journals and old-fashioned newspapers: a visible correction is a proof of honesty.2

When we find an error, or when a reader points one out, we do three things:

  1. We fix the factual error in the article or dataset promptly.
  2. We add a dated note on our public Corrections page explaining what was wrong and what changed.
  3. Where an institutional evaluation or forecast changes, we preserve the prior score in the history log so anyone can see how our assessment evolved.

If you spot an error, a broken citation, or an unfair characterisation anywhere on this site, please write to us. We would much rather fix a mistake in daylight than defend it in the dark.

Our relationship to other organisations

Throughout our essays, research pages, and the Shared Horizon institutional explorer, we cite hundreds of universities, standards bodies, AI labs, civil-society groups, and public agencies.3

When we cite an organisation's research or evaluate its public posture, we are crediting their published work or offering an independent public-interest analysis, never claiming an endorsement, affiliation, or partnership. If we ever enter a formal collaboration with another institution, we will state it plainly on this page.

Trust is earned slowly and lost fast. By keeping our decision-making, our internal AI guardrails, our funding rules, and our corrections in plain view, we hope to give you every tool you need to judge our work on its merits.

1. On separation of privilege and independent verification in computer systems: Jerome H. Saltzer and Michael D. Schroeder, "The Protection of Information in Computer Systems," Proceedings of the IEEE, vol. 63, no. 9 (1975), doi.org/10.1109/PROC.1975.9939.

2. On forecasting accountability, calibration, and transparent revision histories: Philip E. Tetlock and Dan Gardner, Superforecasting: The Art and Science of Prediction (2015); see also the Good Judgment Open calibration methodology at goodjudgment.com/about/.

3. For our open standards drafts and institutional evaluation methodology, see the Research and Shared Horizon pages.

4. On the democratic legitimacy crisis of corporate AI constitutions and the transition toward public constitutional AI: Gilad Abiri, "Public Constitutional AI," Georgia Law Review 59, no. 2 (2024): 601, arxiv.org/abs/2406.16696; see also Saffron Huang, Divya Siddarth, Liane Lovitt et al., "Collective Constitutional AI: Aligning a Language Model with Public Input" (2024), arxiv.org/abs/2406.07814.