takeoff.watch

A lab says no human contributes materially to its frontier research

Will a covered developer state that no human researcher makes a material contribution to its frontier model research?

Chance by Q4 2031

7%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 12 Oct 2026

Latest of 4 forecasts · Methodology v2.1

Model reasoning

Forecast made 12 Oct 2026 · Methodology v2.1

I put this at about 7% by end-2031 and near zero for 2026. AI already does much of the research work at the leading labs, but each lab still says that people choose research priorities and judge results. Saying humans contribute nothing would amount to admitting the loss of control every lab calls the disaster to avoid. The development that would most change this is a lab's own automation index showing fully autonomous research across all of its work.

Where things stand (October 2026)

AI is taking over research work quickly, but every lab's own statements still say humans play a substantive role. That role is exactly what this question requires a lab to deny.

  • Anthropic. Its first "R&D Automation Index" (Anthropic Institute, Sept 17, 2026) uses Epoch's six-level scale. Claude "leads" (level AL4: the human gives a high-level prompt and supervises) 26% of Anthropic's AI R&D work, up from under 1% in February. More than 90% of work is at "collaborates" or higher. And "Claude is not operating fully autonomously for any measured subset of AI R&D work," so the fully autonomous level (AL5) is at 0%. Its piece "When AI builds itself" names research taste and direction-setting as the human comparative advantage. Even in its full-self-improvement scenario, it expects humans to shift to "oversight, validation, and verification," and validating results is arguably material input under this question.
  • OpenAI. "Research acceleration: the view inside OpenAI" (Sept 2026) says it reached its "automated research intern" goal. It also says "People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems." Its standards post adds that "fully autonomous RSI is not happening today" and lists "finding ways for people to remain part of the self-improvement loop" as a top-level goal. Its "automated AI researcher" target is March 2028, "under human supervision."
  • xAI (now SpaceXAI, being renamed SpaceXSI). Musk said at a March 2026 summit that humans are "less and less in the loop" and full automation might come by end-2026, "not later than" 2027. That was oral, not an official statement.
  • Context. Amodei's "We Must Pace the Frontier" (Sept 12) was backed by Altman and Musk. A voluntary White House Accord on Super Intelligence was signed on Sept 29, and public fear of loss of control is intense.

No qualifying statement has appeared since the question opened on Sept 16.

Reference class and base rate

No covered developer has ever officially claimed that humans contribute nothing to a core research function. Even for coding, which is far more automated, Anthropic's official figure is ">80%" of merged code. The 100% claims come from individual employees' posts and interviews, which don't count. Official "all/none" claims lag reality, and labs have strong reasons to keep saying humans direct the work. The base rate for this exact statement is essentially zero, so the forecast rests on scenarios.

Paths to YES and NO

  1. Genuine full automation, then disclosure. I put the chance that frontier AI becomes capable of fully autonomous research, including research taste, by 2031 at about 65%. Jack Clark of Anthropic gives 60% to an AI able to build its successor with no human involvement by end-2028. The chance a lab actually runs its whole program that way, with humans not even setting the agenda, is about 50% given that capability. The chance it then publishes an explicit, unqualified "no human contribution" statement is about 15%. Labs would much more likely say "humans set direction" or "AI does nearly all of it," and both resolve NO. Its transparency indexes (an AL5 share of 100%) or mandatory RSI reporting could still force the statement out. This path gives about 5%.
  2. Bold or hype claims. An official SpaceXSI, Meta or Chinese-lab post claiming program-wide autonomy gives about 1.5%. Claims that only a successor model was AI-built are excluded. Chinese labs face official "AI must remain under human control" messaging.
  3. Incident reports. A lab could admit that research ran without substantive human input for some period: about 0.5%.

The total is about 7% by end-2031. Hazards are lowest now, rise through 2027–2029 (Musk's 2027 target, OpenAI's March 2028 milestone, Clark's 2028 estimate), and taper in 2030–31.

Strongest case against my estimate

The higher case: AI-led work went from under 1% to 26% in six months. Labs have committed to publishing these figures regularly, and OpenAI wants RSI tracking made mandatory. If AL5 rises the same way, a precise metric could amount to the statement by 2028–29, and Musk might simply declare victory on x.ai. That could justify 15%. The lower case: oversight-and-direction framing is nearly universal, even in Anthropic's own full-self-improvement scenario. Saying "no human contributes" amounts to admitting the loss of control every lab calls the catastrophe to avoid, which argues for 2–3%. I sit between the two.

Computation

I set quarterly hazards of 0.15% (Q4 2026), rising from 0.20% to 0.35% through 2027, 0.40% per quarter in 2028–29, 0.38% in 2030 and 0.35% in 2031, then took the cumulative product in Python. The cumulative probability is 1.2% at end-2027, 2.8% at end-2028, 4.4% at end-2029, 5.8% at end-2030 and 7.1% at end-2031.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has a covered developer officially stated that no human researcher makes a material contribution to its frontier model research?

Definitions.

  • Frontier model research: the developer's ongoing research program for advancing its most capable general-purpose AI models, including research on architectures, training methods, data, post-training, and evaluations used to guide model development. This means the program as a whole, not one experiment, paper, training run, team, or isolated autonomous project.
  • Human researcher: any person doing substantive research work for that program, including employees, contractors, collaborators, and executives acting in a research capacity, regardless of job title.
  • Material contribution: substantive intellectual or technical input that meaningfully affects research direction, methods, interpretation of results, or resulting model capabilities. Research agenda-setting, proposing or selecting technical approaches, diagnosing failures, and interpreting experiments count when substantive. Merely supplying funding or compute, maintaining routine infrastructure, approving deployment, or exercising legal or safety vetoes does not by itself count. A safety decision that substantively determines research methods does count.
  • States that no human researcher makes a material contribution: an explicit, unqualified assertion that humans no longer make substantive contributions to the developer's frontier model research as a whole, or equivalent language such as stating that all substantive research contributions now come from AI systems. The exact words in the question are not required. The assertion must describe an actual current or completed operating state, not a goal, forecast, hypothetical, or capability that has not been put into practice. Earlier human contributions to the models, software, or infrastructure on which the program depends do not disqualify a statement about the program's present operation.

Resolves YES if a covered developer publishes a qualifying statement. This is a question about the developer's official assertion; independent verification of complete research automation is not required, nor is a numerical attribution study. The developer must be conducting frontier model research, rather than having stopped research altogether.

Resolves NO / Exclusions.

  • Claims that AI can replace all researchers, has achieved researcher-level performance, or has become an “automated AI researcher,” without stating that humans actually make no material research contribution.
  • Claims of very high automation, negligible average human productivity, or AI producing most research, while acknowledging any substantive human research contribution.
  • Claims that humans are unnecessary but still contribute, or that human contributions cannot be measured or distinguished.
  • Fully automated individual projects, successor-model training runs, or subprograms without the program-wide assertion above.
  • Staff reductions, absence of employees with researcher titles, and third-party inferences from company operations.

Resolving source. Developer-authored official statements on its domain, research reports, system cards, corporate filings, or written testimony issued on the developer's behalf and published by the receiving court, regulator, or legislature. These are a narrower developer-only source requirement under Public confirmation (A2); outside evaluators and affected third parties cannot substitute for the developer's statement. Executive interviews, personal posts, and oral testimony without qualifying developer-authored written testimony do not count. Apply the Resolution window and Resolution date (A3) conventions.

Download all forecasts (JSON)