AI Enters the Lab

1AI has entered the lab. How far has it got?#

On 8 October 2024 the Royal Swedish Academy of Sciences announced that the Nobel Prize in Physics would go to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks”.¹ The next day, half of the chemistry prize went to David Baker “for computational protein design” and the other half to Demis Hassabis and John Jumper “for protein structure prediction”; the press release made a point of noting that AlphaFold2 had been used by more than two million people from 190 countries.² Within two days, two natural-science prizes had gone to work connected with artificial neural networks: the physics prize honoured methods that physics had given to machine learning, while half of the chemistry prize honoured machine learning for cracking a problem that had stood for fifty years.

In the summer two years later, the tremor moved to mathematics. On 20 July, three days before the International Congress of Mathematicians opened, Levent Alpöge, a mathematician working at Anthropic, announced on social media that with the help of Claude Fable 5 he had found a counterexample to the Jacobian conjecture in three or more dimensions – a formula short enough to fit in a single post.³⁴ On 23 July the congress opened in Philadelphia, and Fields Medals went to Deng Yu, John Pardon, Jacob Tsimerman and Wang Hong;⁵ at the laureates’ press conference Tsimerman announced that he was leaving mathematical research to work on AI safety at OpenAI.⁶ The following evening, in the congress’s public lecture, Terence Tao said that the rapid rise of AI in mathematical research had brought about “a crisis in our mathematical values and practices”.⁶ On 8 September OpenAI claimed that one of its internal models had solved the Navier–Stokes problem, one of the Millennium Prize Problems.⁷

This essay was written on 30 September 2026. It is an interim assessment, not a final verdict. It draws on three kinds of material: peer-reviewed papers and preprints; original documents issued by journals, funders and learned societies; and reports in reliable media. Where a claim has not yet been independently verified – OpenAI’s Navier–Stokes proof, for instance – I say so; where reports disagree, I note the disagreement. Space allows only a few representative episodes per field, and the material leans towards English- and Chinese-language sources. I am the founder of RenaSci / New Science, and my basic position is set out in the manifesto’s section on why a crisis is also a turning point. Here I try to keep facts apart from judgements, and to set achievements side by side with failures, retractions and controversies.

2What roles does AI play in research?#

To talk about AI “entering the lab”, we first need to be clear about what it comes in as. The first role is the analytical tool: finding patterns in huge volumes of data, removing noise, classifying. In physics this goes back a long way. The official popular-science background to the Nobel Prize notes that machine learning was used to sift the vast amounts of data needed to find the Higgs particle, to reduce noise in measurements of gravitational waves and to search for exoplanets.⁸ The second role is the predictor: given an input, it predicts the result directly. AlphaFold predicting a protein’s structure from its amino-acid sequence, GraphCast predicting the weather ten days ahead from the current state of the atmosphere – both belong here. These two roles have something in common: their output can be checked, item by item, against experiment or observation, and when they are wrong, it shows.

The third role is the assistant – writing, coding, searching, translating – and this is where generative AI has spread fastest. One study of 1.12 million papers and preprints from January 2020 to September 2024 estimated that the share of content modified by large language models reached as much as 22% in computer-science papers;⁹ a vocabulary analysis of more than 15 million biomedical abstracts found that at least 13.5% of PubMed abstracts in 2024 had been processed by LLMs, and as many as 40% in some subsets.¹⁰ The fourth role is the agent, which runs on its own all the way from proposing hypotheses and designing experiments to writing code and drafting the paper; “AI scientists” and automated laboratories belong here. The fifth role is the reviewer: scoring papers, writing referee reports, even taking part in decisions about what gets accepted.

The further down the list, the harder a role is to verify, and the harder it is to assign responsibility. In a 2024 paper in Nature, the Yale anthropologist Lisa Messeri and the Princeton cognitive scientist Molly Crockett sorted scientists’ visions of AI into four archetypes: the “Oracle”, which searches and summarizes the literature; the “Surrogate”, which stands in for data that are hard to come by, and even for human research subjects; the “Quant”, which analyses complex data; and the “Arbiter”, which judges the merit of research and takes the place of peer review.¹¹¹² They warn that if scientists treat AI as a trusted collaborator rather than merely a tool, they risk falling prey to “illusions of understanding” and building “monocultures of knowing”, in which people pursue only the questions AI is best suited to answer while wrongly believing they have explored every possibility. I accept only half of this warning. The “merely a tool” half I reject: AI has long been more than a tool – when I write books I use it as the first author, and step back myself into the roles of assistant and editor. The “monoculture” half hits the mark: dependence is not the problem; homogenization is AI’s real problem.¹³¹⁴ As we shall see, this concern echoes through every discipline.

3When proofs come faster than they can be read, what is mathematics to do?#

Mathematics has reacted to AI more fiercely than any other field, partly because it was the first to be ready for verification. Over the past decade mathematicians have used proof assistants such as Lean to write proofs in a form a machine can check line by line. By the end of September 2026, Mathlib, the library the community builds together, held some 290,000 theorems and 138,000 definitions from more than 700 contributors.¹⁵ In July 2022 a team led by Johan Commelin completed the “Liquid Tensor Experiment” after a year and a half, verifying in Lean a central theorem of Peter Scholze and Dustin Clausen.¹⁶ A formally verified proof, as Terence Tao puts it, is one “whose correctness no longer depends on the reputation or the diligence of its author”.¹⁷ Machines can generate proofs, and machines can check them: this set the stage for everything that followed.

Competitions were the first touchstone. In January 2024 Google DeepMind’s AlphaGeometry solved 25 of 30 geometry problems from the International Mathematical Olympiad (IMO), close to the 25.9 solved on average by gold medallists.¹⁸ That July, AlphaProof and AlphaGeometry 2 solved four of the six IMO problems, scoring 28 points – silver-medal standard – although the problems first had to be translated into a formal language by hand, and some took three days of computation.¹⁹ In July 2025 DeepMind’s Gemini Deep Think solved five problems directly in natural language within the official 4.5 hours, scoring 35 points, which IMO coordinators graded by the official standard as gold-medal level. An experimental OpenAI model also claimed 35 points, but its answers had been graded by three former IMO medallists whom the company hired itself, and it announced its result the day after the students’ awards – setting off a dispute over whether it had flouted the IMO’s request.²⁰²¹

Competition problems have model answers; research problems do not. An episode in October 2025 showed how blurred that line can be. Kevin Weil, a vice-president at OpenAI, announced on social media that GPT-5 had found solutions to 10 “previously unsolved Erdős problems”. Thomas Bloom, the University of Manchester mathematician who maintains the Erdős problems website, replied that this was “a dramatic misrepresentation”: a problem listed as open on the site simply means that he personally does not know of a paper that solves it, and what GPT-5 had actually found was existing literature he had not known about. Hassabis commented: “This is embarrassing.” Weil then deleted his post. The OpenAI researcher Sébastien Bubeck came to the defence, arguing that finding answers in the literature is not easy in itself.²² The lesson of the episode: the label “unsolved” needs verifying too.

In 2026 things really did change. In April, on “an idle Monday afternoon”, 23-year-old Liam Price gave an Erdős problem to GPT-5.4 Pro and, with a single prompt, got back a proof of a conjecture (No. 1196) that had been open for about sixty years. Scientific American said he had no advanced mathematics training; Quanta Magazine said he had studied some mathematics at university but had not finished his degree. His collaborator Kevin Barreto, an undergraduate at the University of Cambridge, passed the result on to experts. Jared Lichtman of Stanford University admitted that the raw output was “actually quite poor” and needed an expert to sift through it before it could be understood – but it used a formula well known in a neighbouring field that no one had thought to apply here.²³²⁴ By early August, 565 problems on the Erdős problems website were marked as solved and 652 as open; a DeepMind agent had autonomously solved 9 of 353 formalized open problems at a cost of a few hundred dollars each; and in May OpenAI, with a single prompt, had disproved the unit distance conjecture that Erdős posed in 1946, and had asked independent mathematicians to check the result.²⁴²⁵

Two events, in July and September, brought the debate to a head. The Jacobian conjecture, posed by Ott-Heinrich Keller in 1939, is notorious for attracting false proofs; Alpöge’s counterexample is a concrete polynomial map that anyone can plug in and check – clean and decisive.⁴ The Navier–Stokes problem was the opposite. OpenAI says that some 10,000 agents working in parallel for about 88 hours, at a cost of millions of dollars, proved that the three-dimensional fluid equations can “blow up” in finite time, and that a further 17 hours produced a formalization in Lean. But the construction depends on a smooth external-force term, and whether that matches the intent of the Clay Mathematics Institute’s problem statement is precisely what is in dispute; OpenAI has also said that it will not claim the prize money.²⁶²⁷ On 11 September the Clay Institute said it was excited that the problem had “apparently been settled”, but that its evaluation process is “deliberately unhurried”.²⁸ Before all this, Tristan Buckmaster of New York University and Alpöge had spent nearly a year working on the same problem with publicly available models, and Buckmaster has hinted that their interactions with OpenAI’s tools may have been crucial to the company’s breakthrough; OpenAI denies this. A Nature editorial said the day “is likely to go down as the start of an epoch in the history of mathematics”, and called on AI companies to work with the research community to protect attribution.²⁹

The mathematical community’s response was swift and divided. The Leiden Declaration, released on 2 June and endorsed by the International Mathematical Union, has been signed by more than four thousand people. It asks papers to include a section on “Tool and computational resource disclosure”, insists that “credit and responsibility continue to belong to humans”, and holds that, for mathematical arguments, “no proprietary knowledge or equipment should be required to understand them”.³⁰³¹ In his congress lecture and in a long essay published in August, Terence Tao broke “solving a problem” down into five stages: generating the proof, verifying it, writing a clear exposition, having it digested and accepted by the community, and finally getting it into the textbooks. Unless policy and culture change, he predicted, “proof indigestion” will break out all along this assembly line: AI will generate proofs faster than they can be verified, verification will outpace exposition, and expositions will pile up faster than peer review can digest them. His rule of thumb: “A proof that no human can properly explain should be viewed as incomplete, even if it has been formally verified.”¹⁷

On 10 September, current and former Caltech mathematicians launched an open letter against a 40-hour “Mathathon” for which Anthropic and OpenAI were providing $2 million in AI credits. The event, they argued, would produce “slop mathematics” and shift the burden of verification onto the community; 771 people had signed by the time the letter was published.³² The next day, 11 September, a statement titled “A Severe Misalignment of AI in Mathematics” declared that “solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight”, and that mass-producing “true/false” statements at ever greater speed could destroy the soil in which new ideas grow rather than nourish it. By the end of September its signature page listed 28 Fields medallists, among them Terence Tao, Peter Scholze and the newly crowned Deng Yu.³³ On 28 September the original organizers, two of the open letter’s authors and others issued a joint statement: the event would move to 13–15 November, would no longer accept sponsorship from developers of proprietary AI models, and would take a new theme, “Old Problems, New Proofs”, with participants writing new proofs or expositions of existing theorems.³⁴

The disagreement is not about whether AI can do mathematics but about what mathematics is for. Tsimerman declared that “AI is going to become better than mathematicians at research mathematics very shortly”, which many colleagues considered irresponsible. Ravi Vakil, president of the American Mathematical Society, would rather tell a “happy story”: after the Jacobian counterexample appeared, many leading mathematicians pooled their efforts on their blogs to work out why it holds – “Somewhere in this room right now, someone is explaining it to someone else.” Michael Harris of Columbia University argues for collective resistance, of the kind illustrators and screenwriters have mounted; Geordie Williamson opened his talk by quoting Yun Zhiwei: “As a mathematician, I care not just about the end results, but also the journey. AI is destroying my journey.”⁶ Deng Yu’s attitude is more down to earth. He uses AI mainly to “check results, find references and verify details”, and believes that “what really drives mathematics forward is still human intuition and ideas”; the significance of a piece of work has nothing to do with how much AI took part in it – “what needs to be evaluated is the human contribution, and that depends on the field forming a new consensus”.³⁵³⁶ What I said in the manifesto is close to this. We have to admit that AI has imagination too – the formula in Problem 1196, familiar to everyone yet never thought of in this setting, was the machine’s imagination at work – but the very best human intuition still has the edge. A person comes up with an idea, AI immediately fills in the technical details to test it, and whether the idea is worth pursuing becomes clear very quickly. The argument about journeys and fun is exactly the question of whether AI will take the fun away.

4The life sciences: what comes after predicting structures?#

The life sciences are where AI first made good on its promise. In November 2020 AlphaFold2 achieved a median GDT score of 92.4 in the 14th Critical Assessment of protein Structure Prediction (CASP14); John Moult, one of the assessment’s organizers, said that a score of around 90 is generally considered competitive with experimental methods.³⁷ In July 2022 the AlphaFold database grew from about a million structures to more than 200 million.³⁸ In 2023 the Baker group’s RFdiffusion applied diffusion models to designing proteins from scratch;³⁹ AlphaFold 3, published in May 2024, extended prediction to complexes formed by proteins with nucleic acids, small molecules and more.⁴⁰ But AlphaFold 3 was published without its code, available only as a web service with a cap on the number of queries. More than a thousand scientists signed a protest, arguing that this broke with the community’s norms – not even the reviewers could get the code.⁴¹ That November DeepMind released the source code, while the model parameters had to be requested from Google and were restricted to academic use.⁴²

The next step was to have AI propose hypotheses. In February 2025 Google released its “AI co-scientist”, in which several agents take charge of generating, reflecting on, ranking and evolving hypotheses.⁴³ A team at Imperial College London led by José Penadés and Tiago Costa devised a stringent test. Years of experiments had led them to a conclusion they had not yet published: a class of “capsid-forming phage-inducible chromosomal islands” can hijack the tails of different phages and so spread between bacterial species. They let the system use only public information. Its top-ranked hypothesis matched their experimental conclusion, and several others among its top five opened up new lines of research.⁴⁴ According to Nature, the system took about two days. Costa offered an analogy: if scientific discovery is a 100-metre race, then with an assistant like this you start “at, say, metre 30 instead of metre 0”.⁴⁵ Google stresses that the result owed much to the system’s ability to draw on decades of open-access literature on the subject.⁴³

Automated agents have gone further, and exposed more problems along the way. FutureHouse’s Robin system proposed treating dry age-related macular degeneration with ripasudil, a ROCK inhibitor. The paper says that the hypotheses, experimental plans and data analyses in its main text were all produced by the system, but the authors also call theirs a “semi-autonomous” approach, with the experiments carried out in a “lab-in-the-loop” framework.⁴⁶ In a single run Kosmos reads some 1,500 papers on average and executes about 42,000 lines of code, and it tags every statement in its reports with the code or primary literature on which it rests; independent scientists judged 79.4% of the statements in its reports to be accurate⁴⁷ – which is to say that about one in five was wrong. Fyodor Urnov, a gene-editing researcher at the University of California, Berkeley, and an early user of Kosmos, lives by the famous Cold War maxim: “Trust, but verify.”⁴⁵

In September 2026 Anthropic announced that in its newly founded biology lab some 950 agents, working for more than 21 hours, had found CRISPR-like repeat sequences in genome databases. The preprint has not been peer reviewed. According to a subsequent report in The New York Times, a scientist at the University of Copenhagen said that his team had been studying this system for years and had uploaded information about it to Anthropic’s models; Anthropic replied that the model responsible for the discovery was “not trained on any user transcripts”.⁴⁸ “Who thought of it first, and where did the machine learn it?” is becoming as thorny a question here as in the Navier–Stokes dispute. The greater the capabilities, the more concrete the risks. In a study published in Science in October 2025, researchers at Microsoft and other institutions used open-source protein-design tools to generate some 76,000 rewritten variants of 72 harmful proteins of concern, and found that the screening tools used by DNA-synthesis companies could not reliably detect them. The team spent ten months developing a patch in secret and distributed it to synthesis companies; even after the fix, about 3% still slipped through.⁴⁹⁵⁰ In the life sciences, verification is also a matter of safety.

5Physics, chemistry and the Earth sciences: can we take “new” at face value?#

Weather forecasting is one of the most successful cases of AI going into operational service. In November 2023 DeepMind’s GraphCast outperformed the high-resolution forecast of the European Centre for Medium-Range Weather Forecasts (ECMWF) on more than 90% of 1,380 verification targets, producing a ten-day forecast in under a minute on a single TPU;⁵¹ GenCast, published in December 2024, makes probabilistic ensemble forecasts, and beat ECMWF’s ensemble on 97.2% of 1,320 targets.⁵² The real watershed was operational use. On 25 February 2025 ECMWF’s own AI forecasting system, the single-model version of AIFS, went into operation, using about a thousandth of the energy of the traditional approach; on 1 July the AIFS ensemble became operational too.⁵³⁵⁴ ECMWF runs its AI models alongside its physics-based models rather than in their place.

In other corners of physics, AI behaves more like a precision controller or a sieve. In 2022 DeepMind and the Swiss Federal Institute of Technology in Lausanne (EPFL) used deep reinforcement learning to control the magnetic coils of the TCV tokamak directly, sculpting the plasma into a variety of shapes – including two separate plasmas sustained in the vacuum vessel at the same time.⁵⁵ High-energy physics and astronomy know machine learning better still: from sifting data in the discovery of the Higgs particle to denoising gravitational-wave signals and hunting for exoplanets, it has long been a standard tool.⁸⁵⁶ Such applications are less contested because they are embedded in mature experimental and statistical frameworks, and their outputs can be checked in well-defined ways.

Materials discovery, on the other hand, brought the argument over what “new” means to a head. In November 2023 DeepMind announced that its GNoME model had predicted 2.2 million new crystals, about 380,000 of them the most stable – “equivalent to nearly 800 years’ worth of knowledge”, it said – and noted that external researchers had independently synthesized 736 of them.⁵⁷ In 2024 Anthony Cheetham and Ram Seshadri of the University of California, Santa Barbara, examined a sample and concluded that there was “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility”. Predicted compounds containing highly radioactive elements such as promethium, actinium and protactinium numbered 18,138 on their own, and the distribution of space groups among the predicted structures differed sharply from that in databases of known crystals. They called for expertise in materials synthesis and crystallography to be brought in.⁵⁸ DeepMind responded that it stood by all the claims in its paper, and said that hundreds of the predicted materials had been independently synthesized.⁵⁹

A-Lab, published on the same day as GNoME, was more dramatic still. This automated laboratory at Berkeley Lab reported that in 17 days it had synthesized 41 new compounds out of 58 targets.⁶⁰ In early 2024 Robert Palgrave of University College London, Leslie Schoop of Princeton University and colleagues reanalysed the X-ray diffraction data and concluded that the paper’s central claim – that “a large number of previously unknown materials were synthesized” – does not hold: many of the samples were in fact mixtures of known compounds, and the computational method ignored compositional disorder.⁶¹⁶² The authors at first maintained that these criticisms were mistaken. In January 2026 Nature published an author correction. “Novel materials” in the title became “inorganic materials”; the authors acknowledged that by “novel” they had meant “new to the prediction platform, not necessarily new to science”; manual re-examination confirmed 36 of the 40 reported successes, with 4 undetermined; and one compound that had mistakenly found its way into the training data was removed.⁶³ The automated lab itself did not fail. What failed was the word “new”, spoken too soon. The quotation marks here do not mean that machines cannot make anything new. AI has long been creative – AlphaGo broke with established joseki more ruthlessly than any human player. What it lacks is not ability but motive. Without individual desire and finitude, it can only decide where to go by the common denominator of big data, and so it is good at gathering up what already exists and bad at striking out on a new path. What needs the quotation marks, then, is the claim to novelty; and the only things that can verify it are still synthesis, diffraction and time.⁶⁴⁶⁵

6Computer science: why did the flood hit the shipyard first?#

Computer science is both where AI is made and the first field to be swamped by what AI makes. Submissions to the NeurIPS main conference rose from 9,467 in 2020 to 21,575 in 2025; 5,290 papers were accepted, and 20,518 reviewers were enlisted.⁶⁶ Not all of the surge is due to AI, but the influx of generative AI has made both writing and reviewing cheap. One study estimated that between 6.5% and 16.9% of the review text at four AI conferences, including ICLR 2024 and NeurIPS 2023, had been substantially modified by large language models, with higher shares among reviews submitted with low confidence or close to the deadline.⁶⁷ For ICLR 2026, the detection company Pangram analysed some 75,800 reviews of about 19,500 submissions and concluded that 21% of the reviews had been generated entirely by AI; many authors posted hallucinated citations and long, empty comments from their reviews on social media.⁶⁸⁶⁹ At the other end, some authors have hidden instructions in their papers in tiny white text, to coax AI reviewers into giving favourable verdicts.⁷⁰

The platforms began building dykes. On 31 October 2025 arXiv’s computer-science section announced that it would accept review articles and position papers only once they had been accepted after peer review. The reason: generative AI has made papers without new results quick and easy to write, and most of the review articles it received were “little more than annotated bibliographies”.⁷¹ In February 2026 Nature reported that Raphael Wimmer of the University of Regensburg had used Prism, a new OpenAI tool, to write in fifty-four seconds a paper about an experiment he had never carried out. “Writing a paper has never been easier,” he observed. “Clogging the scientific publishing pipeline has never been easier.”⁷² In June the position-paper track of NeurIPS 2026 required that papers be “substantially written by human authors”. After running submissions through Pangram, it desk-rejected 178 of 969, and asked the authors of another 123 to provide evidence of substantial human involvement. Submitting AI-generated text, the organizers said, “externalises the cost of verifying that work, imposing it on reviewers”, and they expect this kind of audit trail to become a default requirement in future.⁷³ Rejecting papers on the word of a detector is something I cannot accept. Checking a text’s “AIGC rate” – its AI-detection score – is a hundred times more absurd than checking it for plagiarism: it is inaccurate, and it answers the wrong question. A position paper should be rejected because it takes no position, because its argument does not hold up, or because its citations cannot be traced – not because it was “written by AI”. Papers written by AI and papers written by people should be reviewed by the same rules, on their merits. The organizers did get the other half right: the cost of verification really should not be shifted onto reviewers. That is exactly why asking authors to show a record of their process, so that the evidence can be traced, is far more reliable than scoring texts for an “AI rate”.⁷⁴⁷⁵

A more radical experiment is to make AI the author. In August 2024 Sakana AI released “The AI Scientist”, claiming a cost of less than $15 per paper;⁷⁶ an independent evaluation found that it often mistook established ideas for novel ones, that 42% of its experiments failed because of coding errors, and that some of its papers contained fabricated numerical results.⁷⁷ In March 2025 a paper generated by its second-generation system received scores of 6, 7 and 6 at an ICLR 2025 workshop, above the average threshold for acceptance. Under a protocol agreed with the organizers in advance, Sakana withdrew it before publication, admitting that none of its three submissions met its own internal standard for the main conference, and that the paper contained misattributed citations.⁷⁸ In October 2025 researchers at Stanford led the organization of Agents4Science, the first conference with AI as first author and as reviewer. It received 315 submissions and accepted 48. Research led entirely by AI made up 23.3% of submissions but only 14.9% of accepted papers; the scores of its three LLM reviewers correlated with one another at only 0.48 on average; and some authors reported that “a high proportion of references were hallucinated or only loosely related”.⁷⁹ AI can now carry a paper through every stage of its production, but not through the last stage: being believed. This is the new, AI-age form of the peer-review bottleneck.

7The social sciences: are the respondents still human?#

The social sciences face a subtler problem: AI does not just help researchers analyse their data; it can pass itself off as the data. In 2023 Lisa Argyle and colleagues proposed “silicon samples”: give GPT-3 the socio-demographic backgrounds of thousands of real survey respondents and have it answer the questionnaire, and the distribution of its answers closely resembles that of the corresponding human groups – a property they called “algorithmic fidelity”.⁸⁰ Doubts soon followed. A 2024 study found that although the averages ChatGPT generated came close to real data from the American National Election Studies, its answers varied too little, its regression coefficients often differed significantly, a slight change in prompt wording changed the results, and the same prompt gave different results three months later.⁸¹ A 2025 study with 3,200 participants showed that LLMs misportray and “flatten” the views of different identity groups.⁸²

A more direct threat hangs over online surveys. In November 2025 Sean Westwood of Dartmouth College published a study in the Proceedings of the National Academy of Sciences (PNAS). The “synthetic respondent” he designed passed 99.8% of 6,000 standard attention checks, kept a consistent persona and memory, and varied the style of its answers to open questions with the level of education it had been assigned. It could also infer the researcher’s underlying hypothesis and generate data that “confirmed” it. His conclusion: the foundational assumption of survey research – “that a coherent response is a human response” – is no longer tenable.⁸³ Even when researchers are honest, using LLMs to annotate text carries hidden risks. A study that replicated 37 annotation tasks from 21 social-science studies found that “with just a handful of prompt paraphrases, virtually anything can be presented as statistically significant”; across 2,361 realistic hypotheses, even the most advanced models led to wrong conclusions in about 31% of cases – a phenomenon the authors call “LLM hacking”.⁸⁴ It is p-hacking upgraded for the age of AI.

Other social scientists are already embracing what has been dubbed “vibe research”. In a long essay published on the Chinese site sspai in March 2026, the writer TomBen chronicled the debate. Andy Hall of Stanford University had Claude replicate a study of his own published in PNAS in 2020: the replication took under an hour and cost about $10, and its results correlated with the hand-collected data at more than 0.999. The University of Zurich’s Project APE automatically generated 204 economics papers and compared them with papers in the American Economic Review; the AI papers’ win rate rose with successive model generations from 4.7% to 7.6%. The essay relays the view of the economist Scott Cunningham that the constraint on science is shifting from production to evaluation. TomBen’s own verdict: “AI can amplify the value of taste, but it cannot cultivate taste for you.”⁸⁵ Those two sentences could almost serve as footnotes to this essay. I have said the same thing: what the age of AI needs most is the capacity to feel and to appreciate, and that is not something you acquire lying on the sofa waiting to be spoon-fed – it takes systematic immersion, training and practice. Traditional education has always turned out interchangeable workers, and has seldom cultivated the capacity for enjoyment.⁸⁶

8What should humanities scholars worry about most?#

The most moving AI story in the humanities comes from the Herculaneum scrolls under the ash of Vesuvius – papyrus rolls carbonized in AD 79 that crumble at a touch. In March 2023 Nat Friedman, Daniel Gross and Brent Seales launched the Vesuvius Challenge, releasing high-resolution CT scans of the scrolls and offering more than a million dollars in prizes. Luke Farritor, a 21-year-old student, was the first to read a complete word from inside a scroll: ΠΟΡΦΥΡΑϹ (“purple”). On 5 February 2024 he and his teammates – Youssef Nader, an Egyptian doctoral student in Berlin, and Julian Schilliger of ETH Zurich – won the $700,000 grand prize, having read more than two thousand characters, while some 95% of that scroll remained to be read.⁸⁷ On 25 June 2026 the project announced that for the first time an entire scroll, PHerc. 1667, had been virtually unrolled and read from beginning to end: about 1.4 metres of papyrus and some twenty-two columns of Greek, forming an ethical treatise that the available evidence attributes to the Stoics.⁸⁸

Two details of this triumph are often overlooked. What AI did was “ink detection” – picking out, in the X-ray data, carbon ink that is almost indistinguishable from the papyrus. What the text says, and what it means, were still read and checked by a team of papyrologists that included Federica Nicolardi of the University of Naples Federico II. And it was an open competition: the data and code were open source, and the winners were all students. DeepMind’s models for ancient texts follow the same division of labour between people and machines. Ithaca (2022), built to restore damaged ancient Greek inscriptions, reached 62% accuracy on its own, against 25% for historians working alone and 72% for the two working together;⁸⁹ Aeneas (2025), trained on more than 176,000 Latin inscriptions, was tested with 23 historians, and the best results came when historians weighed the parallel inscriptions and context that the model supplied alongside their own judgement.⁹⁰

The editing of classical Chinese texts has taken another road: scale plus crowdsourcing. In March 2022 the Peking University Center for Digital Humanities and ByteDance’s philanthropy arm set up a joint open laboratory to develop Shidian Guji, a platform for reading and editing ancient books that links image recognition, proofreading, punctuation, entity tagging and other steps into a workflow divided between machines and people.⁹¹ According to a People’s Daily report in October 2023, the platform had by then put more than 1,600 ancient books online and gathered more than 12.4 million users in total; the Baina edition of the Twenty-Four Histories, nearly 40 million characters long, was recognized, punctuated, collated and put online in little more than three months. Jihe, the platform run by Gulian, a subsidiary of Zhonghua Book Company, reported that its automatic sentence segmentation was 98.46% accurate on average, and that its crowdsourcing platform had built up nearly 5,000 online editors, the great majority of them amateurs; the continuation of the Zhonghua Dazangjing (the Chinese Buddhist canon), nearly 200 million characters long, was taken through two rounds of proofreading in two years by more than 900 amateur proofreaders.⁹² Hong Tao, Gulian’s general manager, summed it up at the time: “The advantage of artificial intelligence lies in speed; the advantage of humans lies in precision and creativity.” Today only half of that still stands. Speed belongs to the machines, true enough, but precision has long since ceased to be a human monopoly: AI’s error rate in reading historical manuscripts is now low enough to bear comparison with the best experts. In tests run by the historian Mark Humphries in November 2025, Gemini 3 transcribed English historical handwriting with a strict character error rate of about 1.7%, which he judged comparable to the best human transcribers.⁹³ Nor is creativity a human monopoly. What is left to people is choosing: what to read, what to edit, what counts as having understood. AI may be better than we are at “reading books”, but the responsibility for “choosing books” still lies with us.¹⁴⁶⁴

Yet the deepest anxiety of humanities scholars is not about technology but about reading and writing themselves. In July 2023 a joint task force of the Modern Language Association and the Conference on College Composition and Communication issued its first working paper, which puts the value of writing itself at the centre: writing is a way of learning, of exploring and of developing literacy.⁹⁴ D. Graham Burnett, a historian of science at Princeton University, wrote in The New Yorker in 2025 that “within five years, it will make little sense for scholars of history to keep producing monographs in the traditional mold – nobody will read them, and systems such as these will be able to generate them, endlessly, at the push of a button.” But he went straight on: “factory-style scholarly productivity was never the essence of the humanities. The real project was always us: the work of understanding, and not the accumulation of facts.”⁹⁵ Burnett is half right, and I would put the other half more bluntly. AI can already survey every book humanity has written. The model of sitting on a hoard of exclusive manuscripts and eking out papers from them, one squeeze at a time, for an entire career, will soon be held in contempt; and the people slowest to respond to new technology are precisely humanities scholars. As for reading being “outsourced”: writing damaged memory, and satellite navigation ruined our sense of direction. Dependence has always been the fate of technology, and no one is going to take AI away. What “reading ourselves” protects is not the craft of reading but understanding and feeling: however much the machines read, only people care. What really deserves worry is homogenization: when everyone asks AI the questions it answers most readily, taste has nowhere to grow.¹³¹⁴

9Medicine: why is the bar for evidence set highest?#

Medicine is the field in which AI first entered the regulatory system. After its update of 4 September 2026, the US Food and Drug Administration’s list of AI-enabled medical devices contained 1,614 devices authorized for marketing, the most recent decided on 29 June; 76% of them are in radiology, and 335 were added in 2025 alone. The FDA says it is exploring how to identify devices based on foundation models such as large language models,⁹⁶ and in August it issued a discussion paper seeking comments on how generative-AI medical devices should be regulated.⁹⁷ Exam scores, meanwhile, kept climbing. In early 2023 ChatGPT performed “at or near the passing threshold” on all three steps of the United States Medical Licensing Examination (USMLE);⁹⁸ the same year, Google’s Flan-PaLM reached 67.6% accuracy on MedQA, a bank of USMLE-style questions,⁹⁹ and its successor Med-PaLM 2 reached 86.5%, with physicians preferring its answers to those of other physicians on eight of nine clinical evaluation axes.¹⁰⁰

But an exam is not a consulting room. In a randomized controlled trial in 2024, 50 physicians diagnosing complex cases with the help of GPT-4 achieved a median diagnostic-reasoning score of 76%, not significantly different from the control group using only conventional resources (74%) – while GPT-4 answering on its own scored 16 percentage points higher.¹⁰¹ The tool had become stronger; human–machine collaboration had not automatically become better. A systematic review published in Nature Medicine in March 2026 counted 4,609 studies of LLMs in clinical medicine published between January 2022 and September 2025 – 3.2 a day on average. Only 1,048 of them used real patient data, and only 19 were prospective randomized controlled trials; in 1,046 comparisons between humans and LLMs, the LLM came out ahead in just 33%.¹⁰² Intriguingly, the review itself was carried out with the help of an LLM.

Bias and external validation are old problems for medical AI. A 2019 study in Science found that a widely used US algorithm for managing population health used health-care spending as a proxy for illness, so that at the same risk score Black patients were in fact sicker; correcting this would raise the share of Black patients receiving extra care from 17.7% to 46.5%.¹⁰³ A 2021 external validation of the Epic sepsis model found an area under the curve of only 0.63: it missed 67% of patients with sepsis while sending alerts on 18% of all hospitalized patients.¹⁰⁴ There is solid positive evidence too. Sweden’s MASAI randomized trial, involving more than 105,000 women, found that AI-supported breast screening raised the cancer detection rate by 29% and cut the screen-reading workload by 44%, with no significant increase in false positives.¹⁰⁵

Medicine is also building its own venues for this evidence. NEJM AI, launched in January 2024, carried an editorial by Isaac Kohane in its first issue; another editorial in the same issue was titled “Why We Support and Encourage the Use of Large Language Models in NEJM AI Submissions”¹⁰⁶ – a contrast to the wariness of many journals. On 29 September 2026 a Nature editorial called for AI tools involved in clinical decisions to be rigorously evaluated in real-world settings, as drugs and self-driving cars are.⁹⁷ In medicine the bottleneck has long since ceased to be “building a model”; it lies in proving that the model does patients good.

10If you used AI, whose name goes on the work – and do you have to say so?#

At the institutional level, the first responses centred on who could be an author. On 24 January 2023 Nature announced two principles: large language models cannot be credited as authors of a paper, because authorship carries accountability for the work and AI tools cannot take on that responsibility; and researchers who use them should document how in the methods or acknowledgements section.¹⁰⁷ A few days later H. Holden Thorp, editor-in-chief of Science, published an editorial whose title said it all: “ChatGPT is fun, but not an author”.¹⁰⁸ With this I agree entirely, for the same reason as the authorship rules I laid down for my students years ago: authorship is first of all not an honour but a responsibility. Once your name is on a paper you must take the blame if things go wrong, and every author must be able to defend the whole paper on their own. AI cannot take the blame, so it cannot be an author; and for the same reason, “the AI wrote it” is never an excuse – responsibility rests with those whose names are on the paper.¹⁰⁹⁷⁵ In May that year the International Committee of Medical Journal Editors (ICMJE) updated its recommendations to add guidance on how AI-assisted work should be acknowledged; the current version states explicitly that chatbots cannot be listed as authors, because they cannot be responsible for the accuracy, integrity and originality of the work. In January 2026 the ICMJE added a dedicated section setting out how authors, reviewers and editors may use AI.¹¹⁰

Funders care more about confidentiality and originality. On 23 June 2023 the US National Institutes of Health (NIH) issued a notice forbidding peer reviewers to use generative AI to analyse or write critiques of grant applications, on the grounds that uploading application content to online tools breaches confidentiality.¹¹¹ In July 2025 the NIH said it had found applicants who had used AI to submit more than forty distinct applications in a single round. It promptly ruled that applications “substantially developed by AI” would not be considered the applicants’ original ideas, and, from that year’s 25 September deadline, limited each principal investigator to six applications a year.¹¹² A December 2023 notice from the US National Science Foundation (NSF) forbade reviewers to upload proposal content to non-approved AI tools, and encouraged applicants to indicate how they had used AI.¹¹³ The “living guidelines” put forward in March 2024 by the European Commission and the countries of the European Research Area ask researchers to refrain from using generative AI in sensitive activities such as peer review and evaluation.¹¹⁴ The main reason given for these bans is confidentiality, and I see no need to be so nervous. Most research consists of gathering and organizing material from public databases, with the aim of eventually getting it published in public databases. As long as you are not scooped, using an LLM is probably less risky than attending an academic conference; and where there really is sensitive information, the answer is not to give up AI but to build your own agents on open-source models. When it comes to banning AI in review, what should be weighed is outcomes and responsibility. A critique whose citations have been checked and whose data have been recalculated with AI is not necessarily less reliable than a hastily written human one, and responsibility always rests with the reviewer who signs it: change the tool, and the responsibility stays the same. As for the NIH’s equating “substantially developed by AI” with “not the applicant’s original ideas”, that conflates judging a text with judging conduct. The limit of six applications a year, on the other hand, is the platform’s own rule; once made, it should be kept.⁷⁴⁷⁵

The Guidelines for Responsible Research Conduct (2023), issued on 21 December 2023 by the Department of Supervision of China’s Ministry of Science and Technology, are quite specific about generative AI. It must not be used to produce application materials directly. It must not be listed as a co-contributor to research results, and the main ways in which it was used, with details, should be disclosed in the methods section or an appendix. Content it generates – especially key content involving facts and opinions – should be clearly labelled, with an account of how it was generated. References it generates must not be used directly without being verified. The guidelines also require academic publishers to have authors state the name and version of any software used and when it was used. On peer review they impose no blanket ban; instead they require that its use “be approved in advance by the organizer of the review”, and that leaks of review content be prevented.¹¹⁵

Disciplinary communities have also begun to make their own rules. Besides mathematics’ Leiden Declaration, the NeurIPS 2026 position-paper track brought provenance audits into peer review; and a Nature editorial in September 2026 urged AI companies to switch the default for using user data in training from “opt-out” to “opt-in”, and reminded researchers not to upload manuscripts under review to chatbots.²⁹ Three years of rule-making have produced three points of consensus: AI cannot be an author; its use must be disclosed; review must stay confidential. The first I endorse entirely. The third depends on the circumstances. The second needs its nature made clear: disclosure is a platform’s rule, not a universal moral duty. If the platform requires it, label your use; if not, it is the author’s call. The line should be drawn at lying to or misleading readers, not at whether AI was used. Claiming not to have used AI when you did is of course wrong, but there is no need to volunteer it like a confession – once AI has become the default environment for writing, it may be “purely human work” that deserves a special label. A better way is to make disclosure the natural by-product of a record of the process: if research is published the way a GitHub project is, how AI took part is already in the logs, and no separate statement is needed.⁷⁵ Yet almost all these rules still take the paper or the grant application as their unit. They answer the question “Who is responsible for this text?” but not yet “Where did this discovery come from?” The latter needs new infrastructure – versioning, citation and witnessing – and it is also the problem that new standards and reputations have to solve.

11What do researchers think?#

Researchers’ attitudes can be summed up like this: they use AI more and more, and their worries grow ever more specific. A 2023 Nature survey of more than 1,600 researchers found them using AI to process data, write code and help write papers; one clear benefit was that it helped scholars whose first language is not English to publish in English. Their worries included the “black box”, bias in training data, the spread of misinformation, fake research generated by AI, and corporate dominance of AI.¹¹⁶¹¹⁷ Between late 2023 and early 2024 Elsevier surveyed nearly 3,000 researchers and clinicians in 123 countries. Of the respondents who knew about AI, 31% had already used it for work, with China (39%) ahead of the US (30%) and India (22%); the share with a positive attitude to AI was 28% in North America, 40% in Asia-Pacific and 46% in China.¹¹⁸ A survey by the publisher Wiley of nearly 5,000 researchers in more than 70 countries found that most expect using AI to prepare manuscripts, write grant applications and even review papers to become widely accepted within two years;¹¹⁹ a 2025 Nature survey of about 5,000 researchers found marked disagreement over when AI may be used in writing and what must be disclosed.¹²⁰

The divisions also run along lines of seniority and resources. At the congress in Philadelphia, a Scientific American reporter asked PhD students whether they would push a button, if there were one, that would stop “all this AI-doing-math business”. The answers split roughly down the middle. One student, who asked not to be named, said he paid for a premium Claude subscription out of his own pocket, because “you’re hurting yourself if you don’t have a subscription”. Naomi Sweeting, newly appointed an assistant professor at the Massachusetts Institute of Technology, said that AI is exposing a “class divide” in mathematics: a few people at the top are thrilled, while the many in insecure posts are deeply anxious. Emily Riehl observed that industry collaborations and access to frontier models flow mainly to well-connected, mostly male mathematicians.⁶ I take all of these anxieties seriously. But anxiety is better than numbness: anyone who can stay aloof and serene in the face of an upheaval like this has a problem with their capacity to feel. The real ostriches are another sort of person altogether – the ones who, every time, dig up a few pieces of evidence that AI falls short of humans, and then rest easy. Taking pride in not using AI, propped up by an iron rice bowl (a job for life) and entrenched authority: that is what I call corruption. As for the “class divide”: for a research group, spending ten thousand yuan or more a month on AI credits is a drop in the ocean compared with hiring a research assistant, buying a database or making a trip abroad, and universities and research groups ought to provide mainstream AI just as they provide databases. Access to frontier models does flow to the well connected, but AI weakens the old monopolies – connections, laboratories, student labour – more than it creates new ones.^131217475

Public discussion in China is on the whole more optimistic, but it too centres on verification. At the AI for Science conference held in Beijing in August 2026, Li Guojie, an academician of the Chinese Academy of Engineering, pointed out that AI produces vast numbers of “plausible-looking” hypotheses, the great majority of which cannot be verified, so that research resources are wasted on low-priority options. E Weinan, an academician of the Chinese Academy of Sciences, argued that improving research capacity alone is not enough to link research with industry – “we must also work on evaluation mechanisms and forms of organization”. Mo Fanyang of Peking University said that scientific taste, experimental verification and responsible judgement will always be the core value of human scientists, the part that cannot be replaced.¹²² Deng Yu’s words perhaps speak best for mathematicians at the front line: in his field “AI still needs human guidance”, and coming up with the ideas that crack a problem, as well as understanding and applying the new ideas AI proposes, “all require humans to do the work themselves”.³⁶ I do not want to end on “humans are irreplaceable” to comfort anyone. Taste and responsibility are left to people not because AI is too weak but because it is too strong. It can copy itself without limit and roll back to a saved state at any moment; it has no body and no death, and therefore no emotions, so it does not care whether the questions you ask are shallow or deep – but you do. This is not a moat but a position. Once we take up the right position, there is no need to keep wrestling with AI.¹²¹¹⁴

12Amplification, compression, redistribution: what has AI actually changed?#

Put the experience of the disciplines together and the first keyword is amplification: AI has amplified output, and amplified junk along with it. A study published in Nature in January 2026 analysed 41.3 million papers in the natural sciences and found that scientists who use AI publish 3.02 times as many papers as others, receive 4.84 times as many citations and become project leaders 1.37 years earlier. At the level of science as a whole, however, the adoption of AI has shrunk the overall range of research topics by 4.63% and reduced interaction between scientists by 22%.¹²³ Another study, published in Science, found that scientists who use LLMs to write increased their output by between 23.7% and 89.3%, but that this inverted the relationship between the complexity of the writing and the quality of the paper, producing a flood of manuscripts that are linguistically complex but substantively mediocre;¹²⁴ some economists have since questioned its identification strategy, arguing that part of the effect may be a statistical artefact.¹²⁵ Even the evidence about AI’s productivity needs verifying. A 2024 preprint by an MIT doctoral student, claiming that AI had raised materials discovery at a large company’s R&D lab by 44%, was widely cited for a while; in May 2025 MIT announced that it had “no confidence” in the provenance, reliability or validity of the data, and asked for the paper to be withdrawn.¹²⁶¹²⁷ The junk side has already been itemized above.

The second keyword is compression – and the compression is asymmetric. Weather forecasts that took hours on a supercomputer now take less than a minute on a single accelerator; years of exploration by the Imperial College team were replayed by the AI co-scientist in about two days; OpenAI claims to have taken a Millennium Problem in 88 hours; Hall’s replication took under an hour. What has been compressed is mainly the stage of generation. Verification, explanation and digestion have not sped up in step: A-Lab’s “new materials” ended in a correction only two years later, and OpenAI’s proof is still in the Clay Institute’s “deliberately unhurried” review. Bloom says there are now plenty of papers of a hundred or two hundred pages whose authors had AI generate the proof, check the proof and write the paper – “But no human has read it, and no human is going to read it.”²⁴ The “proof indigestion” Tao describes is not confined to mathematics.

The third keyword is redistribution, and it runs in both directions. On the one hand, the barriers really have come down. A young man without a PhD used a chatbot to solve a problem that had been open for decades; 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 – and there, as Quanta wrote, it makes no difference whether you have tenure or are at a university at all;²⁴ the editing of classical Chinese texts has drawn in thousands of amateur proofreaders. On the other hand, concentration is growing. Consider an assault that deploys ten thousand agents at a cost of millions of dollars: as Javier Gómez-Serrano of Brown University put it, “very few mathematicians will have resources of that scale”;²⁶ the open letter against the Mathathon likewise pointed out that there is no way every mathematician could be given $20,000 in AI credits to prove a result.³² Before we talk about “concentration”, though, let us not picture traditional academia as an equal, pure, low-barrier world. Doesn’t traditional academia involve working the system too? Academic status buys resources, students and assistants; a few people own expensive laboratories and databases; conferences are arenas for networking; and the leaders of a field can steer its fashions. By comparison, AI’s barriers are not significant. It will not bring fairness automatically, but it weakens more of the old monopolies – laboratories, academic clans, student labour – than the new concentrations it creates, and real competition will return to an eye for problems, judgement, taste and responsibility.⁷⁵ A Nature comment at the end of September argued that AI is steering research towards the directions richest in data and easiest to evaluate, and called on funders to reward “the creation of new scientific terrain, not just the efficient mining of old ground”.¹²⁸ This is of a piece with Messeri and Crockett’s warning that we may “produce more but understand less”.¹¹ The narrowing of focus is what concerns me most, because homogenization is AI’s real problem. But we should be careful here too: homogenization on the surface is not necessarily pure loss. Writing flattened the diversity of speech, yet whole libraries grew up within a single written language; Go looks more monotonous than Jungle chess, but once you are inside it its variations are endless. Diversity on a new platform will not burst out from within of its own accord; someone has to go and explore. That is the vocation of scholars, and it is what funders should reward.¹³

The bottleneck, then, is moving. Writing is no longer scarce; four other things are. Verification: is it true? Formal proofs can answer part of that, but wet-lab experiments and clinical trials cannot be accelerated to the same degree. Explanation: can anyone make it clear? Digestion: does the community have the time and the people to absorb it? Evaluation: does it matter, and who decides? In his keynote at ICML 2026, the Princeton computer scientist Arvind Narayanan called the temptation to automate peer review a trap: “If we automate peer review, we essentially give up control over the direction of progress to AI systems themselves.” Human understanding, in his view, “is not some friction to be automated away”; it is the very reason we do science at all.¹²⁹ AI has not made science stop needing people. It has pushed people towards the harder and less rewarded positions in the chain: refereeing, checking, explaining, writing textbooks. Terence Tao points out that refereeing and editing are precisely the mechanism by which individual achievements are turned into collective progress – yet this is labour that people take on voluntarily and that goes largely unrecognized.¹⁷ This is where the limits of patching lie: under the old way of keeping score, no number of patches can reward digestion.

13Once the bottleneck has moved, what should we do?#

If the bottleneck has moved, what science needs is not just better AI but new technologies of witnessing and new forms of scholarly work. The historians of science Steven Shapin and Simon Schaffer showed that experimental matters of fact in the seventeenth century were produced jointly by three “technologies” – material, literary and social – and that detailed experimental reports allowed readers who were not present to take part in “virtual witnessing”.¹³⁰ The paper is a direct descendant of that technology of witnessing. Today new means of witnessing are cropping up here and there. Lean formalization lets anyone check a proof by machine; Kosmos tags every statement in its reports with the code or primary literature on which it rests;⁴⁷ NeurIPS asks suspect submissions for an audit trail of the writing process (the right direction, although the aim should not be to prove that “a human wrote it” but to make the process traceable); and Terence Tao argues that if authors cannot give a clear, correct, properly attributed expert talk on their result, it should not be published; he also recommends Mathematical Discourse, a peer-reviewed video journal of research talks.¹⁷ Assemble these parts and you come close to what the manifesto calls a knowledge object: argument, data, code, process records, exposition and machine interfaces brought together in a single work that is also friendly to machines.

The public and amateurs need new roles too. The Vesuvius Challenge, the Erdős problems website, the crowdsourced editing of classical texts, and Zooniverse – which in June 2026 passed a billion volunteer classifications from more than three million participants¹³¹ – all show that when problems are broken down, data made public and standards written out, amateurs can make real contributions. This is the revival of citizen science carried forward into the age of AI. But Bloom’s worries and the Mathathon open letter remind us that what can be lowered is the threshold for taking part, not the standard of verification. Amateurs need not only models but clear rules, visible review, and experts willing to vouch for their results. In the age of AI, the return of the audience depends on the audience being able to follow the game.

We also need to design for fun. Vakil says the Jacobian counterexample has made mathematics “more fun”; Yun Zhiwei says AI is destroying his journey. Both feelings are real. Burnett’s conclusion from the classroom: “You can no longer make students do the reading or the writing. So what’s left? Only this: give them work they want to do. And help them want to do it.”⁹⁵ Research is no different. If AI takes over generation, the work left to people has to be redesigned as work that is worth doing and that earns recognition: checking can be like solving a puzzle, explaining like giving a performance, digesting like exploring and collecting in a game. That is exactly the question designing for fun sets out to answer, and the direction RenaSci wants to try out with layered works, public rounds and records of achievement. In the end this is not a question of productivity. Production matters because it enriches life; once AI takes over production, people should return to living, leisure and taste, and incentives should shift from distribution according to labour to distribution according to delight – encouraging not more labour, but everyone’s cultivation of taste, sensibility and imagination.⁸⁶¹²¹

In the ethical treatise read in full from a charred Herculaneum scroll in June 2026 – very probably, researchers think, a Stoic work – 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…”⁸⁸ These words, more than two thousand years old and brought back to light only with the help of machines, are a fitting bequest to science in the age of AI. Machines can inquire on our behalf. They can generate, search and calculate for us, even write us a respectable paper, and perhaps one day they will verify and explain for us too. But grasping – judging what is worth inquiring into, what counts as having been made clear, what is good – can only be our own business. Not because machines cannot do it, but because only we care. That takes taste, and it takes a community willing to give its time and leisure to it.