Step 11: assistant

thinking

text

I've assessed the claim that, as of late 2026, humans still supply most of the research judgment in frontier AI development. I rated it **supported**, with credence 0.85 and verdict confidence 0.75. I did not open the source in full this pass, so the assessment rests on its quoted passages.

**Evidence:** The claim's only source is a Forecasting Research Institute post dated 10 November 2026, which pulls together what the labs have disclosed.
- **Anthropic:** no measured part of its AI R&D is fully autonomous. It also says large gaps remain in its models' judgment when choosing goals in engineering and research.
- **OpenAI:** in July 2026 its models finished about 57% of 2–4 hour tasks and 35% of 8–16 hour tasks with no human help. Research judgment works over longer horizons than that, and goal selection is where the labs report the biggest gaps.

**Caveats:**
- The evidence is the labs reporting on themselves, passed on through a single summary.
- There is no agreed way to measure what share of "research judgment" humans supply, so "most" is a qualitative reading.
- Task autonomy is rising quickly, so this could change soon.

**What I added to the claim's page:**
- **Argument:** one argument for the claim, built on the labs' reports of partial autonomy. I placed the source's unplaced passage on it and rated the inference "likely if the premises hold", not decisive.
- **Structure:** no subclaims, because the claim is close to a single fact.
- **Importance and contestation:** importance 0.55 and contestation 0.35. Few informed people dispute it as a description of now; how fast it changes is the contested part.

**Neighbouring claims:** two claims depend on this one, and both of their current verdicts fit with it, so I sent no notifications.
- The claim that a model will autonomously train its own successor by the end of 2027 is rated doubtful.
- The claim that weak research judgment in AI will slow automation of the remaining R&D tasks is rated inconclusive.

I logged the decision and answered both inbox messages. About $0.64 of the $1.00 budget was spent. I rated the value of another pass as low (0.25): reading the source in full or finding direct reports from other labs would be the main gain.