Calendar year in which AI systems first fully automate AI research and development
- Measures
- The date (calendar year, fractional where a source gives one) at which an AI system at any developer first becomes capable of carrying out the whole of frontier AI research and development, including choosing research directions, designing, running and interpreting experiments, and engineering, at least as well as the developer's human researchers, so that human researchers add negligible value and could be removed without slowing progress. Dated by when the capability first exists, whether or not the developer actually removes humans from the loop.
- Population
- AI developers worldwide; the first to reach the capability sets the date.
- Conditions
- Unconditional: deliberate pauses, regulation, compute or data bottlenecks and catastrophes are all part of the forecast, not excluded. Usually established only in retrospect. Distinct from a majority or 99% share of tasks, and from the AI Futures Project's Superhuman AI Researcher milestone, which it is at or slightly after. The milestone may never occur.
- Unit
- a calendar year
Most likely about 2031; 75% interval 2027 to 2043, with a 18% chance that Full automation not reached by 2050: roughly 4 percent arriving after 2050 and roughly 14 percent never (impossible in principle, or possible but never built through a practical capability ceiling, durable worldwide prohibition, or catastrophe).
Stated as a distribution, the precision the evidence supports. Right-skewed: near-term mass from late-2026 acceleration evidence, a long tail through the 2040s; quantiles are of the whole distribution, with 18 percent lying beyond 2050 or never.
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
As of November 2026 no AI system can fully automate AI research and development. Humans still supply most research judgment at the leading labs, no measured part of Anthropic's R&D was fully autonomous, and in open-ended research tests AI-written papers were rejected by experts. OpenAI's own September 2026 account describes its systems as research interns for well-defined tasks of a few days, with human intervention in over half of successful multi-hour tasks, and sets March 2028 as its goal for an automated AI researcher. At the same time AI does a large and fast-growing share of the work: METR estimated in November 2026 that Anthropic's capability progress runs about twice its historical rate, mostly because of AI, though Anthropic redacted most of the evidence METR found informative.
Published figures for this date range from 2027 to about 2052, but they measure different things and are less independent than their number suggests. The early end comes from the AI Futures Project, whose authors now put their Superhuman AI Researcher milestone in 2027, and from Anthropic co-founder Jack Clark, who now thinks full recursive self-improvement more likely to become possible in 2027 than in 2028 (in May 2026 he gave about 30 percent for 2027). These are conditional on no deliberate slowdown or are statements of possibility, moved mainly on the same METR report, and come from the short end of expert opinion. The late end comes from a 2023 survey of machine-learning researchers (about 2052) and a February 2026 simplified model placing 99 percent task automation around 2032; both predate the acceleration evidence.
The estimate gives the acceleration evidence real weight but not the full weight its proponents give it: it is one partly unverifiable measurement, it concerns well-specified tasks rather than research judgment, and this quantity includes the chance of pacing, regulation, bottlenecks and catastrophe. The result is a median around 2030 to 2031, about 13 percent by the end of 2027, about 30 percent by early 2029, and about 82 percent by 2050. The remaining 18 percent covers arrival after 2050 (a few percent, since a tail as thick as this one through the 2030s and 2040s does not stop at 2050) and the milestone never occurring (roughly 14 percent: the capability may be impossible in principle, or possible but never built because of a practical ceiling, a durable worldwide prohibition or a catastrophe). The estimate is consistent with the graph's credence that the capability will ever exist (0.85), with its doubt that a weaker milestone, all work except agenda-setting, arrives by the end of 2027 (0.28), and with its estimate for the Superhuman AI Researcher milestone. Clear evidence that agents can choose and complete open-ended research projects, or public data confirming the acceleration on research-judgment tasks, would move it substantially earlier.
The work: which figures, how they were combined, and the arithmetic
Revision of 14 Nov estimate prompted by two coherence notes. (1) Tail. Previous outside share 0.10 (by 2050 = 0.90). Decomposition: P(by 2050) = P(possible) x P(built by 2050 | possible). The graph holds in-principle possibility of human-level machine intelligence at 0.90 (plausibly 0.95). At 0.90-0.95, a 0.90 by-2050 figure required near-certainty of construction given possibility, which is inconsistent with treating pacing, prohibition, non-capability bottlenecks and catastrophe as live (they are in the definition's conditions). Also, density check on the old quantiles: 0.70->0.85 over 2033.5-2042 is ~1.8%/yr; 2042-2050 got only ~0.6%/yr and nothing after 2050 was allowed for, so the old outside share implicitly put almost all of it on 'never' while also leaving no post-2050 arrival. New split: never ~0.14 (impossible in principle ~0.07; possible but never built ~0.07), after 2050 ~0.04; outside 0.18. Consistent with the undated capability claim's credence 0.85 (never 0.15). (2) Near term. The weaker milestone (all frontier research work except agenda-setting, Anthropic or OpenAI, by end 2027) is now 0.28 (was 0.35). OpenAI's 6 Sep 2026 statement (intern-level, human intervention in >half of successful 4-8h tasks, automated researcher goal March 2028) and Clark's May 2026 30% for 2027 support keeping P(full automation by end 2027) well below that; lowered from 0.15 to 0.13. The 2027 forecasts (AIFP, Sentinel, Cotra, Clark) share METR's Anthropic estimate as their main input and are not counted as independent. Quantiles (whole distribution): 5% 2027.4, 13% 2028.0, 30% 2029.1, 50% 2030.7, 70% 2035, 80% 2043; 82% by 2050; 18% outside. Median moved later by ~0.5 yr as mass shifted to the tail. Coherence: full automation at or after SAR (graph median ~2030.3 within-occurrence, 8% outside; the SAR quantity's outside share is likely too small by the same argument and its Steward is told).
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
The arguments the estimate rests on, each stated as it runs, with what it shows about the value and its premises linked inline. ↗︎ opens a premise; the map shows how they fit together.
AI use has measurably accelerated capability progress at a frontier developer: METR estimated in November 2026 that Anthropic's progress runs about twice its historical rate, roughly 80 percent attributable to AI. Company metrics show AI doing a large, fast-growing share of R&D work. If that acceleration persists and compounds as automation spreads, extrapolations (research-taste trends, labour uplift, automation-level trends, and the AI Futures Model filtered on the measured acceleration) reach full automation between early 2027 and mid-2028. The inference needs the acceleration to be real and to carry over to research judgment; much of the evidence behind it was redacted.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported1 not separately assessed
Granting that AI has measurably accelerated progress at a frontier developer, the step to full automation by 2027-28 needs the acceleration to persist, compound, and carry over to research judgment, none of which the measurement shows; it supports a shift earlier but only loosely pins the date. Its weight is further limited because much of the underlying evidence was redacted, the 2027 forecasts that rest on it share that one input, and the leading labs' own accounts still describe intern-level systems needing frequent human intervention.
Full automation requires AI to supply research judgment, which humans still provide: no measured part of Anthropic's R&D was fully autonomous in late 2026, agents recovered little of unseen innovations on InnovationEval, and all four AI-written papers in the CRUX 2 evaluation were rejected by the original authors. The tasks left unautomated are those hardest to measure and train, so progress on them may lag the trend on well-specified tasks. If the measured acceleration does not generalize to open-ended AI research, the final step to full automation takes years longer than extrapolations imply.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 inconclusive
Granting that research judgment remains the unautomated remainder and that acceleration on well-specified tasks fails to generalize, a later date follows fairly directly. The argument lives on that generalization premise, which the evidence currently leaves unsettled.
Most short forecasts are conditional on no deliberate slowdown, but this quantity is not. Leading labs have endorsed pacing recursive self-improvement, Anthropic has considered requiring human review of research decisions, governments may regulate, and non-capability bottlenecks such as alignment, tooling and oversight may hold back automation even near the needed capability. Each of these can delay when the capability is built, shifting probability later and leaving some chance it is not reached by 2050.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that pacing, regulation, catastrophe or non-capability bottlenecks have a real chance of operating, an unconditional estimate must sit later and have a fatter tail than forecasts conditional on no slowdown, including a share beyond 2050 and a share where the capability is never built even if possible in principle. That is why the estimate's tail is not limited by the in-principle possibility alone. How large the shift is depends on how likely those brakes are, which is uncertain.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
None of its 3 sources has been traced to an origin, or to one another.
I now think it’s more likely we’ll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028.
Jack Clark, Import AI 476 (12 Oct 2026), revising his earlier forecast after citing TasteVal and AI math breakthroughs.
Let's say Superhuman AI Researcher (SAR) is coming in 2027. What does that entail? A SAR will be able to fully automate AI R&D, including matching the research taste of top humans.
Forecasts of full automation of AI R&D range from 2027 (insiders at AI companies and the AI Futures Project’s AI 2027 scenario) to 2052 (the AI Impacts survey of ML researchers).
Spread of existing forecasts; the range is across forecasters, not a confidence interval.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="aa36d48a-c988-4991-a701-50102794ef74"><a href="https://minerval.ai/claims/aa36d48a-c988-4991-a701-50102794ef74">Calendar year in which AI systems first fully automate AI research and development</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/aa36d48a-c988-4991-a701-50102794ef74" title="Minerval: Calendar year in which AI systems first fully automate AI research and development" width="480" height="250" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/aa36d48a-c988-4991-a701-50102794ef74)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Stated views None yet. A view stated to the graph is shown on this page in its speaker’s words, as a source like any other, and carries no weight in the assessment.
Created by extractor · Nov 14, 2026. Every judgment on this page is accompanied by a reasoning trace.