RenaSciQuestion network · The Renaissance of Science E5AI Enters the Lab

E5Interim assessmentFurther topic

AI Enters the Lab

How the Disciplines Responded (2023–2026)

?Now that AI is in every discipline, where is the real bottleneck?

Topic page
~1,430 words · 8 screens
Deep read
~10,635 words · 131 sources
Questions
19 out · 16 in
Start with the first question ↓Go straight to the deep read ↗
  1. 01Is AI a tool, or a colleague?
  2. 02When proofs arrive faster than anyone can read them, what is mathematics to do?
  3. 03When a machine announces a “new discovery”, should we take it seriously?
  4. 04Why did the flood hit the shipyard first?
  5. 05If AI reads faster than we do, what are humanities scholars for?
  6. 06Who gets in?
  7. 07If you used AI, whose name goes on the paper – and do you have to say so?
  8. 08Once the bottleneck has moved, what should we do?

◇The network re-centres on “AI Enters the Lab”. Every screen leaves a few follow-up questions; open one to walk into another topic.

01/08Medals

Is AI a tool, or a colleague?

Two medals and a summer

In October 2024, within two days of each other, the Nobel prizes in physics and chemistry both honoured work bound up with artificial neural networks; half the chemistry prize went to Demis Hassabis and John Jumper, the makers of AlphaFold. Less than two years later the tremor reached mathematics, the discipline that cares most about proof. In July 2026 a mathematician working with Claude found a counterexample to the Jacobian conjecture in three or more dimensions – a formula short enough to fit in a single post. In September OpenAI claimed to have cracked the Navier–Stokes problem with some ten thousand agents.

AI plays at least five roles in research: analytical tool, predictor, assistant, agent and reviewer. The further down that list, the harder the work is to verify and the harder it is to say who is responsible. Depending on AI is not the problem; an environment without AI no longer exists. Homogenization is the problem: if everyone asks the questions AI answers most readily, we will fool ourselves into thinking we have explored every possibility.

Follow-up questions
Read deeperAI has entered the lab. How far has it got?↘
Asked from elsewhereMRenaissance

02/08Maths

When proofs arrive faster than anyone can read them, what is mathematics to do?

Proof indigestion

In April 2026 a 23-year-old with no professional training in mathematics put a single question to an AI and got back a proof of an Erdős problem that had stood open for about sixty years. The raw output was “quite poor” and needed an expert to sift it before it made sense, yet it used a formula that everyone knew and no one had thought to apply here. By August, 565 problems on the Erdős problems website were marked as solved. Mathematics was the first field to be ready for verification: Lean and Mathlib let machines check a proof line by line.

Terence Tao breaks solving a problem into five stages – generation, verification, exposition, digestion and entry into the textbooks – and warns that AI will cause “proof indigestion”. His rule: a proof that no one can properly explain is incomplete, even if it has been formally verified. More than four thousand people have signed the Leiden Declaration, which calls for tools to be disclosed and keeps credit and responsibility with humans. More than seven hundred signed a letter against a forty-hour “Mathathon”, which was eventually recast as a search for new proofs of existing theorems, written up as clear expositions.

Follow-up questions
Read deeperWhen proofs come faster than they can be read, what is mathematics to do?↘
Asked from elsewhereMFloodE3ReturnE4AgonE4Response

03/08Lab

When a machine announces a “new discovery”, should we take it seriously?

Fast predictions, “new” in scare quotes

The AlphaFold database holds more than 200 million protein structures. Google’s “AI co-scientist” took about two days to propose a hypothesis that matched years of unpublished work by a team at Imperial College London. The European Centre for Medium-Range Weather Forecasts has put AI forecasts into operational service. But “new” is often claimed too soon. The automated laboratory A-Lab said it had synthesized 41 new compounds; in early 2026 its authors corrected the paper, conceding that “new” meant new only to their prediction platform. And independent scientists judged about a fifth of the statements in the reports of Kosmos, an “AI scientist”, to be inaccurate.

A prediction can be checked against experiment; a discovery takes time to confirm. AI has long been creative. What it lacks is not ability but motive: with no individual desire and no finitude, it is good at gathering up everything that already exists, and bad at striking out on a path of its own.

Follow-up questions
Read deeperThe life sciences: what comes after predicting structures?↘
Asked from elsewhereE3Theatre

A chemical laboratory with a table of chemical affinities. Engraving by Prévost after Goussier, 1763. Wellcome Collection · Public domain

04/08Flood

Why did the flood hit the shipyard first?

The faster we write, the slower we review

Computer science is both where AI is made and the first field to be swamped by what AI makes. Submissions to NeurIPS more than doubled in five years. One detection company estimates that about 21% of the peer reviews for ICLR 2026 were written entirely by AI. arXiv now requires review and position papers in computer science to pass peer review first. The position-paper track of NeurIPS 2026 desk-rejected 178 papers on the word of a detector, and that I cannot accept: what should be judged is the paper itself, not who wrote it.

The social sciences have it harder still. An AI “synthetic respondent” passed 99.8% of attention checks, and could guess the researcher’s hypothesis and generate data that “confirmed” it; with a few rewordings of a prompt, almost any conclusion can be made “significant”. The economist Scott Cunningham put it well: the constraint on science is shifting from production to evaluation.

Follow-up questions
Read deeperComputer science: why did the flood hit the shipyard first?↘
Asked from elsewhereE2Sluice gate

05/08Humanities

If AI reads faster than we do, what are humanities scholars for?

Reading charred scrolls – and ourselves

In June 2026 the Vesuvius Challenge announced that, for the first time, an entire Herculaneum scroll had been virtually unrolled and read from beginning to end. AI found the ink in the X-ray data; deciding what the text said was still the job of papyrologists. The editing of classical Chinese texts has taken the road of scale plus crowdsourcing, with thousands of amateurs helping to proofread. And AI’s error rate in reading historical manuscripts is now low enough to bear comparison with the best experts.

D. Graham Burnett, a historian of science at Princeton, writes that “factory-style scholarly productivity was never the essence of the humanities”. I would put it more bluntly. The model of sitting on a hoard of exclusive manuscripts and eking out papers from them, one squeeze at a time, for a whole career, will soon be held in contempt; and the people slowest to respond to new technology are precisely humanities scholars. Dependence has always been the fate of technology. AI may be better at reading books than we are, but the responsibility for choosing which books to read still lies with us.

Follow-up questions
Read deeperWhat should humanities scholars worry about most?↘
Asked from elsewhereE1PreludeE6Curator

The eruption of Vesuvius on the night of 8 August 1779. Coloured etching by Fabris, 1779. Wellcome Collection · Public domain

06/08Threshold

Who gets in?

Amateurs and ten thousand agents

In the comments section of the Erdős problems website, a researcher who works at a customer-service company by day can collaborate with Terence Tao; whether you have tenure, or a university post at all, makes no difference. At the other end, very few mathematicians can afford an assault that deploys ten thousand agents and costs millions of dollars. And a study of 41.3 million papers found that scientists who use AI publish three times as many papers as others, while the range of topics that science as a whole explores has narrowed.

But let us not picture traditional academia as an equal, low-barrier world. Laboratories, academic clans, student labour: AI weakens more of these old monopolies than the new concentrations it creates. What really deserves our worry is the homogenization that comes from a narrowing focus.

Follow-up questions
Read deeperAmplification, compression, redistribution: what has AI actually changed?↘
Asked from elsewhereMTurning pointE3AmateursE7Exit

07/08Rules

If you used AI, whose name goes on the paper – and do you have to say so?

Authorship, disclosure, confidentiality

In January 2023 Nature announced that large language models could not be credited as authors, and an editorial in Science was bluntly titled “ChatGPT is fun, but not an author”. On this point I agree completely: authorship is first of all not an honour but a responsibility, and AI cannot take the blame. The US National Institutes of Health bars peer reviewers from using generative AI to write their critiques; China’s Ministry of Science and Technology rules that it must not be used to generate application materials directly, nor be listed as a co-contributor to research results.

Three years on, three points of consensus have emerged: AI cannot be an author, its use must be disclosed, and review must stay confidential. The third depends on the circumstances. The second needs its nature spelled out: disclosure is a contract with a platform, not an author’s confession, and it works best as the natural by-product of a record of the research process.

Follow-up questions
Read deeperIf you used AI, whose name goes on the work – and do you have to say so?↘
Asked from elsewhereMStandardsE2Sluice gate

A page of the Philosophical Transactions. Royal Society, from 1665. Wellcome Collection · Public domain

08/08Shift

Once the bottleneck has moved, what should we do?

Amplification, compression, redistribution

AI has amplified output, and amplified junk along with it. It has compressed the time it takes to generate, but not the time it takes to understand. It has lowered the barriers, and bred new concentrations. What is scarce is no longer getting things written, but verifying, explaining, digesting and evaluating them. Science needs new technologies of witnessing and new forms of scholarly work, and it needs to redesign checking and explaining as work that is worth doing and that earns recognition.

In the ethical treatise read from a charred Herculaneum scroll in 2026 there is a broken sentence: “…we will inquire into something, but we will not grasp it, if in some way we depart from ourselves and from our own nature…” Machines can inquire on our behalf; grasping is something only we can do.

Follow-up questions
Read deeperOnce the bottleneck has moved, what should we do?↘
Asked from elsewhereMTurning point

E5The discussion goes on

The discussion goes on

?Now that AI is in every discipline, where is the real bottleneck?

In October 2024, within two days of each other, the Nobel prizes in physics and chemistry honoured work bound up with artificial neural networks. Less than two years later, AI has stirred up a storm in the place mathematicians hold most dear: proof. This essay is an interim assessment as of 30 September 2026. It distinguishes the roles AI plays in research, looks in turn at how seven fields have reacted, and discusses how institutions and researchers have responded. Let me give the conclusion first. AI has amplified output, and amplified junk along with it. It has compressed the time from idea to result, but has not compressed the time needed for verification and understanding to the same degree. It has redistributed who gets to take part in research, loosening some old monopolies while creating new concentrations at the other end. The bottleneck is moving away from “getting it written” and towards verification, explanation, digestion and evaluation.

Deep readAI Enters the Lab: How the Disciplines Responded (2023–2026)~10,635 words · 13 questions · 131 sources↗
  1. 1AI has entered the lab. How far has it got?
  2. 2What roles does AI play in research?
  3. 3When proofs come faster than they can be read, what is mathematics to do?
  4. 4The life sciences: what comes after predicting structures?
  5. 5Physics, chemistry and the Earth sciences: can we take “new” at face value?
  6. 6Computer science: why did the flood hit the shipyard first?
  7. 7The social sciences: are the respondents still human?
  8. 8What should humanities scholars worry about most?
  9. 9Medicine: why is the bar for evidence set highest?
  10. 10If you used AI, whose name goes on the work – and do you have to say so?
  11. 11What do researchers think?
  12. 12Amplification, compression, redistribution: what has AI actually changed?
  13. 13Once the bottleneck has moved, what should we do?

→Questions from here

The follow-up questions left on this page, grouped by the topic they lead to.

←Questions from elsewhere

Questions asked in other topics that lead here. Open one to return to the screen where it was asked.