AI in Science

Where it stands

• Use is near universal. Most researchers use general-purpose assistants for reading, coding, writing and ideas, but few feel trained to use them responsibly.
• More output, narrower science. The best evidence shows individual gains alongside a contraction of topics and less engagement between scientists.
• The bottleneck has moved to checking. Review, moderation and validation are strained, from arXiv to grant panels.
• Governments are betting on AI for science. The US, China and the UK build it into strategy; few test what AI does to assessment itself.
• Access and ownership are open questions. Models, data and instruments are concentrated, and access within universities is uneven.

1. What AI changes

AI predicts what used to need experiments, generates candidates, simulates and reads the literature at scale. Progress is fastest where data is good.

• From open problem to infrastructure: AlphaFold’s creators shared the 2024 Nobel Prize in Chemistry (Nobel Prize).
• Uneven by data: wins cluster where data is clean, structured and verifiable, and stall in messier or regulated areas (Benchling 2026); progress is “broad but shallow”, with data infrastructure the binding constraint (Charting AI, July 2026; Malliaraki).
• Tools for observing: James Evans argues AI’s most important role may be helping design new instruments, since about a quarter of fields were opened by a new tool (Nature, August 2026).
• Expected pace: Kevin Bryan argues AI will speed up science more slowly than many expect, because experiments, institutions and adoption set the pace (Macroscience).

2. The researcher’s work

Use is close to universal and mostly through general-purpose assistants; the craft and the burden of checking are changing.

• Adoption: 84% of 2,430 researchers used AI tools in 2025, up from 57% a year earlier (Wiley ExplanAItions); only 27% felt they knew how to use them responsibly (OUP).
• Daily use: nearly half of 637 scientists use AI daily and save about seven hours a week, and almost half say it steers them towards safer, incremental questions (Google DeepMind, October 2026).
• Access: without institutional access to frontier models, good practice becomes a privilege of the well-funded (Hendriksen, Altinget).
• Writing and origin: journals have started judging work by suspected AI origin, which detectors cannot settle; Allison Stanger argues for judging analytical writing by its quality (Chronicle, October 2026).

3. Productivity and narrowing

AI makes individual scientists more productive while the field narrows: the central tension in the evidence so far.

• The causal study: across 41.3 million papers, AI-augmented scientists publish 3.02 times more and are cited 4.84 times more, while the range of topics shrinks 4.63% and engagement between scientists falls 22% (Hao, Xu, Li & Evans, Nature 2026).
• The mechanism: AI pulls research towards data-rich, benchmark-friendly questions; many fields lack the data and incentives to produce it as a public good (Malliaraki).
• Ownership: who owns AI-generated results and “knowledge capital” will shape the future knowledge ecosystem (Fosgerau).

4. Validation, review and publishing

Generation is cheap; checking is slow. The bottleneck is moving to validation, review and moderation.

• Conjecture machines: agents make hypotheses abundant while validation stays physical, institutional and slow (Google DeepMind, July 2026).
• Review capacity: arXiv now limits submitters to two papers a month after submissions doubled (Avasthi); grant applications rose 57% at twelve funders from 2022 to 2025, and in the Marie Skłodowska-Curie scheme a fifth of proposals scored above 95% (Wilsdon).
• AI reviewers: the best single AI reviewer caught 71 of 100 planted errors and an ensemble 93, but omissions escaped all systems (Litvak); manuscript-only review barely catches flaws in AI-scientist papers, so full workflow logs and code are needed (Luo, Kasirzadeh & Shah, PNAS).
• Shared infrastructure: working scientists name coordination failure and missing shared infrastructure for data, methods and null results as the main bottleneck (Astera).

5. Automated research and new institutions

Systems that run the whole loop, self-driving labs and AI-first organisations are early but moving fast.

• End-to-end systems: AI scientists draft hypotheses, run experiments and write papers; FutureHouse builds agents for biology (FutureHouse); self-driving labs close the experimental loop in chemistry and materials.
• Mathematics: Terence Tao calls for a “Maths 2.0” as AI reshapes proof and discovery (Nature, 22 October 2026).
• Measuring the shift: Wilsdon suggests universities report the metric AI labs already use, agent-workdays per human workday (Wilsdon).

6. Funders and governments

Governments are building AI for science into strategy, and funders are starting to study AI in their own processes.

• United States: the Genesis Mission and $2.4bn in compute credits from eleven companies (8 October 2026) (White House); the science adviser calls “superintelligence for science” the biggest unlock in history (NYT interview).
• China: the AI+ plan puts AI for science first among six areas and targets over 90% adoption of intelligent agents by 2030 (Caixin; see the China tracker).
• United Kingdom: the Metascience Unit funds about £5m for seven projects on AI-driven indicators of research quality, as evidence before any use (UKRI).
• A norm for assessment: no funder, journal or university should use AI in assessment without disclosing it, measuring its effects and testing whether it spots what humans miss (Wilsdon).

7. Training, universities and uptake

AI reaches the whole chain from training to application.

• Training: AI removes the entry-level work through which judgement used to be learned, so the apprenticeship model of research training is under pressure.
• Universities: the bundle of research, teaching and credentials is under strain (Gilman); the teacher’s value shifts towards judgement, mentoring and verification.
• Uptake: results travel faster, so the bottleneck moves to the absorptive capacity of firms, clinics and public bodies.

8. Risks

Three risks recur across the evidence.

• Integrity: AI-generated papers and fabricated results outpacing validation.
• Concentration: frontier models, data and instruments held by a few organisations, deciding who can do AI-driven science.
• Epistemic: narrowing of topics and the automation of error at scale.

Timeline

1. Oct 2024
   AlphaFold’s creators share the Nobel Prize in Chemistry.
2. Jan 2025
   DeepSeek-R1 released as open source.
3. 2025
   Wiley: 84% of researchers use AI tools, up from 57%.
4. Aug 2025
   China’s State Council AI+ plan, with AI for science first.
5. Jan 2026
   Hao, Xu, Li & Evans (Nature): more output per scientist, narrower science.
6. May 2026
   Malliaraki on AI’s data blindspot.
7. Jul 2026
   Charting AI (Renaissance Philanthropy) and Conjecture Machines (Google DeepMind).
8. Aug 2026
   Litvak’s planted-error benchmark of AI peer review; Evans on AI as a designer of instruments.
9. 2 Oct 2026
   arXiv caps submitters at two papers a month.
10. 5 Oct 2026
    PNAS: AI-scientist systems hide flaws that manuscript review misses.
11. 8 Oct 2026
    UK Metascience Unit funds AI-driven research evaluation; Google DeepMind’s survey of 637 scientists; US Genesis Mission compute credits.
12. 13 Oct 2026
    Stanger on judging AI-assisted writing by quality.
13. 17 Oct 2026
    Fosgerau on knowledge capital and AI.
14. 20 Oct 2026
    Wilsdon: application volumes up 57% at twelve funders, and a call for a norm on AI in assessment.
15. 22 Oct 2026
    Kratsios on “superintelligence for science”; Terence Tao on a “Maths 2.0”.

What to watch

• Do funders and journals disclose and test how they use AI in review and assessment?
• Does the narrowing of topics continue, and do funders act to keep exploration diverse?
• How do review and moderation systems cope with rising volumes: caps, fees, AI filters or new models?
• Do self-driving labs and AI scientists produce validated discoveries beyond benchmarks?
• Who gets access to frontier models, compute and data in universities, and on what terms?
• How does research training change when AI does the entry-level work?

My reading