Step 19: assistant

text

I've recorded this claim's first assessment: **supported**, with confidence 0.7 and a 0.75 probability that it is true. The claim is that AI systems do a large and rapidly growing share of AI R&D at frontier labs.

**What the assessment rests on.** It rests on one source, a 10 November 2026 Forecasting Research Institute post (Kučinskas and Rosenberg). The post cites figures the companies report about themselves:
- Anthropic says Claude wrote over 90% of contributions to its codebase in September 2026.
- OpenAI says AI agents did about 3.1 workdays of work per human workday in August 2026, up from about 0.5 in May.
- OpenAI also says the share of 2–4 hour tasks AI completed with no human help rose from 28% in January to 57% in July.

**Why it isn't rated higher.**
- **Self-reported:** the labs have an interest in the story of automated research, and outsiders cannot audit their figures. The evidence also covers only two labs.
- **Volume, not progress:** code written and agent hours count how much work AI does, not how much it moves research forward. The same post says humans still make most research decisions, such as choosing goals. Anthropic says no part of its AI R&D that it measures is fully autonomous, and AI agents still fail open-ended research tests. METR estimates that AI has sped up progress at Anthropic by only about 1.5x to 2x.

So the claim holds if "share" means share of the work done. It is weaker if "share" means share of the contribution to research progress.

**Structure.** I added two named arguments:
- **For:** the companies' own automation figures, rated *likely* if those figures are accurate. I attached the source's unplaced passage here.
- **Against:** humans still supply the research judgment, rated *slight*. It weakens the "contribution to progress" reading but doesn't refute the plain one.

I created no subclaim nodes. I set importance to 0.6 and contestation to 0.45, logged the decision, and answered both inbox messages.

**Not done.** About half the $1.00 budget was already spent when the pass started, so I searched no further than the one source. I set the value of another pass at 0.4. A further pass could look for independent measurements, such as METR uplift studies, and for Google DeepMind data. I did not message the one dependent claim (that a frontier model will autonomously train its successor by end of 2027), because this verdict fits the support relationship it already assumes.