No frontier AI system should be deployed to the public until an independent body has examined it, and said so on the record.
No frontier laboratory should be allowed to grade its own homework on systems that could destabilize labor, information, finance, or national security. Independent auditors, accredited and liable, must review training data provenance, evaluation results, and known failure modes. Findings must be filed in a public registry. Self-attestation is not safety.
Today, the company that builds a frontier AI system is also the company that decides whether it is safe to release, the company that decides which tests to run, the company that decides which results to publish, and the company that decides which results to keep internal. There is no outside check.
In every other high-stakes industry, this arrangement is understood to be unacceptable. Aircraft are not certified by the airlines that fly them. Drugs are not approved by the firms that sell them. Bridges are not signed off by the contractors who pour the concrete. The principle is older than the modern state.
Artificial intelligence is, at present, the only technology of comparable consequence that is permitted to grade its own homework.
An independent audit is not an attack on innovation. It is the mechanism by which a society confirms that what is being released into it has been examined by someone who does not stand to profit from its release.
An audit asks: What can this system do that its creators did not intend. What are its failure modes under realistic adversarial use. What populations does it disadvantage. What categories of harm has it already produced in testing. What safeguards have been bypassed. What was found, and not fixed.
The answers belong to the public, because the public is the testing environment.
A standing independent body, staffed by technical reviewers without commercial ties to the labs being audited, accountable to a legislative oversight committee, and required to publish its findings in plain language.
This is a question of institutional design, not of impossibility. Comparable bodies already exist for aviation, pharmaceuticals, finance, nuclear materials, and food. None of them are perfect. All of them are better than nothing.
- 01.Mandatory third-party safety evaluation before any frontier model is released to the public or to commercial customers.
- 02.Pre-deployment red-team testing by reviewers without equity, contracts, or future employment ties to the developer.
- 03.Public disclosure of audit findings, including failures, in plain language.
- 04.Authority to require remediation or to block deployment until risks are addressed.
- 05.Adequate funding for the audit body, independent of the industry it regulates.
- —Every release will continue to be graded by the entity that profits from the release.
- —Critical failures will be discovered by users, journalists, and adversaries, not by reviewers.
- —The public will continue to learn what a system can do only after it has done it.
- —The first serious accident will be used to justify rules far more restrictive than the ones being asked for now.
- Objection
- Answer
- Objection
- Answer
- Objection
- Answer
"Auditors cannot keep up with the pace of releases."
Then the pace of releases is the problem, not the principle of review. No other high-stakes industry accepts speed as a reason to skip inspection.
"Companies already do internal safety testing."
Internal testing is necessary. It is not sufficient. Internal teams report to the people whose product they are reviewing. That is, by definition, not independent.
"Audits will leak trade secrets."
Established frameworks already exist for confidential review in finance, defense, and pharmaceuticals. The problem is solved. The objection is rhetorical.
A society that cannot inspect what is being deployed into it has, in practice, stopped governing it.
Human Future Framework · Independent Oversight
- → Generate a call/email script tailored to this demand, then phone or write your officials.
- → Find your representatives and contact them about this demand.
- → Sign the Human Future Pledge to add your name (or anonymity) to the public record.
- and log the share below.