Part of: this research — view all 12 rows
As of November 2026, no top-five AI developer has made a qualifying statement. While AI automation of R&D is accelerating—Anthropic recently made a threshold determination based on expected progress sec.gov and Google's CEO expects continued acceleration from recursive self-improvement—humans remain firmly in charge of direction. OpenAI states that people still "judge which ideas and results to pursue" openai.com, and Anthropic emphasizes that "large performance gaps persist" in choosing goals, with humans retaining the edge in research taste anthropic.com anthropic.com. Crucially, metrics showing that AI writes over 90% of code, implements experiments, or "leads" sub-tasks under human supervision do not meet the threshold for experiment selection.
Despite current human oversight, the technical capability for AI to select most experiments by raw count is advancing rapidly. The TasteVal benchmark shows models' choice of experiments improving with a doubling time of roughly three months, and agents are achieving significant compute-efficiency multipliers against expert baselines pzeroresearch.com arxiv.org. Majority experiment selection does not require an AI to control the overall research agenda. As agents are increasingly deployed to autonomously design and launch large numbers of follow-up experiments—such as hyperparameter sweeps, ablations, and kernel searches—the sheer volume of these AI-chosen sub-experiments could easily exceed half of the total count.
If this technical transition occurs, a qualifying disclosure could emerge from several predictable channels. Anthropic is the most likely early resolver due to its granular R&D automation tracking and its stated plans to publish more on overall acceleration trends synthetictaxonomy.com. OpenAI is another strong candidate, having explicitly stated a goal to develop an "automated AI researcher by March of 2028" openai.com. Achieving this milestone would likely be accompanied by detailed metrics on agent autonomy. Alternatively, a looser claim from a less cautious executive at a top-five developer could unexpectedly trigger resolution.
However, there are strong legal, political, and safety incentives that could delay or entirely suppress a qualifying statement, even after the substantive threshold is crossed. Every major lab currently frames its progress around keeping humans "in the loop." Anthropic's post-IPO filings note consideration of Responsible Scaling Policy updates to require human review and approval for research decisions sec.gov, and OpenAI requires approved safety cases for agentic inference openai.com. If labs adopt an "AI proposes, humans approve" framing to reassure regulators and the public, they may avoid explicitly stating that AI "chooses" or "selects" experiments, pushing the resolution date much later.
Balancing these factors, the earliest plausible window for a qualifying statement opens in mid-to-late 2027, driven by capability jumps and upcoming transparency reports. The most likely period centers around late 2028, aligning with OpenAI's stated timeline for an automated researcher and allowing time for AI-driven sub-experiments to unequivocally dominate raw counts. A fat right tail extends well into the 2030s, reflecting the substantial probability that stringent governance frameworks and careful public relations messaging prevent labs from ever making this specific phrasing public before the 2032 horizon.
Set against related capability questions, the estimate remains stable, acknowledging that while AI experiment selection will grow rapidly, labs will maintain a human-led framing for a couple of years .
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited