Slowdown.AI Center for AI Safety Add your name

An international call for restraint · Opened for signature 2026

Slow down AI development.

We want time. Not to abandon artificial intelligence. Not to unplug what already exists. To stop building past the frontier until the people building it can show — not assert — that the next system will remain under human control.

The Slowdown Statement is a public register. It asks governments for four ordinary instruments of governance — a threshold, a licence, an audit, and an inspectorate — applied to the one technology that has so far escaped all four.

Convened by the Center for AI Safety · San Francisco · Open to every signatory category

Seal of the Slowdown Statement THE SLOWDOWN STATEMENT ◆ PROOF BEFORE POWER ◆ MMXXVI
10,000
First milestone
Signatories sought before the statement is formally transmitted to governments.
Scientists & engineers
Researchers and technical staff in any field, at any institution.
Policymakers & officials
Serving and former legislators, ministers, regulators, and staff.
Civil society, faith & labour
Leaders of organisations speaking for a constituency of members.
Citizens
Anyone, anywhere, who is willing to be counted. No qualification required.

† Note on this draft This page is a design draft. The milestone above is a target, not a count. No signatures have been collected, no figures on this page are real, and nothing submitted here is transmitted anywhere. On opening, the register publishes verified counts by category and country, updated daily.

Article  I

The Statement

Text as opened for signature

Opened for signature · San Francisco

A Call for a Verified Slowdown in the Development of Frontier Artificial Intelligence

Convened by the Center for AI Safety

Whereas artificial intelligence has, within a single decade, moved from a specialised research discipline to a general-purpose technology whose most capable systems are not understood in detail by anyone, including the engineers who train them;

Whereas the pace of that development is set not by evidence of safety but by competition among a small number of firms and states, each of which has said in public that it would prefer to move more carefully if its rivals would do the same;

Whereas the leading laboratories have themselves warned that these systems may pose risks of catastrophe, including risks to the survival of our species, and have nonetheless accelerated;

Whereas no government today possesses the technical capacity to evaluate, licence, or halt a frontier training run, and no international body holds any authority over the technology whatsoever;

Whereas the burden of proof in every other field of consequence — medicine, aviation, nuclear power, the handling of pathogens — rests on the party that would deploy the hazard, and in this field alone it has been placed on the public;

And whereas humanity has, on prior occasions and under harder conditions, chosen to defer a capability it was fully able to build, and has not, on reflection, regretted the delay;


Now, therefore, the undersigned call upon the governments of the world, acting together, to adopt the following measures without further delay.

Article I

A halt at the frontier.

No training run should proceed beyond the capability level of today's most advanced deployed systems until independent, publicly specified safety criteria have been met. The limit should be defined in terms that can be measured and audited — computation, capability evaluations, or both — and should bind every developer equally, public or private, in every jurisdiction.

Article II

Licensing before, not review after.

Frontier development should require a licence issued in advance, as it does for a new medicine, a new reactor, or a new aircraft. The applicant should carry the burden of demonstrating safety. An absence of known harm is not a demonstration of safety, and self-certification is not an audit.

Article III

Accounting for compute.

Advanced AI training hardware should be registered, tracked, and reported at scale, on the established model of safeguards for fissile material. Computation of this kind is physical, concentrated in a handful of firms, and passes through a supply chain of a few chokepoints. It is among the most countable dangerous inputs humanity has ever regulated.

Article IV

An inspectorate with standing.

Verification should be carried out by an international body with a right of access to training facilities and hardware, empowered to confirm compliance and to publish its findings. Restraint that cannot be verified will not be trusted, and restraint that is not trusted will not hold.

Article V

Conditional, and reviewable.

These measures are a sequencing decision, not a prohibition. They should be reviewed on a fixed schedule against published criteria, and lifted in step with demonstrated progress on alignment, interpretability, evaluation, and security. The slowdown ends when the science catches up. That is the point of it.

We do not ask that artificial intelligence be abandoned. We ask that its most dangerous capabilities not be built before anyone knows how to control them. The order in which we do these two things is not a detail. The order is the whole question.

Convened and maintained by the Center for AI Safety · Open to signature by scientists, policymakers, civil-society leaders, and citizens

Article  II

Two terms, defined

Who asks, and for how long

The campaign's demand fits in four words — slow down AI development — and its reason in three: we want time. Two words in those sentences carry the weight, and both deserve a precise answer.

Who is "we"?

The public at large.

Not only the researchers, though many researchers have signed. The register is open to four categories of signatory, and the largest of them — citizens — requires no qualification at all. The engineers' warnings began this conversation; it is not theirs to finish. The people who would bear the consequences of an uncontrolled system are entitled to a view on its schedule, and this register exists to record that view.

How much time?

Until safety demonstrably keeps pace.

The statement names no date, because a date is the wrong instrument. Article V sets the terms: the measures are reviewed on a fixed schedule against published criteria, and lifted in step with demonstrated progress on alignment, interpretability, evaluation, and security. The slowdown ends when the science catches up. That is the point of it.

Article III · The register

Put your name to it.

A statement is a measurement of who is willing to be counted. Signatures are counted by category and country; individual names are never published, sold, or shared without your separate and explicit instruction.

The register is read by the people it is addressed to. Legislative staff, ministries, and journalists ask us one question more than any other: how many, and who? So the register is built to answer it — every entry verified, counted in its category, and published as a running total by country.

  • Scientists & engineers Open
  • Policymakers & public officials Open
  • Civil society, faith & labour leaders Open
  • Citizens Open

Signatures in the policymaker and scientist categories are checked against the affiliation given before they are counted. Everyone else is counted on confirmation of their email.

Sign the Slowdown Statement

Form no. SD–01 · Register of signatories

By signing you affirm that the details given are your own. The Center for AI Safety is the data controller for this register. Your email is used to confirm your signature and, only where you have ticked the box above, to send campaign updates or to introduce you to allied organisations. Nothing is shared otherwise. You may withdraw at any time.

Entered into the register

Thank you.

Your signature has been recorded. A confirmation has been sent to the address you gave; it must be opened for the entry to become final.

Reference
Category
Received
Status
Awaiting email confirmation
Register
The Slowdown Statement

The second thing you can do takes four minutes. A signature is a count; a letter to your legislator is a constituent. Write to your representative →

Article  IV

Proceedings of the campaign

The rest of the record, in four parts

The statement and the register live on this page. Everything else the campaign publishes is filed under four headings.