Make scrutiny independent.
Let qualified outsiders challenge the claims, and report what they find.
Prevents A builder choosing its own examiner until it gets the answer it wants.
Proposed
What it asks
Builders should enable qualified outside scrutiny, with access proportionate to the claims and risks being examined. For higher-risk systems, the builder should not be able to make an unfavourable assessment disappear by choosing a different evaluator.
Independence needs an operating model. Who chooses the examiner? Who pays? Can the examiner run its own tests? Can it report a serious concern to an authorised oversight body without the client’s permission? A company-funded audit can still be useful, but these questions determine how much confidence it deserves.
Access to evidence and the freedom to publish findings are meaningful safeguards in their own right. They are not the same as authority to require a response. An assessment should make a decision better informed; it should not give its author unchecked power to stop work. The responsible authority, its duties and the route to challenge its decision must be specified too.
Public accountability does not require publishing private training records, security vulnerabilities, or every model artefact. Sensitive evidence can be examined through controlled access, with public summaries explaining the scope, findings, and unresolved concerns. Restrictions should have stated reasons and a route to challenge them, because commercial confidentiality can otherwise become a permanent answer to every difficult question.
Evaluators should also face scrutiny. Their methods, conflicts of interest, and material errors need examination. Independent assessment improves the basis for a decision; it cannot turn uncertainty into a guarantee.
Proposed
What compliance looks like
- A published assessment scope and disclosed conflicts of interest.
- A route for evaluators to report adverse findings.
- Public explanations of concerns that remain unresolved.
Illustrative
The case that tests it
An outside team finds a serious risk. What happens next?
Two labs find a reason to spend longer testing. Neither wants to be the only one that waits. A government may see leadership as a source of security, prosperity and influence, and slowing down alone could surrender those advantages without reducing a rival’s risks. But a failure can harm people across borders, including countries that had little say. Shared checks and agreed responses could make it easier to wait without acting alone.
An outside team finds a serious risk. What happens next?
These arrangements can work together. The difference is whether a finding comes with an agreed duty to respond.
Neither arrangement guarantees safety. Some companies or states may refuse to participate, and checks can miss hidden activity. People outside the leading companies and countries need a voice in setting the terms.
Our inferenceThe test of any institution is whether it can obtain evidence, act on it, resist pressure, and be held accountable itself.
Reported
What the case draws on
Provider’s commitment
A concrete proposal, not just a call to slow down.
- Reported
- September 2026
Dario Amodei committed Anthropic to bringing outside reviewers inside the company, with continuing access comparable to internal risk teams and the right to publish key findings without the company’s editorial control, subject to limited redactions for security, legal privilege, commercial sensitivity and third-party confidentiality.
These are announced intentions for an arrangement to be established, not proof of its implementation. They do not establish that reviewers have stop-work authority, or that a binding agreement between companies has been signed. Amodei also couples pacing with preserving a US and allied lead over China; that is his position, not a condition this proposal places on who deserves protection.
Existing institutions
What international oversight can build on.
The IAEA’s safeguards rely on technical verification under agreements accepted by states. An AI institution would likewise need a basis for access, a defined remit and legitimate authority.
The UN established an AI scientific panel and global dialogue in August 2025. Assessment and discussion are valuable functions; they do not themselves confer global enforcement powers.
IAEA: Basics of IAEA safeguards
United Nations: AI scientific panel and global dialogue
Reported
On the record
- Legislation signed
Outside scrutiny enters a legal framework.
California signed SB 813 and AB 1405, establishing an independent-verification framework and a registry with standards for AI auditors.
The governor’s announcement records legislation, not proof that oversight works in practice. It does not establish influence by this proposal.
Proposed
Where it can fail
- Evaluators can be captured, mistaken, or selected for favourable conclusions. Their independence needs scrutiny too.
- Restricted access programmes need transparent criteria and a route to challenge exclusion.
- Outside review will leak secrets.
- Sensitive evidence can be examined under controlled access, with public summaries of scope, findings and unresolved concerns. Restrictions need stated reasons and a route to challenge them.
- Reviewers would gain unchecked power to stop work.
- An assessment should make a decision better informed, not hand its author a veto. The responsible authority, its duties and the route to challenge it must be specified too.