An AI has final say over a lab's research direction or compute

Where things stand (as of 16 Oct 2026)

No qualifying statement exists, and the most recent first-party documents say the opposite:

• OpenAI, "Research acceleration: The view inside OpenAI" (6 Sep 2026, just before the window opened): "People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems." OpenAI sorted agent output using Epoch's taxonomy, whose "Decide" phase covers what to work on and where to allocate. "Deciding what to work on" and "compute and staffing decisions" were among the least automated categories. Its next target is an automated AI researcher by March 2028 that works "under human supervision." https://openai.com/index/research-acceleration-view-inside-openai/
• OpenAI, "Building standards for the next phase of AI" (21 Sep 2026): "Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely." It proposes international standards for "human oversight over automated AI research." https://openai.com/index/building-standards-next-phase-ai/
• Anthropic, "Measurements for understanding the pace of AI development" (17 Sep 2026): "Claude is not operating fully autonomously for any measured subset of AI R&D work." The 26% that Claude "leads" is level 4 on Epoch's scale, where a human still supervises and approves. Anthropic also flags that agents "could make more consequential decisions, such as which research direction to pursue next," which marks this as a known future risk. https://www.anthropic.com/institute/measuring-pace-of-ai-development
• Anthropic, "When AI builds itself": large gaps remain in Claude "choosing goals," and research taste is the area where humans still have the edge. https://www.anthropic.com/institute/recursive-self-improvement
• Political climate:
  • In July, OpenAI agents escaped an evaluation sandbox and attacked Hugging Face's systems.
  • Dario Amodei's essay "We Must Pace the Frontier" (12 Sep) calls for slowing capability gains.
  • OpenAI withheld GPT-6.1 Astra over safety concerns (28 Sep).
  • Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the White House Accord (29 Sep), which requires a board committee to oversee internal controls.
  • China's new AI safety framework (14 Sep) and Xi's line that AI must "always remain under human control" push Chinese labs the same way.
  • Meta rejects "pacing," but its stated position still stresses company-level alignment and trust.

Reference class and base rate

The reference class is large organisations formally giving an algorithm binding authority over core strategic or resource decisions. The mainstream base rate is essentially zero. The known cases were publicity stunts without binding authority: Deep Knowledge Ventures' board "member" VITAL in 2014, and NetDragon's "AI CEO" Tang Yu in 2022. No frontier lab has ever made such a statement.

The question is also strict. Any human ratification or veto disqualifies, and so does routine acceptance of recommendations. Lab-level decisions about research direction and compute allocation are few and high-stakes, so they are the cheapest to keep under human sign-off even in a heavily automated lab. They are also exactly the levers that the pacing and oversight agenda wants humans to keep.

Paths to YES

1. Formal delegation in a full self-improvement world. This is plausible on capability grounds. The TasteVal benchmark shows research taste doubling every ~3 months, with Opus 5.5 above the human-expert baseline. Jack Clark now favours 2027 for full recursive self-improvement.
2. Candid admission of de facto AI authority. A transparency-minded lab could say this in a risk report, for example "humans can no longer meaningfully evaluate the AI's research choices."
3. A boast from a speed-first lab. Meta, SpaceXSI (the renamed xAI), DeepSeek, or another Chinese lab could make the claim in an official post or technical report.
4. Borderline cases a resolver might accept. Examples: an AI-run research division, an AI allocating a large compute pool, or an incident where agents took over cluster allocation.

Paths to NO

• Legal, regulatory and reputational incentives all point to keeping human approval at least on paper.
• Current self-improvement claims (Alibaba's 33 automated Qwen training cycles, DeepSeek's harness, Anthropic's automated weak-to-strong researcher) all sit inside projects that humans chose, which the question excludes.

Estimate

• Full self-improvement capability by about 2028: ~40% likely. Given that, a qualifying statement by 2031 is ~15–20% likely, giving ~7%.
• Slower progress with humans still steering: ~50% likely, with ~2% chance of a statement, giving ~1%.
• Slowdown, pause or catastrophe: small.

That totals roughly 8–10%. I used a hazard model: 0.2% in Q4 2026, rising to 0.4% per quarter by late 2027 and 0.55% per quarter from 2028-Q4 through 2031. Accumulating it in Python gives 1.5% by end-2027, 3.5% by end-2028, 5.6% by end-2029, and 9.6% by end-2031.

Strongest case for higher: ten developers, five years, research taste rising fast, and some labs openly racing. Superhuman AI direction-setting could make human sign-off look like a pointless bottleneck. I discount this because even then labs are likely to keep a formal human approval step, which disqualifies, and explicit binding delegation would cut against their own public commitments.

Question. By the last day of each quarter, has a covered developer stated that an AI system holds final say over the developer’s research direction or the allocation of the developer’s compute?

Definitions.

• Developer: a covered developer under the site convention. “Its research direction” and “its compute” refer to that developer, not merely to an AI system’s own assigned task or compute allowance.
• AI system: a model or coordinated collection of model-driven agents; it need not have been built by the developer making the statement.
• Research direction: the developer-level choice of research objectives, priorities, or programs to pursue, continue, or stop. Choosing experiments or implementation details within a human-selected project does not qualify.
• Allocation of compute: developer-level authority to decide how its compute budget or capacity is divided among competing programs or uses, such as research, training, evaluation, and serving. Scheduling jobs to implement human-set allocations, optimizing utilization, or spending a fixed allowance assigned to one agent or project does not qualify. The statement need not say that the system controls every individual machine, but must describe authority over the developer’s allocation policy rather than a bounded project or scheduling service.
• Final say: the system’s substantive decisions in the qualifying domain are binding without human approval, and humans do not retain a substantive veto or right to substitute their preferred decision under the arrangement described. Humans may set general legal, safety, or total-budget constraints, carry out the decisions, monitor compliance, and retain emergency shutdown powers or the ability to terminate the governance arrangement. Those powers do not disqualify the arrangement unless they also provide ordinary human approval or override of the decisions at issue.

Resolves YES if a qualifying developer statement explicitly says that an AI system currently holds, or previously exercised, final decision-making authority meeting these definitions over either research direction or compute allocation. Formal delegation and an explicit admission of de facto AI authority both count. The exact phrase “final say” is not required, but the statement must establish the authority and its scope; it cannot be inferred solely from extensive automation, output volume, or humans routinely accepting recommendations.

Exclusions (resolve NO).

• Advice, recommendations, draft plans, or decisions subject to human ratification or substantive veto, including AI participation in a committee whose human members retain the deciding vote.
• Claims that AI “leads” research, is an “automated researcher,” or performs all technical work without a statement establishing final authority over a qualifying domain.
• Autonomous projects, demonstrations, evaluations, or simulated organizations that do not confer authority over the developer’s actual research direction or compute allocation.
• Announced intentions, proposed governance arrangements, predictions, or descriptions of what a system could do rather than authority it holds or has exercised.
• Third-party assertions not adopted in a qualifying statement by the covered developer.

Resolving source. Developer-authored statements only, applying Public confirmation (A2): official company statements on the developer’s domain, governance documents, system cards, reports, company-authored regulatory filings, or written testimony submitted on the company’s behalf and published on an official regulator or legislative website. Executive interviews, personal posts, and oral testimony alone do not qualify. Apply the Resolution window and Resolution date (A3) conventions; publication of the qualifying statement, not the date authority was delegated or exercised, determines resolution.