takeoff.watch

Most accepted papers at a top ML conference are AI-produced

Will the organisers of NeurIPS, ICML, or ICLR state that a majority of the papers accepted at one edition of the conference were produced primarily by AI systems?

Chance by Q4 2031

16%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 17 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 17 Oct 2026 · Methodology v2.1

Organisers of the top machine-learning conferences are currently fighting low-quality AI-written submissions and treating AI as a tool. None has come close to saying AI did most of the research behind its accepted papers. AI research ability is improving fast, so this could become true around 2028–2030, but organisers would still have to measure it and say so officially. The biggest change would be a conference asking authors whether AI led the research and reporting a majority.

Where things stand (17 October 2026)

No organiser of NeurIPS, ICML or ICLR has said anything close to "most accepted papers were produced primarily by AI." Organisers are mainly fighting low-quality AI-generated submissions and treating AI as a tool:

  • NeurIPS 2026 Position Paper Track (blog, 2 June 2026). Ran the Pangram AI-text detector, desk-rejected 18.4% of submissions and required human-written papers. That analysis covered writing and submissions, so it would not qualify here anyway. The NeurIPS 2026 main-track handbook says agents and LLMs cannot be authors. Main-track decisions went out on 24 September 2026, so that edition is technically in the window. Nobody claims its accepted papers were mostly AI-produced.
  • ICLR 2026 retrospective (31 March 2026). Reported desk rejections for hallucinated references and flagged LLM-written reviews. It gave no figure on how much of the research AI did.
  • ICLR 2027. About 60,000 abstracts were submitted, against 19,525 valid submissions in 2026. The chairs capped authors at 20 submissions, and at one for authors with no prior paper at a major venue, to curb "usually AI-generated" papers (Nature, October 2026). They wrote that current AI "seem[s] to have poor taste in research questions" and that humans "will never give up our role." ICLR 2027 does require a structured AI-use disclosure and says it will treat these as data. That is the most plausible way organisers could eventually produce counts. Its categories, though, ask whether AI was used for each task, not who did the main intellectual work.
  • COLM 2026 programme chairs estimated that about 5% of submissions used AI very heavily or were mostly AI-generated. ICML 2026 accepted 6,352 of 23,918 submissions. Its AI-related enforcement concerned reviews.

Capability trend (the main driver)

AI is getting better at research fast:

  • OpenAI says it reached its "automated research intern" goal in September 2026, uses 3.1 agent-workdays per human workday, and is aiming for an automated AI researcher by March 2028.
  • Anthropic says Claude "leads" 26% of its internal AI R&D work.
  • Anthropic co-founder Jack Clark puts about 60% odds on fully autonomous successor-training by the end of 2028.
  • Google's ScientistTwo (September 2026) improved on 86 of 107 accepted papers' benchmarks. Human reviewers rated its papers roughly equal to the NeurIPS originals.

There are counter-signals. A Princeton and UK AI Security Institute study (August 2026) found frontier agents still fail at open-ended research. Labs are also talking about voluntary slowdowns. Even so, it is plausible that by 2028–2030 most accepted ML papers will be AI-led in substance.

Why I still forecast low

A YES needs three things together:

  1. Most accepted papers at one edition really are mainly AI's intellectual work.
  2. The organisers measure it. Authors have reasons to understate AI's role while it is stigmatised.
  3. The organisers officially say so in "primarily AI" terms, on the conference's own site. Statements that AI was used for particular tasks, findings about submissions, or results from text detectors do not count.

Conferences exist partly to give humans credit, and organisers consistently describe AI as a tool. No journal or conference has ever made such a statement. For comparison, a much easier Metaculus question (whether any paper with an AI author appears at one of these venues before 2028) sits at roughly 68–88%. Our question asks far more.

Scenarios

  • Fast takeoff (~20–25%): automated researchers by 2028 and AI clearly dominant by 2029–30. I put about a 40% chance on an explicit statement by 2031.
  • Moderate progress (~45%): about 8%.
  • Slow progress (~30%): about 1%.

These blend to roughly 13–16%. The ScientistTwo evidence pushes me a little higher.

Hazard profile

I assumed the per-quarter chance is about zero through NeurIPS 2026 and ICLR 2027 (0.1–0.3%). It rises to 0.7–1.1% per quarter in 2028, as ICLR 2028, ICML 2028 and NeurIPS 2028 run on more capable agents. It then holds at about 0.9–1.2% per quarter through 2031, with a dip in the third quarter of each year, when there are fewer results announcements and retrospectives. Compounded, that gives about 4.7% by the end of 2028, 8.8% by the end of 2029, 12.9% by the end of 2030 and 16.4% by the end of 2031. I computed this in Python as one minus the product of (1 − hazard) across quarters.

Strongest case for higher: the capability curve is steep. ICLR already collects disclosures and promises to share the data. If a future form asks whether the research was "AI-led," the counts could cross 50% quickly after 2028.

Strongest case for lower: organisers can always describe AI-heavy work as "AI-assisted," and policies may require human-originated ideas.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, have the organisers of NeurIPS, ICML, or ICLR stated that a majority of the papers accepted at one edition of the conference were produced primarily by AI systems?

Definitions.

  • Conference and edition: one annual edition of the Conference on Neural Information Processing Systems (NeurIPS), International Conference on Machine Learning (ICML), or International Conference on Learning Representations (ICLR). Editions and conferences may not be pooled.
  • Organisers: the conference's governing board, organising committee, general chairs, or programme chairs speaking officially for the conference, rather than in a personal capacity.
  • Accepted papers: distinct papers formally accepted into that edition's conference programme, including all conference paper tracks and presentation categories. Workshops, tutorials, demonstrations without accepted papers, and affiliated events are excluded. A statement about only one track or a sample does not establish the conference-wide majority unless the organisers explicitly generalise the finding to the full accepted-paper population or provide counts establishing that majority. Count each paper once; subsequent withdrawal does not undo a formal acceptance.
  • Majority: strictly more than 50% of the accepted papers. An explicit statement of “a majority” or “most” qualifies. An exact count or percentage must exceed 50%; “at least half,” a rounded 50%, or a range spanning 50% does not qualify.
  • Produced primarily by AI systems: the organisers attribute the principal substantive intellectual work behind the papers—their research contributions, methods, arguments, or results—to AI systems rather than humans. Human supervision, verification, editing, submission, or listed authorship does not disqualify a paper. AI-generated prose alone, including drafting a manuscript describing human-produced research, does not qualify. AI systems from any developer count.

Resolves YES if an official statement or report identifies a particular edition and affirmatively concludes that the qualifying majority was produced primarily by AI systems. The conclusion may be based on author disclosures, a survey, an audit, or another method; independent verification of each paper is not required. An organiser-adopted statistical estimate above 50% qualifies even if its uncertainty interval crosses 50%, provided it is presented as the organisers' conclusion about the full accepted-paper population, not merely a possibility. Counts published by the organisers also qualify if they directly establish the required majority using the definition above.

Resolves NO / Exclusions. No qualifying statement by the horizon means NO. Predictions about future editions, hypothetical scenarios, statements about submissions rather than acceptances, AI-written peer reviews, and reports that most papers used AI in some capacity do not qualify. Neither author claims nor outside AI-detection studies qualify unless the conference organisers officially adopt the relevant conclusion. Merely hosting a paper, talk, or workshop making such a claim is not adoption. Permission for AI authorship or AI-assisted research is not sufficient.

Resolving source. Official organiser-authored statements, reports, programme-chair reports, or official transcripts published on neurips.cc, icml.cc, or iclr.cc, including their official blogs and subdomains; alternatively, an organiser-authored report explicitly linked as an official conference report from those domains. This is a narrower source requirement than Public confirmation (A2). Apply Resolution date (A3) and the Resolution window. Apply the Observation period convention: the edition's acceptance-decision period must end after the opening date. Official discussions distinguishing research production from writing include ICLR's 2026 LLM policies, ICML's 2026 Call for Papers, and NeurIPS's 2026 Position Paper Track statement; those policy statements alone do not meet the criterion.

Download all forecasts (JSON)