The Renaissance of Science

1Science never stops progressing – so how can it have a renaissance?#

We call this undertaking RenaScience: “Renaissance” and “Science” fused into one word, meaning the renaissance of science. The name invites an obvious question. The arts can be reborn: statues and poems from two thousand years ago are still models today, and later generations can go back to them and learn afresh. Science seems different. It progresses: any secondary-school student now understands the motions of the heavens better than Ptolemy did, and nobody wants to go back to Aristotle’s physics. So how can science have a renaissance?

We are certainly not proposing to turn scientific knowledge back to antiquity, or to bring back the heliocentrism of the ancient Greeks. Interestingly, though, modern science itself began by borrowing the authority of the ancients. In 1543, in the preface to On the Revolutions of the Heavenly Spheres addressed to Pope Paul III, Copernicus wrote that he had first read in Cicero that Nicetas (that is, Hicetas of Syracuse) believed the Earth moved, and had later found in Plutarch that Philolaus and others held the same view.¹ In his manuscript he had also mentioned Aristarchus, the first to argue that the Sun stands at the centre, in the same breath as Philolaus, but he struck the passage out before the book went to press.² Historians have not settled whether Aristarchus really inspired him;³ what is certain is that he feared his peers would mock a doctrine so new, and so took pains to stress that the idea of a moving Earth was ancient, not his own invention.⁴ Many of the Renaissance’s creations were like this: done in the name of the ancients, yet never done before.

Nor has learning been reborn only once. In antiquity, scholarship would often flower in a golden age when a hundred schools of thought contended, and in its final phase be gathered up into magnificent compendia. Then the lively, varied but imperfect doctrines were gradually lost, the great compendia hardened into dogma, and everything had to wait for the next age of wonders to revive it. The Romans, the Arabs and twelfth-century Europe all had revivals, some long, some short. The Renaissance we all know became the last of them – after it, learning never declined again – because it no longer depended on the chance patronage of a king. It had a new social setting and new media equipment: the rise of civil society, and the movable-type printing press.⁵ Printing meant that books were no longer easily lost; later generations could always stand on the shoulders of earlier ones, and continuous progress became the norm (see manuscripts and books).

Today AI has brought another upheaval in media. It can survey almost every book humanity has produced – and it may also make people stop reading. It can break down the walls between specialisms and carry human knowledge into a new golden age – and after that golden age it may lead us into mediocrity. How can we make this upheaval open an era of continuous progress, as the Renaissance did, rather than fall into decline after its peak, like the scholarship of the Ptolemaic kingdom?⁵ Our answer: what needs reviving is not any particular body of knowledge, but a spirit.

2What exactly are we trying to revive?#

At the opening of the Metaphysics, Aristotle says that all human beings by nature desire to know; that people begin to philosophize out of wonder, moving from perplexity to understanding; and that mathematics first arose in Egypt because the priests there had leisure. He adds that we pursue this knowledge for no other advantage: just as a free man is one who lives for his own sake and not for another’s, so this is the only free science, the only one that exists for its own sake.⁶ The Greek word for leisure, scholē, later gave us the word “school”. Wonder and leisure, together with the free debate peculiar to the Greeks, happen to be the very hallmarks of play: recreation and contest driven by intrinsic delight (see wonder and leisure).

This delight was not high-minded. Debate, athletic contest and war were all ways for the Greeks to compete for honour; there were fame-hungry Sophists before there were philosophers in the sense of Socrates and Plato (see contests and competitions in knowledge). When people talk about “the scientific spirit”, they like to list indifference to fame and gain, a willingness to endure obscurity, even self-sacrifice. I don’t like a single one of these words. Look at the real history of science: the great figures mostly had wide circles of acquaintance, went with the tide and sought recognition within their community. Rather than lone warriors, they were people who chased fame and fortune, could not bear obscurity and knew how to keep in with every side.⁴ Curiosity and the will to win have always come as a pair, and beneath both lies the same ground: seeking knowledge is an end in itself, a delight in itself.

The scientific spirit did not appear in an instant. Like a person’s character, it has a history of growth. Ancient Greece laid down science’s first impulse: an inward way of seeking knowledge, driven by the desire to know and by curiosity, together with a tradition of free, disinterested inquiry and criticism. The Middle Ages learned to value experience and to draw careful distinctions between terms. The Age of Discovery brought the idea of exploring the unknown and finding new knowledge – the scholars of antiquity did not think of themselves as making “discoveries” that went beyond their predecessors. Printing made learning public. The Scientific Revolution established experiment and quantification. The Enlightenment believed in progress and in science’s power to benefit humanity. The Industrial Revolution allied science with technology. The twentieth century added ethical reflection and social responsibility.⁷ None of these later-grown traits needs to be thrown away. Staying true to a first impulse does not mean staying frozen in it; someone who has learned to keep their feet on the ground need not lose the urge to look up at the stars. The renaissance we mean is for science, now grown up, to find its childlike heart again.

3How did science become a job?#

Science has not always kept this spirit. The first alienation it has suffered is specialization and professionalization.

The English word “scientist” was proposed only in 1833, at a meeting of the British Association for the Advancement of Science; the following year William Whewell used it in a book review, and it appeared in print for the first time.⁸ Before then, those who studied nature were mostly aristocrats, clergymen, physicians and amateurs from every walk of life. Afterwards, the German-style research university, the laboratory and the doctorate turned science into a profession with a way in, a ladder and a salary (see how being a scientist became a job). In 1917, lecturing to students in Munich on Science as a Vocation, Max Weber warned the young people in his audience that whether a Privatdozent – an unsalaried lecturer – ever rose to a full professorship was “simply a gamble”, and that academic life was “a wild gamble”.⁹

Behind professionalization lies the logic of industrial society as a whole. Technology has always tended towards specialization, and after the Industrial Revolution the division into specialisms and the assembly line became the underlying logic of the entire technological system: a band of all-round craftsmen working from their own interest is no match for a crew of dull, regimented workers, each specialized in one particular motion.¹⁰ Industrial production is at heart about replication, and it demands that people, too, be replicable and interchangeable, like parts; so education policy and the education system came to aim at mass-producing “cogs”.¹¹ Science was drawn into the same logic. In the 1930s Ernest Lawrence used the cyclotron at Berkeley to launch “Big Science”. The machines were so complex that neither the best theorist nor the ablest experimenter could master them alone, and keeping them running day to day took a great deal of tedious labour: the “research labourers” of today – in Chinese, literally “research migrant workers”, the people who do a lab’s routine grind – had already appeared.¹² During the Second World War, Oak Ridge used devices Lawrence had invented to separate uranium isotopes, and recruited some ten thousand young women from rural Tennessee to watch the dials. According to Kenneth Nichols, who was in charge of the project, the women beat Lawrence’s PhDs in a production race that Nichols had proposed, because they were “trained like soldiers ‘to do or not to do—not to reason why’”, while the scientists could not help stopping to investigate even the slightest fluctuation in the readings.¹³ An assembly line has no need of curiosity; on an assembly line, curiosity is even a nuisance.

The accumulation of knowledge made specialization worse still. In 1963 Derek de Solla Price estimated that eight or nine in ten of all the scientists who had ever lived were alive at that moment.¹⁴ The more knowledge there is, the heavier the pack a person must shoulder before reaching the frontier. The economist Benjamin Jones analysed more than two million US patents and found that inventor teams were growing larger, that people were older when they made their first invention, and that inventors were becoming more specialized; he called his paper “The Burden of Knowledge and the ‘Death of the Renaissance Man’”.¹⁵ The “all-sided man” of the Renaissance described by Jacob Burckhardt is ever rarer in modern science (see why more knowledge makes people narrower).

Specialization has brought astonishing efficiency, but it has also created disciplinary islands and expert blind spots. As early as 1930 José Ortega y Gasset was attacking “the barbarism of specialization”;¹⁶ in 1959, in a lecture at Cambridge, C. P. Snow lamented that science and the humanities had become “two cultures”.¹⁷ But what Snow described was the mutual disdain of literary intellectuals and scientists, a quarrel within the elite over who gets to speak, and I said long ago that this debate is out of date.¹⁸ Today the real rift runs between elite culture as a whole and the general public. A paper assumes its readers are peers, so peers can read it fast; but beginners, readers from other disciplines and the public find it hard to get in. It is a narrow-necked bottle (see the triumph and cost of specialization).

More important still is the change in motive. Once research became a profession, it also became a way for many people to earn a living. Learning for pleasure is something natural and deeply human; learning for marks, grade-point averages, certificates and jobs is its alienated form.¹⁰ When research can be exchanged for a salary and a professional title, many people who do not love knowledge will also walk into the laboratory for the sake of a steady income. You cannot blame them: an assembly line never needed love, only compliance (see why people who don’t love knowledge do research anyway).

4Why does knowledge need a body?#

Knowledge never exists naked. It must lodge in some medium: spoken words, characters written on bamboo slips and parchment, chapters printed in books, letters sent to distant places, a paper in a journal, a diagram, a piece of code. The medium is not a transparent pipe. It determines what can be preserved, passed on and put to the test – and therefore what will count as knowledge.

In the Phaedrus, Plato has Socrates worry about writing: it will make people look learned without real understanding, and it cannot answer a reader’s questions.¹⁹ Seen from more than two thousand years later, the worry was both right and wrong. Writing did change knowledge – not by making it less, but by giving it a different shape. Speech excels at dialogue and rebuttal: the exchanges between Confucius and his disciples were written down as recorded sayings, and medieval universities trained scholars through lectures and disputations. Books excel at building systems: Copernicus, Newton and Darwin each changed the picture of the world with a whole book. Letters excel at weaving networks: in the seventeenth century, scholars scattered across Europe kept one another informed by correspondence, and even staked priority claims with anagrams (see the Republic of Letters). But letters were private in the end; it was printed books and journals that turned this private web into a public “republic of the literati”. The journal article excels at something else: compressing a piece of research into a unit that can be registered, circulated, cited and counted.

The Chinese tradition offers yet another body. The Nine Chapters on the Mathematical Art set mathematics out as problem after problem, each with its “procedure”; in the Wei–Jin period Liu Hui wrote a commentary on it, supplying reasoning and diagrams to prove why the algorithms were correct. Knowledge grew as problems, procedures and commentaries accumulated, layer upon layer.²⁰ This reminds us that the same learning can take very different bodies, and that once a body is chosen, it dictates how later comers enter the subject and how they add to it. And some knowledge is not in writing at all. “We can know more than we can tell,” as Michael Polanyi put it.²¹ The craft skills of the laboratory, the trained eye that reads a scan, the intuition that tells whether a proof “feels right” – all of these pass from master to apprentice and between colleagues. Any medium can carry only part of knowledge.

Every medium has things it does well and things it cannot see. Conversation cannot be kept; books take too long to write; letters are private. The paper’s blind spot is that it keeps only the path that worked. Put these media side by side and it becomes clear that the paper is not the natural shape of knowledge, but a body that knowledge took on for a while in one particular media environment. In The Strong Programme in Media History I argued that media are not tools for transmitting knowledge but its a priori conditions: the possibility of “recalling” depends on the possibility of “recording”.²² Media help to shape how we think and how we act, and also what is taken to be true. The media of science are no exception. When the environment changes, the body changes with it – not because we want a new one, but because the new environment no longer rewards the old body.

5How did the paper win for three hundred years?#

In 1665 two periodicals appeared almost at once. In January the Journal des sçavans was launched in Paris; in March Henry Oldenburg, secretary of the Royal Society, founded Philosophical Transactions in London.²³²⁴ At first they read like miscellanies of news, book reviews and extracts from letters, but they soon revealed an unprecedented power: a complete stranger could get hold of the same stable text and raise an objection on the same page. Journals took on four jobs: registering who discovered something first, spreading the discovery, giving it a preliminary certification and archiving it (see the birth of the journal). At around the same time, Robert Boyle wrote exhaustively detailed reports of his air-pump experiments, so that readers who had not been there could “virtually witness” what happened inside the glass receiver.²⁵ From the very beginning, the paper was a technology of witnessing.

In the nineteenth century, as science became a profession, the paper won its decisive victory. Specialist societies and specialized journals multiplied, and research universities began to measure people by what they published. The Royal Society’s Catalogue of Scientific Papers turned one scholarly life after another into a list of papers.²⁶ The paper became the scientist’s standard output and the basic unit of academic credit (see the paper and the professionalization of the scientist). The genre itself was standardized in its wake. The four-part structure of Introduction, Methods, Results and Discussion gradually spread through medical journals in the mid-twentieth century, and by the 1980s had become the standard format for original research papers;²⁷ the methods section lets others repeat the work, and the impersonal voice makes a paper read as though nature itself were speaking. Nor did scientists ever care any less about who got there first: as early as 1957 Robert K. Merton pointed out that priority disputes run through the whole history of science,²⁸ and the journal happened to provide a registry with a date stamp.

In the twentieth century this unit was fitted with meters. In 1963 Price observed that the scientific literature was doubling roughly every ten to fifteen years;¹⁴ in 1955 Eugene Garfield proposed the citation index, and the impact factor followed later.²⁹ Peer review, by contrast, arrived much later: only in the Cold War did it become a symbol of scientific legitimacy,³⁰ and it was not until the 1970s that Nature systematically sent submissions out to referees.³¹ In 1936, when Physical Review sent one of Einstein’s papers to an anonymous referee, he was so indignant that he withdrew it and submitted it elsewhere.³² Peer review, now regarded as a cornerstone of science, is in fact a rather young institution.

The paper won not because it came closest to the truth, nor because printing destined it to win, but because it best suited the whole environment of its time. It could be printed and posted; it could register priority; it let strangers check and rebut one another on the same page. A paper is not a monologue but a conversation: every citation says to one’s peers, come and check me, come and refute me. Later it could also be cited and counted, and written into CVs and grant applications. It is modular, standardized and easy to administer. That modern science could divide its labour, collaborate across the globe and grow cumulatively owes a great deal to the paper. And precisely because it has been so successful, we easily mistake it for knowledge itself.

6The web arrived – so why is the paper still paper?#

The internet had the chance to change all this. In 1989, at CERN, Tim Berners-Lee wrote a proposal to use hypertext to link physicists’ scattered documents into a web.³³ The World Wide Web was born in a laboratory; it was invented for scientists in the first place (see from the Memex to the World Wide Web). Yet more than thirty years later, the main form a paper takes online is still the PDF: a digital page that imitates paper, with page numbers and columns but almost no links and no layers. The data and code lie somewhere else, and the failed paths are still nowhere to be seen. Digitization merely swapped one sheet of paper for another.

The information age also made counting faster and finer-grained. In 2003 Shanghai Jiao Tong University published the first Academic Ranking of World Universities;³⁴ in 2005 the physicist Jorge Hirsch proposed the h-index, which sums up a person’s scholarly output in a single number;³⁵ databases now tally every paper’s citations in real time. Publishing, meanwhile, became more concentrated: one study found that in 2013 the five largest publishers produced more than half of all the world’s papers.³⁶ Subscription fees and article-processing charges rise year after year, and university libraries pay to buy back the papers their own researchers wrote while also paying those researchers’ publication fees (see scale and concentration: the publishing industry).

The information age brought real progress, too. arXiv, born in 1991, let physicists share papers before review; the open-access movement let more people read research free of charge; during the Covid-19 pandemic, preprints let research circulate at unprecedented speed (see preprints and open access). The evaluation system has also been correcting itself. The 2012 San Francisco Declaration on Research Assessment called for journal impact factors no longer to be used to evaluate individual researchers;³⁷ the very first principle of the 2015 Leiden Manifesto for research metrics states that quantitative evaluation should support experts’ qualitative judgement, not replace it.³⁸ Registered reports, open data, open review: one patch after another has been applied.

But every patch is built around the same unit: the paper. The patches make the paper faster, cheaper and more open, but they do not make it grow new organs. Twenty years ago people were already saying that print journals ought to be obsolete; twenty years on, scholars search and read online, yet the public forum for scholarly exchange is still the print journal – and the humanities have, if anything, gone backwards. This is not loyalty to scholarly tradition but a betrayal of it. The form of the journal article was itself born of the postal age, as a means of communication; to cling to the form when the environment has changed is to preserve nothing but the form.³⁹ Digitization changed how papers circulate; it did not change the body of the paper (see the limits of patching).

7Why did publishing become a matter of life and death?#

The second alienation science faces is an evaluation regime centred on publishing papers, together with a university job system that comes close to a “Ponzi scheme”.

When a single medium has to perform every function at once – communication, certification, archiving, evaluation and allocation – it is bound to be squeezed out of shape. Papers can be counted, so academia distributes according to labour: the more you write, the more you get. And if the more you write, the more you get, then everyone writes more and everyone is caught in the rat race – each step the result of rational calculation. The phrase “publish or perish” was already going round American academia in the first half of the twentieth century; later it became the law of survival for scholars all over the world (see where the pressure to publish comes from). In 1981 William Broad, a reporter for Science, wrote that in some circles the “least publishable unit” was a euphemism for fragmenting data: a researcher publishes four short papers rather than one long one.⁴⁰ In 2018 John Ioannidis and his colleagues searched a literature database and found 265 hyperprolific authors who, in at least one year, put their names to a paper every five days on average; between 2001 and 2014 the number of such authors grew about twentyfold. Of the 27 who completed a questionnaire, 19 admitted that in more than a quarter of their papers they had not met at least one of the criteria for authorship.⁴¹

The costs come in four forms. The first is a “city of survivors”. To save space and keep the story tidy, a paper keeps only the path that worked; failed experiments, null results and abandoned hypotheses are cut out and buried underground. An analysis of more than two thousand papers found that 91.5% of papers in psychology and psychiatry reported “positive” results;⁴² a large collaborative project repeated a hundred psychology studies and obtained significant results in only about a third of them;⁴³ in a 2016 Nature survey, more than seven in ten researchers said they had at some point failed to reproduce someone else’s experiment.⁴⁴ Above ground, the world of the literature is ablaze with light; below lie vast ruins that no one has recorded (see the replication crisis and publication bias).

The second is the “tyranny of counting”. When numbers of papers, impact factors and citation counts are used to allocate jobs, funding and honours, the metrics begin to devour the goals. In 2023 more than 10,000 research papers were retracted worldwide, a record, and more than 8,000 of them came from journals belonging to a single publisher that had been infiltrated by paper mills.⁴⁵ Since 2018 China has been pushing to “break the Five Onlys” – the exclusive reliance on papers, talent “hats” (honorific titles), professional titles, diplomas and awards – and in 2020 it issued measures aimed specifically at ending the “papers only” orientation in the evaluation of science and technology. These are corrections to exactly this tyranny.⁴⁶ Counting itself is not wrong; what is wrong is letting counting take the place of judgement (see metrics and “papers only”).

The third is the pyramid. The whole counting game rests on a pyramid propped up by newcomers. Richard Larson and colleagues estimated that an engineering professor in the United States trains an average of 7.8 PhDs over a whole career, and that in a steady state only one of them can take over a faculty post; for half of all PhDs to stay in academia, the field would have to grow by 14% a year.⁴⁷ In 2010 The Economist wrote that “the fiercest critics compare research doctorates to Ponzi or pyramid schemes”, and that universities “have discovered that PhD students are cheap, highly motivated and disposable labour”.⁴⁸ In autumn 2023 about 68% of faculty at American colleges and universities were off the tenure track; in 1987 the figure was about 47%.⁴⁹ China’s doctoral graduates rose from 55,000 in 2016 to 112,900 in 2025,⁵⁰ while births over the same period fell from 18.83 million to 7.92 million.⁵¹ As I have said, even without the AI revolution the Ponzi structure of humanities education was close to collapse. If the best future for the students you train is to become teachers themselves, you need an ever-growing number of posts to accommodate your students and your students’ students. In the past there was population growth and university expansion; now the population is shrinking and university participation has peaked, so in future there will be fewer and fewer chairs³⁹ (see why academic careers look like a Ponzi scheme).

The fourth cost is the one least talked about: the draining away of delight. A study of Australian researchers estimated that a single round of project-grant applications to the National Health and Medical Research Council consumed about 550 years of working time;⁵² graduate students suffer anxiety and depression at more than six times the rate of the general population.⁵³ Science began in wonder and leisure, in the pleasure of puzzles and contests; today it looks more and more like a piecework assembly line (see the fun that got squeezed out).

These costs are not the result of a few people’s moral failings; they are structural. More than a century ago, Weber put a question to every young person who hoped for an academic career: do you really believe you can bear it, year after year, as one mediocrity after another climbs past you, without growing embittered and being corrupted inside?⁹ Today there is one more thing to bear: year after year, slicing the passion for knowledge into papers that can be counted.

8After the knowledge explosion, why are researchers more lost than ever?#

The third alienation comes from the information age, and from the explosion of knowledge that AI has brought.

Databases and the web have made searching easier than ever, but reading has not got any easier: the papers added in every field each year far exceed what one person could read in a lifetime. Generative AI has pushed the flood to new heights (see the influx of generative AI). One study estimates that at least 13.5% of biomedical abstracts in 2024 were processed with large language models;⁵⁴ NeurIPS 2025 received 21,575 valid submissions;⁵⁵ in October 2025 arXiv announced that review articles and position papers in its computer-science category would be accepted only after passing peer review;⁵⁶ a detection company estimated that about 21% of the reviews for ICLR 2026 were entirely AI-generated.⁵⁷ In February 2026 a Nature news story told of a researcher at the University of Regensburg in Germany who used a newly launched writing tool to write up a paper describing a fictitious experiment in fifty-four seconds.⁵⁸ These figures should themselves be taken with a pinch of salt: “AI rates” – AI-detection scores – simply cannot be measured accurately, and checking them is institutional laziness. What the figures really show is that the old system was designed on the assumption that writing things up is hard, and that assumption no longer holds.

Mathematics has changed most dramatically of all (see mathematics). In April 2026 a 23-year-old with no professional training in mathematics handed an Erdős problem to a chatbot. The raw output was “actually quite poor”, yet it pointed to a route nobody had taken before.⁵⁹ By August, 565 problems on the Erdős problems website were marked as solved, with AI involved in more and more of them. Yet the site’s founder, Thomas Bloom, worried that more and more papers a hundred or two hundred pages long were having their proofs generated by AI, checked by AI and written up in full by AI – and of such a paper, “no human has read it, and no human is going to read it”.⁶⁰ In June the Leiden Declaration on Artificial Intelligence and Mathematics was published with the endorsement of the International Mathematical Union. It calls for the use of AI tools to be disclosed and for authorship and responsibility to be left to humans, and more than four thousand people have signed it so far.⁶¹ In July, in a public lecture at the International Congress of Mathematicians, Terence Tao spoke of a crisis in the values and practice of mathematics. Soon afterwards, in a long essay, he broke “solving a problem” down into stages – generation, verification, exposition, publication and digestion, and finally entry into the textbooks – and pointed out that AI generates proofs far faster than humans can verify and digest them, a condition he called “proof indigestion”. “A proof that no human can properly explain,” he wrote, “should be viewed as incomplete, even if it has been formally verified.”⁶² In September OpenAI claimed that its agentic system had solved the Navier–Stokes problem, one of the Millennium Prize Problems; the Clay Mathematics Institute said that the problem had “apparently been settled”, but that its review would be “deliberately unhurried”.⁶³⁶⁴ More than twenty Fields medallists signed a statement warning that the goals of AI companies and those of the mathematical community are “severely misaligned”;⁶⁵ more than seven hundred people signed an open letter opposing a forty-hour “Mathathon” at Caltech, fearing it would mass-produce “slop mathematics”.⁶⁶ At the end of the month the original organizers, two of the open letter’s authors and others issued a joint statement: the event would instead look for new proofs of existing theorems and write expositions of them, and would no longer accept sponsorship from developers of proprietary AI models.⁶⁷ Deng Yu, who had just won a Fields Medal, put it plainly: AI can help people check results, find literature and verify details, but what really drives mathematics forward is still human intuition and ideas.⁶⁸

There is also a quieter finding. A study covering more than forty million papers showed that scientists who use AI publish about three times as many papers as others, receive nearly five times as many citations and become project leaders earlier; yet science as a whole has narrowed its focus, and researchers interact with one another less.⁶⁹ Individuals are expanding; the collective is contracting.

In the flood, researchers grow ever more lost. Specialization leaves each person guarding a small patch without seeing where it sits in the whole; the flood means one cannot even finish reading one’s own patch; and AI makes it hard to say which work is still worth doing by humans. Under the logic of the assembly line, experts need no general education: they need only know “how”, not “why”.¹⁰ Yet the yardstick for what is precious and what is rubbish must ultimately rest on a human measure. If people give up cultivating their own taste and sense of what is interesting, AI can only sift things for us according to the average palate of the past.⁷⁰

9Why is the crisis also a turning point?#

But the third crisis is also a turning point. Information technology and AI are creating the flood, but they are also washing away the walls built by specialization and paper supremacy.

Take specialization first. Industrial-age education and research set out to mass-produce specialists because industrial production is at heart about replication, and people, too, had to be replicable and interchangeable, like parts. But copying information costs almost nothing: build one AI and it can be copied for everyone to use. Once replication is no longer society’s main task, education and research have a chance to stop aiming to turn out replaceable consumables.¹¹ More directly, any skill that large numbers of people can be drilled into by rote, step by step, AI will soon learn too – and learn better.¹⁰ The burden of knowledge can be shared with machines. With AI, one person can get the lie of the land in an unfamiliar field in a few days, check the derivations and get the code running; the literature of the discipline next door is no longer a closed book, and interdisciplinary researchers can, for the first time, go deep enough into other people’s territory. I am rather optimistic about AI: I believe it will overturn the whole system of scholarship and education and break down the walls between specialisms.⁵

Then the paper. Writing costs almost nothing now, so being able to write something up no longer vouches for anything, and the paradigm of paper supremacy has begun to loosen. In 2025 researchers at Stanford held Agents4Science, the first conference at which AI served as first author and as reviewer: 315 submissions, 48 accepted. Research led entirely by AI made up 23.3% of submissions but only 14.9% of accepted papers.⁷¹ Humans were clearly more involved in the papers that were accepted, and where they contributed most was precisely in the early stages – proposing hypotheses and designing the research. The bottleneck has moved from writing things up to verifying, explaining, digesting and evaluating (see amplification, compression and redistribution) – and not one of these four is something the paper, as a genre, is good at carrying. It is good at submitting a conclusion; it is not good at showing the process, sharing the work of verification or explaining things to different readers, and it cannot show at all who asked the question, which detours were taken, or what the machine did at which step.

Different disciplines have responded differently (see AI Enters the Lab). The life sciences reaped the first harvest: half of the 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for predicting protein structures with AI;⁷² in 2025 a hypothesis about a mechanism of bacterial gene transfer, put forward by Google’s “AI co-scientist”, matched a laboratory’s as yet unpublished experimental results.⁷³ The social sciences were the first to sense danger. Researchers used large language models to simulate survey respondents; the resulting “silicon samples” looked much like real people on average, but flattened the differences between individuals.⁷⁴ Another study found that AI bots could pass almost every attention check and pose as humans filling in online questionnaires.⁷⁵ Humanities scholars have witnessed the miracle of machines reading ancient scrolls, and they are also the most worried that the basic skills of reading, writing and interpretation will be outsourced (see the humanities). That worry is not new: writing once outsourced memory, and the cost was real, but nobody went back to oral tradition because of it. What should worry us is that humanities scholars have been the slowest to respond to this wave.

Relying on AI was never the problem, because an “environment without AI” no longer exists; the problem is homogenization. Ideally, the arrival of AI will turn many professions back into “liberal arts”: pursued not to earn a living but for the sake of knowledge and taste. Reality could also turn out worse: once the skills people made a living by are obsolete, many may not want to learn anything at all, and the professions will become elite arts again, practised by a tiny few.¹⁰ The turning point will not arrive of its own accord; it has to be designed (see whether science can go from profession back to passion).

10Can everyone be a scientist?#

The word “amateur” comes from the Latin for “lover”. Many major discoveries in the history of science were made by amateurs. William Herschel started out as a musician; in 1781 he discovered Uranus with a telescope whose mirror he had ground himself.⁷⁶ Darwin never held a university post, and through his letters he mobilized a network of pigeon fanciers, gardeners and plant enthusiasts (see the age of the amateur). In the early days of the Scientific Revolution, any amateur outside the institutions might make a major discovery; once modern science had matured, untrained “folk scientists” – the Chinese pejorative for cranks – had almost nowhere left to stand.⁷⁷

In recent years the door has opened again. Players of the game Foldit solved, within three weeks, the structure of a viral protease that had defeated researchers for more than a decade;⁷⁸ the citizen-science platform Zooniverse reached its billionth classification in June 2026, with more than three million volunteers;⁷⁹ the Vesuvius Challenge offered prizes open to anyone, and contestants used machine learning to read scrolls that the volcano had buried for nearly two thousand years – in 2026 a research team drawn mostly from the contestants read an entire scroll for the first time (see the revival of citizen science).⁸⁰ In the comment threads of the Erdős problems website, it makes no difference whether you have tenure, or whether you work at a university at all.⁶⁰ AI can also keep a layperson company through the literature of an unfamiliar field and check the derivations with them. For anyone who is truly curious, it is the best private tutor there has ever been.

“Everyone can be a scientist” needs two clarifications. First, it does not mean that everyone has to be a scientist. Ordinary people have no duty to advance science, and it is perfectly normal that most of them simply want to live their own lives well.¹⁰ Second, it does not mean that every idea counts. The style of science has always been to join the community first and break with orthodoxy afterwards. Those in history who put forward revolutionary theories could usually already speak the language their peers accepted, and could read and make good use of earlier work; if you want your peers to accept your theory, you must first enter this shared, mutually compatible forum.⁴ What is wrong with cranks is not that they are amateurs, but that they refuse to learn this common language. AI happens to be able to help a layperson learn it faster; the platform’s job is to hold the line on checking and judging (see dividing the roles without lowering the standards).

Within these two limits, science can turn from a rigid profession back into a lively passion. The telegraph put messengers out of work, yet the marathon became a sport; humans have long been unable to beat AI at chess, yet more people than ever play, commentate and watch the game (see whether AI will take the fun away). When the pursuit of knowledge is no longer just a way to make a living – when some make it their profession and many more make it their pleasure – ordinary people with a genuine passion for knowledge, and researchers who cross disciplines, will have the chance to return to the frontier of research. That, we believe, is the renaissance of science.

11If not in papers, where should knowledge live?#

We propose upgrading the unit of scholarly output from “a paper” to “a knowledge object”. A single piece of research should have seven layers at once: a core argument compatible with journals; data, code and figures that can be clicked open; a process record that keeps the failed paths and the differences between versions; a multimodal explanation made of videos, diagrams and interactive models; a governance record that discloses ethics, funding and the use of AI; a structured interface for machines to read; and a shared space where those who come later can criticize, replicate and fork the work (see the seven layers of a work). The seven layers open onto six doors: the public comes in through the video, beginners through the concept map, peers through the paper, replicators through the data and code, overseers through the governance record, and AI through the structured interface. The paper need not disappear; it returns to its proper place, as one of the entrances.

Imagine a concrete case. A psychologist wants to replicate a famous experiment. Under the old system, what she finally hands in is an 8,000-word paper stating that “this study failed to replicate the original effect”, and then she waits for some journal willing to publish the negative result. Under the new system, she hands in a knowledge object. The core argument can still be submitted to a journal; the preregistered protocol, the raw data and the analysis code are a click away; the process record keeps her reasons for revising the experimental design three times; a ten-minute video explains why the effect matters and why it is hard to reproduce; the governance record states the sources of funding, the ethical review and how AI was used in cleaning the data; a structured interface lets other researchers and AI assistants call on her data directly; and in the shared space, the original authors’ response and another team’s further replication hang on the same work. A paper has become a place you can enter, check and carry on building.

This is no novel demand. In 1665, the year Philosophical Transactions was founded, Robert Hooke published Micrographia, engraving a flea and the compound eye of a fly as large copperplates; the diarist Samuel Pepys sat up reading it until two in the morning and called it “the most ingenious book that ever I read in my life”.⁸¹ Science was multimodal from the start; it was only the cost of printing that made later papers more and more a matter of text. Today, multimodality is no longer expensive.

Going further, we propose organizing knowledge objects around questions. One page per topic, making clear what it asks; the follow-up questions on the page lead to other topics, and anyone who wants to go deeper can dive down to the deep read, the sources, the data and the process; and every topic can in turn become a new starting point. A work of scholarly curation is therefore not a paper but a question network. The idea has a long prehistory. In 1945 Vannevar Bush imagined a machine called the Memex that would let researchers leave “associative trails” through the literature that could be retraced;⁸² today mathematicians already break large proofs down into clickable dependency graphs that show at a glance which steps have been proved and which have been formally verified (see precedents). The paths readers take through the network are themselves a kind of understanding.

The new body must also cater for a new kind of reader: AI. More and more people first have an AI read the literature for them, and only then decide whether to read the original themselves. A knowledge object should give its machine readers a structure: which parts are claims, which are evidence and which are mere conjecture; where the data are; what the licence is; which version this is (see friendly to machines). Being friendly to machines is ultimately being friendly to people: only if the AI reads accurately will what it passes on to humans stay true.

The set of topics you are reading is a small example: the manifesto is the starting point, each of the seven further topics has its own curated page and deep read, the questions are tied to one another, and every part will go on being revised (see this set of topics as a case in point). Networks have their risks, too: readers get lost, content fragments, and maintenance is exhausting. So every question network needs a clear starting point, a map, and versions that can be frozen and cited (see labyrinths, fragments and attention).

12How can ordinary people get hands-on with research?#

Science has always had an audience. The anatomical theatre at Leiden, built at the end of the sixteenth century, opened in winter to townspeople who had bought tickets; in the nineteenth century, Michael Faraday’s lectures at the Royal Institution were packed (see when science was a public show). It was later specialization, and the regime of the paper, that pushed audiences and amateurs further and further away. AI means that, for the first time, ordinary people can get their hands on research and “play” it.

The research we have in mind works like a turn-based game. The researcher writes a direction brief setting out the question, the constraints, the no-go zones, the criteria for judging success and the budget. A team of AIs takes a turn – searching, reasoning, experimenting, plotting, hunting for errors, keeping records – and at the end of the turn hands back a battle report. The human then decides whether to carry on, switch to another branch, reload an earlier save point, or call a halt. Every save point is a version of the work, so the version history no longer has to be written up after the fact.

In research of this kind, humans have at least five positions. The Terminator decides when to call a halt; the Commander gives the AIs their heading and marks the no-go zones; the Warden keeps watch over ethics, funding and disclosure; the Dreamer throws out hunches for the AIs to test; and by far the largest group is the Audience (see dividing the roles). Spectators can watch a whole game of research, post suggestions as live comments, vote on branches and tip in resources to get new work started. Their voices go into the “atmosphere layer” of the battle report, and become instructions only if the researcher adopts them. Go did not die after AlphaGo: one study found that once superhuman AI had arrived, human players’ decisions became more novel, and better.⁸³ AI will not necessarily take the fun out of research.

The audience has its problems, of course, but the problem is not the noise. Spectators may not make a substantive contribution, yet they give researchers continuous, steady feedback. Tips are not a traffic metric either: they are distribution according to delight in its most direct form. In the livestreaming economy, the streamer who collects the most tips is never the hardest-working one, but the one who gives everybody the most fun. The real risks are the logic of traffic, and cheating: algorithms tailoring the audience’s tastes, AI accounts flooding the live comments, a research contest watched by crowds mass-producing conclusions nobody has checked – exactly what mathematicians feared about the “Mathathon”. That is why spectators must be identifiable and their voices go only into the “atmosphere layer”, never turning directly into instructions; and that is why every game of research must leave a complete battle report, and every conclusion must pass the three rules set out below. The buzz is a way in, not a standard.

Some will say this turns science into entertainment. But science was born of play in the first place: wonder, leisure and debate are all features of play, and Johan Huizinga found that at the sources of civilization, knowledge and play were one. Play is not opposed to seriousness; on the contrary, it is within play that the demand for seriousness is highest – as the old rule of the board has it, watch in silence and never take back a move. A good game makes every choice meaningful. Hand over to AI everything that machines can run, keep asking, conjecturing, choosing and judging for humans, and the delight of research may be preserved, even amplified. AI doesn’t care whether a problem is deep or shallow. You do (see designing for delight).

13The bar is lower – so where do standards come from?#

What has been lowered is the barrier to doing the work, not the standard. Nor does decentralization mean flattening: we need city walls, we need gates, and the gates need thresholds.⁸⁴ Science for all does not mean that anything counts as a result; on the contrary, when production becomes cheap, standards have to become clearer.

Seventeenth-century science established facts through witnessing: Boyle invited credible people to be present, and wrote detailed reports so that those who were absent could witness too. Twentieth-century peer review was another technology of witnessing. The age of AI needs new technologies of witnessing, and we sum them up in three rules: evidence you can click open; results someone can explain; AI in plain sight. The first means that every key datum, figure and step of reasoning can be traced to its source. The second means that every result must have a person who takes responsibility for it and can explain it in their own words – which is exactly what Tao’s standard means. The third means that every contribution AI makes – which model was used, what it did, how much computing power it consumed – is logged automatically by the platform in the process record. This is the platform’s contract, not the author’s confession: using AI is no stain, it needs no separate statement, and it need not be volunteered like an admission of wrongdoing. The line is drawn not at whether AI was used, but at whether anyone has lied or misled readers.

Rules are shifting everywhere, but not necessarily in the right direction. Major journals stipulate that AI cannot be listed as an author,⁸⁵ and on this we agree: authorship is first of all not an honour but a responsibility, and AI cannot carry the can. The US National Institutes of Health, for its part, prohibits the use of generative AI in peer review;⁸⁶ the Guidelines for Responsible Research Conduct (2023), issued by the Department of Supervision of China’s Ministry of Science and Technology, forbid using generative AI to produce application materials directly, or listing it as a co-contributor to research results (see the rules of journals, funders and societies).⁸⁷ Neither bans nor detection will work: can a ban hold? Can detection catch anything? What should be examined is the work itself, and whether its process can be traced. Papers produced with AI and papers written by people should be reviewed under the same rules, on their merits. Formal verification is one of the strongest forms of witnessing, but it answers whether something is right; it cannot answer whether it is understood, or whether it is worth anything (see trust in a network).

Nor can reviewing itself be handed over to machines wholesale. In a keynote in 2026 the computer scientist Arvind Narayanan warned: “If we automate peer review, we essentially give up control over the direction of progress to AI systems themselves.” Human understanding, he added, “is not some friction to be automated away”.⁸⁸ AI can do far more than help out: checking citations, recalculating data, finding similar work, spotting holes in an argument – it will do all this more carefully than most reviewers. But it can judge what is interesting only in the sense of a statistical average. The judgement of whether a piece of research is worth doing, and whether it has been explained clearly, can only be made by people who care about it and take responsibility for it.

Beyond standards, there is reputation. We propose that academic reputation should consist of achievements that can be accumulated, displayed and assessed by peers, not of paper counts. Everyone can have an academic calling card. It does not list papers and impact factors; it lists achievements: which problems you have solved, what role you played in which team battle, how many people have understood which of your works, which of your chains of evidence has survived replication. The long-undervalued labour of curating, explaining, editing and replicating should be recorded too (see the curator: a new scholarly role). Achievements must be tied to verification: if a result is retracted, the corresponding trophy is void. Otherwise the achievement system will simply become a new counting game and repeat the mistakes of the impact factor.

Models will be replaced generation after generation, and prompts will go out of date. What does not reset to zero with each new version number is what researchers have done in their community, the trust they have earned and the work they have left behind. An online game that has run for twenty years has never lived on having the best graphics engine, but on the characters, achievements, guilds and friendships its players have built up. That is the real moat of a research platform.

14What are we going to do?#

We are going to build three things, plus a calling card.

First, a scholarly curation platform that hosts seven-layer knowledge objects and everyone’s achievements: every work has versions, disclosures, reviews and an audience; invited curators publish first, then registration opens to all. Second, a new journal: each issue is a selection from one season, published in digital and print editions at the same time. The print journal draws readers in and guides them; the platform hosts, traces and reproduces. The magazine edition of our manifesto is its first specimen issue. Third, a research-game engine called Wendao (“asking the Way”): you write a direction brief, the AI team plays its turns, and the battle reports are the versions. It supports several large models, and you can use the platform’s credits or bring your own API keys. Around these three things come the calling card, achievements and guilds, so that those who grind and level up solo and those who team up for group battles can all see what they have built up. The platform’s incentive principle is distribution according to delight, not distribution according to labour: share and get fun in return; achievements, not counts.⁸⁹ We are also experimenting with a six-dimensional assessment tool called Wenheng (“the scales of letters”), which is only an instrument to aid judgement, not a judge.

We will proceed season by season. We are now in Season 0: publishing the manifesto and this set of topics. Season 1 opens a closed beta of Wendao and launches the first knowledge objects. Season 2 brings the first live research stream and the journal’s first issue. Season 3 launches calling cards, achievements and personal home spaces. Season 4 opens group battles, guilds and registration. The order may change; the principles will not: open formats, data you can take with you, no lock-in to any one model company. I do not know whether this road will get us there; I only know that the old road has become a dead end. Nor do we expect universities and journals to reform themselves. I have said that the future of humanities scholarship lies not in universities but outside the ivory tower, in companies and independent associations;³⁹ other disciplines may not move so fast, but this platform is a step in that direction.

We also want to be clear about what we will not do. We will not put up paywalls: knowledge objects are open by default. We will not run black-box rankings: every score must disclose its basis, and those who are scored may respond. We will not fence users in for any model company: Wendao can connect to different models, and researchers can take their work and data away with them in full. Every contribution on the platform – writing, explaining, editing, replicating, curating – will be recorded and credited. We welcome three kinds of people: curators who want to turn their research into knowledge objects, players who want to try a game of research for themselves, and those who would like to sit in the audience for a while first.

15In the end, why do science at all?#

What is research for? A secure university post that pays the bills? Developing technology and raising productivity – in short, making money for humankind? If those utilitarian ends were all there is, AI replacing humans might well be the natural course of things. But what drew us to research in the first place? Curiosity and the desire to know: the wish to understand this world, the mysteries of the universe and the richness of life. We can never let AI do our understanding for us.

It is like reading a mystery novel. There may be a stretch of the plot we cannot follow, and we would like someone to explain it, someone to talk it over with – and that someone can be an AI. But we would never hand the book to an AI to read and have it report back: “Found the murderer.” Perhaps it could crack the case faster than any human, but so what? Would we stop reading?

The mysteries of the universe are like the ending of a mystery novel: perhaps already written, perhaps already read by other people or by machines. Still, we want to read it for ourselves. Following the reasoning is the pleasure itself, and that is the original impulse of scientific inquiry and its deepest meaning.

Plato’s Socrates complained that written words are like a painting: they look alive, but ask them anything and they just say the same thing over and over again.¹⁹ For three hundred and sixty years, papers have been something like that painting. Today, for the first time, we have a chance to make writing answer back: evidence can be clicked open, processes replayed, explanations tailored to each reader; AI can accompany readers as they question their way down, layer by layer, and authors and peers can carry on the conversation on the same work. A new body may give writing back the power of dialogue.

Science has always had a theatrical side: the anatomical theatre at Leiden sold tickets, Boyle invited people to witness his air-pump experiments, and Faraday’s Christmas Lectures were packed. We are not afraid of a crowd; we are only afraid of the crowd taking the place of checking. When every turn of the research is seen, every piece of evidence can be clicked open and every AI contribution is logged, science becomes more honest than it was when hiding behind the paper.

Giving science back to curiosity is not a sentimental slogan but a reordering of values. Ever since science became a profession, academia, like every other trade, has practised distribution according to labour: the more you write, the more you get, so everyone writes more and everyone is caught in the rat race. AI has made labour no longer scarce, and I have argued for replacing “distribution according to labour” with “distribution according to delight”:⁸⁹ giving more to those who ask better questions, who can see what is interesting, and who can make things interesting for others. This begins with restoring pleasure’s good name. Curiosity was never high-minded. The Greeks sought knowledge partly to beat their opponents in debate and win honour, competing in earnest over useless things; scientists are still pugnacious overgrown boys, and priority disputes have never stopped.⁹⁰ We do not intend to whitewash all this as “pure curiosity”; we only want it to happen in the open, with rules, opponents and an audience. Delights come in higher and lower kinds, too – the sot and the sommelier both say they love wine – so taste must be cultivated, and standards must not drop. And the final yardstick for all of this is still “benefiting ordinary people”: the benefits of science and technology must in the end reach ordinary people’s daily lives, and neither curiosity nor the audience may run counter to that.

Science once belonged to the few who could read papers and survive the pyramid. Now it can belong again to everyone willing to ask, to watch and to judge. The paper was a great invention of the age of print, but it should not become the last idol of the age of AI. Science need not go back to ancient Greece, but it can recover the spirit in which the Greeks sought knowledge for the joy of it. Let us give knowledge a new body and give science back to curiosity. This is RenaScience: the renaissance of science.

Fourteen theses#

1. 1

   What we revive is the joy of inquiry, not the knowledge of the ancients.

2. 2

   The paper is a historical invention, not the destiny of knowledge.

3. 3

   Research can be a profession, but it must not be only a profession.

4. 4

   Journals change from containers for print into instruments of certification, curation and display.

5. 5

   Results are curated, living knowledge objects.

6. 6

   Replications and failures deserve to be recorded too.

7. 7

   AI is a transparent assistant, not a ghost author.

8. 8

   What humans contribute: questions, judgement, taste, responsibility and imagination.

9. 9

   The public, amateurs and border-crossers are all legitimate participants.

10. 10

    Friendly to people, and friendly to machines.

11. 11

    Publication is the beginning of public correction.

12. 12

    Judge work by whether it is credible, open, reproducible and something others can build on.

13. 13

    Science should be playable, watchable and open to participation.

14. 14

    Reputation is achievement, not a count; distribution according to delight; the moat is people.