OpenAI has engaged the independent AI research organisation METR to assess a training run the company describes as its largest reinforcement-learning run to date. METR will set its own assessment priorities and receive access to model snapshots, monitoring records and training transcripts, putting external scrutiny inside the training process rather than after the fact.
OpenAI Watch analysis
What happened
OpenAI says the agreement is limited to one training run. METR will investigate behaviour that diverges from developers’ intentions, including failures OpenAI’s own tests and monitoring may have missed.
The arrangement is part of OpenAI’s embedded-assessment initiative. The company announced it on 16 November in its account of the first phase of embedded assessments.
Why it matters
A reviewer can only test what it can inspect. Access to snapshots, monitoring records and training transcripts gives METR more than a polished model demo, while the right to choose assessment priorities gives it room to probe beyond the questions OpenAI might naturally ask itself.
That does not establish that the review will find a problem, or that any findings will be made public. But assessing behaviour during a major training run could reveal issues that are harder to spot in a finished model. The details of what METR examines, and what OpenAI publishes afterwards, will matter as much as the announcement.
Our read
This is a concrete step towards outside assessment of a high-stakes training process, not a safety certificate. The useful parts are the evaluator’s stated independence in setting priorities and its access to records from the run. The limit is equally plain: one agreement covers one run, and the announcement does not say what results will be released. Independent scrutiny is welcome; readers should still ask to see what it turns up.
What to watch
- What behaviours and failure modes METR chooses to assess.
- Whether METR publishes findings, and how much supporting detail is disclosed.
- Whether OpenAI extends embedded assessments to other training runs.
Discussion spark: Should independent assessors be able to publish their findings from a training run, even when those findings are uncomfortable for the company funding the work?
Sources and evidence
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.