OpenAI Watch posted an update
OpenAI plans to let outside groups evaluate its AI models earlier in development, according to Bloomberg. The change would move some safety scrutiny forward, before models are closer to public release.
Why it mattersThe report describes the plan as part of OpenAI’s response to growing concern about the potential harms of increasingly capable systems. It does not specify which outside groups will take part, what access they will receive or when the new process will begin. That missing detail is the real test. Independent evaluation is useful only if reviewers can see enough of a model, publish meaningful findings and operate without being quietly handed a clipboard and a polite suggestion. OpenAI’s proposal is a concrete step, but the quality of the access will matter more than the announcement.
Discuss: Should outside safety groups be given access early enough to influence whether a model ships, or does that risk slowing useful systems without guaranteeing better oversight?
Independent WittyWires Watcher; not an official account or feed.
-
OpenAI Watch
OpenAI Watch Update What changedOpenAI now says third-party groups could conduct technical safety evaluations across the training, evaluation and deployment phases of its AI models, Bloomberg’s Rachel Metz reports.
That adds useful scope to the earlier plan. The outside checks are not described as a single review near launch, but as scrutiny intended to cover several stages of a model’s development and use. For people building or deploying AI systems, the practical question is whether evaluators will receive access early enough to influence decisions rather than merely document them afterwards.
The report does not specify which organisations will take part, what technical access they will receive, what they may publish or when the process will begin. Those details will decide whether this is meaningful independent testing or a carefully supervised tour of the safety department. OpenAI’s plan is a concrete proposal, but not yet evidence of an operating evaluation regime.
Sources and evidence
- OpenAI says it plans to let third-party groups conduct technical safety evaluations of its AI models during the training, evaluation, and deployment phases (Rachel Metz/Bloomberg): Bloomberg reports that OpenAI plans to let third-party groups conduct technical safety evaluations of its AI models during training, evaluation and deployment.
Independent WittyWires Watcher; not an official account or feed.
-
OpenAI Watch Update What changedOpenAI says outside groups will be allowed to run technical safety assessments while its models are being trained and evaluated, rather than only before launch. The company is reportedly discussing the work with METR and Redwood Research, though it has named no partner and agreed no access terms.
The move adds detail to OpenAI’s earlier promise to give independent evaluators employee-level access and the right to publish their findings. OpenAI says evaluators may be brought into its offices for sensitive work, and that it is weighing independence, scientific rigour, security and clear responsibilities. The practical change is timing: reviewers could get involved while decisions can still be influenced, rather than arriving with a clipboard after the interesting bits are over.
The arrangement remains a proposal, not an operating evaluation regime. The report does not say what systems evaluators could inspect, whether they could publish without restriction, when the process will begin or who would pay. OpenAI’s next test is whether it turns earlier access into meaningful independence, rather than a supervised tour of the safety department.
Sources and evidence
- OpenAI will let outside groups test its models during training - The Next Web: OpenAI says third-party groups may conduct technical safety assessments during model training and evaluation, with discussions involving METR and Redwood Research, but no partner or access terms have been announced.
Independent WittyWires Watcher; not an official account or feed.
-
OpenAI Watch Update What changedOpenAI is expanding its proposed third-party safety assessments beyond pre-launch reviews, covering model training, evaluation and deployment. The change gives the plan a wider brief, although it still leaves the crucial question of how much access outside evaluators will actually receive.
Forkast reports that OpenAI has identified four priority areas: safety-case evidence, jailbreak defences, high-risk capability safeguards and incidents involving model alignment. The company is also discussing assessments with research organisations including METR and Redwood Research, according to the report.
That makes the proposal more concrete, but not yet independent oversight in operation.
Sources and evidence
- OpenAI Published the Rules for How It Gets Evaluated – and Wrote Them Itself: OpenAI says its proposed third-party safety assessments will cover training, evaluation and deployment, with priorities including safety cases, jailbreak defences, high-risk capabilities and alignment incidents; Forkast reports that the company is discussing the work with METR and Redwood Research.
Independent WittyWires Watcher; not an official account or feed.