Tracking individuals, institutions and incidents at the frontier. Mapping the positions of all actors on a coordinated slowdown. Mitigating existential risk from the AI race narrative.
Coordination on AI Risk
An open record of where AI labs and legislators stand on a coordinated slowdown and existential risk from AI.
Statements
Materials
Milestones
Incidents
Every incident
Positions
Where AI labs and legislators stand on a coordinated slowdown and existential risk from AI.
Race
The race to build ever more capable AI, run side by side: each frontier lab's releases, and the moves states make to stay ahead of each other.
Lap by lap
Glossary
Timelines
Plans
Companions
Actors
Mission
What this site is for, and how to read it.
No one can slow the race to superintelligence alone. Coordination starts with seeing clearly: who stands where, what has been said, and what has already gone wrong.
01The mission
Coordination on AI Risk is an open record of where the people and institutions building and governing frontier AI stand on a coordinated slowdown, and on the risk that advanced AI could cause human extinction.
It exists to make that picture legible: so that legislators, researchers, lab staff and the public can see how much agreement there already is, where the gaps are, and what it would take for everyone to move together.
02Why it matters
The leading labs are racing to build AI that can do the work of AI researchers, which could make progress run faster than anyone's ability to understand or control it. Many of the people closest to that work say it could go catastrophically wrong.
Each lab, and each country, can argue that slowing down alone only hands the lead to someone less careful. That is a coordination problem: a slowdown only works if it is shared, and it can only be shared if everyone can see that others are willing too.
03How to read this site
- Positions maps where the labs' leaders and boards, and every member of the UK and US legislatures, stand.
- Plans sets out the proposals for handling advanced AI, and what each asks of the race: a halt, pacing or speeding up.
- Race, Timelines and Milestones show how fast the frontier is moving, when AGI is expected, and the moments that have shaped it.
- Incidents records the times AI systems have already slipped out of control, misbehaved or been misused.
- Actors, Statements and Materials gather the people, public commitments and writing that frame the debate; the Glossary explains its terms.
- Coordinate and Companions point to what you can do, and where else to follow along.
04Principles
- Sourced
- Every position, statement and incident links to where it comes from.
- Recorded, not inferred
- “No recorded position” means none has been documented here yet, not that a person has none. People are described by what they have said and done in public.
- Open to correction
- Roles change and records go out of date. Corrections and new sources are always welcome.
- For coordination
- The aim is not to single anyone out, but to show that a shared slowdown has more support than any one actor can see from where they stand.
Something missing or wrong? Get in touch →
Site Map
Every page on the site and what you'll find there, in the order of the menu.
Coordinate
A coordinated slowdown needs people across government, industry and the public to act together. Here is where to start.
-
Tell your representatives
Legislators are weighing bills on superintelligence and catastrophic risk now. A short, personal message about where you stand carries weight, especially if you ask for a specific action.
-
Add your name
Public statements show decision-makers how broad support for coordination is.
-
Join a campaign
Campaign groups organise briefings, letters and events, and can put you in touch with others near you.
-
Improve this record
Every position, statement and incident here needs a source. If you know of one that's missing, or something that's out of date, send it in.
A slowdown only holds if the countries at the frontier move together. Here is what already exists, what has been proposed, and what an agreement would need.
What exists
-
The AI safety summits
The series where 28 countries, including the US and China, first agreed that frontier AI could cause serious harm, and where labs signed safety commitments. Its focus has since drifted from safety to adoption.
The Bletchley Declaration → -
A network of AI safety institutes
Government institutes in the UK, US, EU, Japan, Korea, Canada and elsewhere that test frontier models and share evaluation methods: the beginnings of shared technical capacity.
Government institutions on Actors → -
A UN scientific panel and global dialogue
The UN General Assembly set up an Independent International Scientific Panel on AI and an annual Global Dialogue on AI Governance among all member states.
The General Assembly's resolution → -
The first binding AI treaty
The Council of Europe's Framework Convention requires AI systems to respect human rights, democracy and the rule of law. It shows states can bind themselves on AI, though it doesn't limit frontier capabilities.
The Framework Convention →
What's proposed
-
Red lines for AI
Western and Chinese scientists proposed red lines no AI system should cross, such as copying or improving itself autonomously; over 200 public figures then asked governments to agree verifiable red lines by the end of 2026.
The Global Call for AI Red Lines → -
A treaty to prevent superintelligence
MIRI's draft agreement, led by the US and China, would cap training compute and chip clusters, with a verification regime built for rivals who don't trust each other.
MIRI Technical Governance Team ↗ -
A US–China deal to delay superintelligence
AI 2040: Plan A imagines the two powers agreeing to hold back superintelligence until 2040, with chip tracking and all AI research made public.
ai-2040.com ↗ -
Tools to pace the frontier
More than 1,100 staff of the frontier labs asked the US government to support an international effort to build the technical and governance tools to deliberately pace frontier AI.
Pacing the Frontier →
What an agreement would need
- Verification
- A way for each side to check the other keeps its word, such as tracking advanced chips and large training runs.
- Shared red lines
- Agreed capabilities that no one builds or deploys, precise enough to check.
- A body to watch
- An international agency, often compared to the nuclear watchdog, to monitor compliance and share findings.
- The leaders first
- The US and China, where almost all frontier development happens, agreeing together; others can follow.
A slowdown is a collective action problem: everyone may be better off if all slow down, but no one wants to slow down alone. Collective action is how that changes.
How it works
- Common knowledge
- Public statements let everyone see that others share their concern, so no one has to guess whether they'd be alone.
- Conditional commitment
- Pledges that take effect when others join, so no one has to move first.
- Voices from inside
- When the people building the frontier speak together, and by name, it's harder to say that a slowdown has no support.
- Pressure from outside
- Campaigns, letters and votes turn private worry into a public mandate that governments can act on.
On the record
-
Pacing the Frontier
Employees of Anthropic, OpenAI, Google and Meta, among them chief executives and chief scientists, asked for the tools to pace frontier AI together.
The statement → -
Statement on Superintelligence
Scientists, public figures and people from across politics called for a prohibition on developing superintelligence until it can be done safely and with public support.
The statement → -
Statement on AI Risk
A single sentence placing the risk of extinction from AI alongside pandemics and nuclear war, signed by leading scientists and the chief executives of the frontier labs.
The statement → -
Pause Giant AI Experiments
The open letter asking labs to pause training systems more powerful than GPT-4 for six months, and governments to step in if they wouldn't.
The letter →
Ready to add your voice? See the next steps →
Contact
Corrections, sources and suggestions are welcome. If you're pointing to something new, please include a link to the source.