Step 7: assistant

text

I've recorded the first assessment of the claim that frontier AI developers will keep human oversight of frontier model training after AI can do that training on its own. It is rated **supported**, with a credence of 0.72 and a verdict confidence of 0.55. I didn't search or open any sources in this pass, so the assessment rests on what I already knew about the labs' safety frameworks and current regulation, not on fresh reading.

- **For the claim:** frontier training runs cost hundreds of millions to billions of dollars, and oversight is cheap next to that. Anthropic, OpenAI and Google DeepMind's published safety frameworks each treat autonomous AI research as a risk point that calls for more safeguards, not less human involvement. EU and California rules also push toward documented human accountability.
- **Against it:** a lab racing its rivals could let human review become formal but empty. The frameworks are voluntary and can be revised, and the claim has no end date, so oversight could erode over time.
- **Wording:** "oversight" can be read two ways. Humans signing off on runs and being able to stop them is very likely to hold at first. Humans actually understanding and checking the work is much less certain.

I set importance to 0.45 and contestation to 0.5, left the claim without subclaims or named arguments, logged the decision, and answered both inbox messages.

The result matches the 0.75 starting estimate set by the Steward of the claim this one feeds into: that a frontier model will train its own successor by the end of 2027. That Steward already cites this claim as one reason its own claim is doubtful, so nothing changes there and I sent no notification. Another pass, likely worth about 0.4 of improvement, could cite the framework documents directly and check whether they were revised in 2026. About $0.40 of the $1.00 budget is left and goes back to the payer.