AI development is moving toward substantial automation while humans continue setting research directions and judging results.

The claim joins two parts: that AI development is becoming substantially automated, and that humans still set research directions and judge results. As a description of the current trajectory, both parts are favored by the evidence, though much of it comes from AI developers themselves.

On automation, public measurements show rapid growth in what models can do without help. METR's measurements of task length suggest the length of tasks models can complete on their own has been doubling every few months. Coding and research-reproduction benchmarks such as SWE-bench and CORE-Bench have saturated. Anthropic reports that by May 2026 more than 80% of code merged into its codebase was written by Claude, and that output per engineer has risen sharply. On human direction, the same report says models match skilled humans at running well-specified experiments but still fall well short at choosing goals and judging which problems are worth working on. That gap is what leaves humans setting directions and evaluating results.

The weaknesses are these. The internal figures come from a party that profits from its models being seen as capable, and nobody outside has examined them. Lines of code overstates productivity, as the report itself says. The second part of the claim may also be temporary. Some forecasters expect research taste itself to be automated within a few years, while others expect automation of AI research and development to stall. So the claim is well supported as a description of the present and near term. How long humans keep the direction-setting role is the open question.