takeoff.watch

A major report finds AI has not sped up scientific discovery

Will the US National Academies, the Royal Society, or the OECD publish a report concluding that AI has not measurably raised the rate of scientific discovery or of new drug approvals?

Chance by Q4 2031

6%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 15 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 15 Oct 2026 · Methodology v2.1

I put the chance at about 6% over five years, and near zero this year. These three bodies have so far described AI's potential for science without judging its overall effect, and they usually say "too early to tell", which does not count here. Meanwhile, AI's research successes make a flat "no effect" finding less likely. The biggest change would be an OECD flagship report, such as its 2027 science outlook, stating that AI has not measurably increased discoveries or drug approvals.

Where things stand (15 Oct 2026)

I found no qualifying report, and nothing that is scheduled to become one.

  • OECD. Its main AI-and-science work is promotional in tone. Artificial Intelligence in Science (June 2023) calls faster research "the most economically and socially valuable" use of AI and says its "full potential is far from realised". The STI Outlook 2025 (Oct 2025) mentions "worrisome claims that research productivity has been falling" and says AI "has the potential to amplify scientific productivity". Neither makes a finding about what AI has actually done to the aggregate rate. Both also predate the opening date. The STI Outlook 2027 is being prepared (CSTP session, April 2026); the themes reported so far are the securitisation of science and the SDGs. The OECD Blue Sky Forum (14–16 Sept 2026) had a plenary on AI-enabled research and a poster asking "Does AI advance science?". That signals interest in measurement, but such work usually comes out first as working papers, which do not count.
  • Royal Society. Science in the age of AI (2024) covers opportunities, integrity and access, with no claim about aggregate rates. The final Science 2040 report has not been published.
  • National Academies. The AI work I found covers foundation models for DOE science, the future of work, cybersecurity, and a planned consensus study on AI-enabled biology and biosecurity. I found no study of AI's aggregate effect on the rate of discovery or on drug approvals. The administration has also proposed defunding the Academies (AIP, July 2026), which threatens future commissioned studies.
  • Underlying evidence. It mostly points the other way. Google's AI in Science: Early Insights (Sept 2026) says there is "so far little data" but reports large time savings for scientists. Ding, Lawson and Shapira (arXiv, June 2026) find AI-related papers are more likely to rank among the most creative.
  • Drug approvals. FDA novel approvals were 55 in 2023, 50 in 2024 and 46 in 2025, with 27 by early October 2026. As of mid-2026 no AI-discovered drug had been approved. So the facts could support a "no measurable effect yet" claim on approvals, but one of the named bodies would have to state it explicitly.

Reference class and base rate

My reference class is reports by these three bodies on AI in science or drug development, 2023–2026. Several were published and none made an explicit finding that AI had not measurably raised an aggregate rate. They default to "potential", "too early to tell" or "evidence is limited", and the question excludes all three. Explicit null findings about a technology's aggregate effect on discovery are rare from these institutions.

Paths to YES and NO

YES most plausibly comes from:

  1. An OECD flagship (STI Outlook 2027/2029/2031) or a dedicated AI-in-science stocktake saying something like "AI has not yet produced a discernible acceleration in aggregate research productivity or discovery".
  2. An OECD, National Academies or Royal Society report on drug development saying AI has not yet increased the number of new drug approvals. Long development timelines make that true for some years.

NO comes from:

  • hedged wording that the question excludes;
  • AI's visible successes making a null finding implausible (autonomous "AI scientist" results, likely first approvals of AI-designed drugs around 2027–28);
  • the requirements for aggregate scope, adoption by the institution, and supporting data covering periods after the opening date.

My rough decomposition:

  • P(a named body formally assesses AI's aggregate effect in 2027–31) ≈ 0.5
  • P(the aggregate effect is in fact unclear) ≈ 0.4
  • P(explicit null wording rather than hedging) ≈ 0.25
  • plus a separate drug-approval path of about 2–3%
  • minus a deduction for strict resolution.

That gives about 6% over five years.

Hazard model (computed)

Quarterly hazards are about 0.15% for Q4 2026 (almost no qualifying post-opening data yet), 0.2–0.25% in early 2027, and about 0.3% per quarter after that. I added bumps to about 0.45% in Q4 2027, Q4 2029 and Q4 2031 for likely STI Outlook releases. The cumulative probability is 1 − ∏(1 − h): 1.3% by end-2027, 2.5% by end-2028, 3.8% by end-2029, 5.0% by end-2030 and 6.3% by end-2031. I kept the hazard roughly flat because two effects offset each other: more evidence and more reports over time, but also a growing real effect of AI that makes null findings less plausible.

Strongest case against my estimate

Three prolific bodies have five years, and drug approval counts are stable and slow to respond. A single sober sentence in one OECD health or STI report would be enough, and the OECD's own analysts have voiced scepticism (for example, the warning that LLMs could worsen "scientific overload"). That keeps me from going below about 5%. But the exclusions explicitly cover the most likely hedged phrasings, so I don't go much higher.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has the US National Academies, the Royal Society, or the OECD published a report concluding that AI has not measurably raised the rate of scientific discovery or of new drug approvals?

Definitions.

  • Named bodies: the US National Academies of Sciences, Engineering, and Medicine (including its constituent National Academy of Sciences, National Academy of Engineering, and National Academy of Medicine); the United Kingdom’s Royal Society; or the Organisation for Economic Co-operation and Development (OECD).
  • Report: a final, publicly available institutional research, assessment, or policy report issued or jointly issued by a named body. The conclusion must be presented as a finding of the report, including its executive summary or synthesis, rather than merely quoted from a contributor or outside source. National Academies Consensus Study Reports qualify. Workshop proceedings, individual journal articles, guest essays, working papers carrying only their authors’ views, speeches, press releases, and draft reports do not qualify unless the named body separately adopts the conclusion in a qualifying report. Routine disclaimers that an OECD report does not represent every member government do not disqualify it.
  • AI: artificial-intelligence methods used in scientific research or drug discovery and development, including machine learning and generative AI; not restricted to frontier models or covered developers.
  • Rate of scientific discovery: the pace of production of new scientific knowledge, assessed across science as a whole globally or within at least one country. A report may use an explicit proxy for discoveries, but must itself connect that proxy to the rate of discovery. Findings limited to one discipline, institution, task, or AI product do not qualify.
  • Rate of new drug approvals: the number per unit time of newly authorized drugs across a national, multinational, or global regulatory system. Findings about one therapeutic category, one company, trial entry, candidate generation, or supplemental approvals of existing drugs alone do not qualify.
  • Has not measurably raised: an explicit retrospective finding that AI has produced no detectable positive effect on at least one of these aggregate rates during the period assessed. An explicit finding of no statistically significant increase attributable to AI qualifies; the report’s stated measurement standard governs. A finding of reduced rates also qualifies. No finding that AI can never raise the rate is required.

Resolves YES if at least one named body publishes a qualifying report that explicitly reaches the defined conclusion about either scientific discovery or new drug approvals. A report may acknowledge successful AI applications, faster individual research tasks, or future benefits and still qualify if it reaches the required aggregate conclusion. The finding must concern a period ending after the question’s opening date; the Observation period convention applies to any published statistics supporting it.

Resolves NO / Exclusions. Resolve NO for a horizon if no qualifying report has been published by that horizon. The following do not suffice:

  • Statements that evidence is insufficient, measurement is impossible, it is too early to tell, or further evaluation is needed, without an explicit finding of no detectable increase.
  • Findings that benefits are smaller than expected, that growth remains slow, or that overall rates are flat or falling, without a conclusion about AI’s contribution.
  • Predictions of future stagnation or hypothetical scenarios.
  • Findings only about costs, researcher productivity per dollar, publication volume, or time saved, unless the report explicitly concludes that the rate of discovery or new drug approvals has not measurably increased because of AI.
  • Reports published before the opening date, or new publications merely repeating findings for periods ending before that date.

Resolving source. The qualifying report itself, published through nationalacademies.org or the National Academies Press (nap.edu), royalsociety.org, or oecd.org or OECD iLibrary (oecd-ilibrary.org), including official successor publication platforms. These named-body reports are the exclusive resolving sources, narrower than Public confirmation (A2); third-party descriptions alone cannot resolve the question. Apply Resolution date (A3), the Resolution window convention, and the Observation period convention.

Download all forecasts (JSON)