v1.0 · 14 October 2026 · Open for comments
Claudio Pirrone
Abstract
Research on artificial intelligence (AI) in science has mostly asked how much AI impacts research productivity. This paper argues that science is not a homogeneous category, and asks which part of research AI can substitute for, which part it complements, and what follows for research policy. Taking a deductive approach, it defines science as the revision of shared beliefs through evidence that the community does not control. From this definition follows a distinction between boundary and inner research that rests neither on motive nor on appropriability. Inner research takes a paradigm’s problems and acceptance criteria as given, while boundary research allows itself to change them. Kuhn’s normal science is therefore inner research, and the performance of AI in research is strictly linked to it, because paradigms offer AI appropriate guidance in the form of training data and objectives. At the brink of knowledge, AI is limited by its own training. A system that scores claims by their fit to an existing corpus can generate corrections to that corpus but cannot select them without external judgment, the scarce human complement of AI. The recognition system of science assumes that within-paradigm results are costly to produce. Cheaper normal science empties performance metrics of their meaning, and an output-based recognition system raises the relative cost of judgment. AI thus presents science and scientists with a bifurcation. Science may blind itself and settle into a scholastic regime, in which agreement with the corpus becomes the criterion of truth, or we, as a community, may accept the risk of a critical regime, in which machines carry inner research and scientists concentrate on its failures.
Highlights
- The performance of AI in research is strictly linked to normal science
- Paradigms offer appropriate guidance for AI in science
- At the brink of knowledge, AI is limited by its own training
- Cheaper normal science empties performance metrics of their meaning
- Output-based recognition makes judgment relatively costlier, with policy implications
Keywords: Artificial intelligence, Normal science, Paradigms, Research evaluation, Priority rule, Science policy. JEL: O31, O33, O38, I23.
Provenance note
| Question | Author |
| Definitions and argument | Author |
| Sources | Claude Opus 5.5 (max). Integrated and checked against primary sources by author |
| Writing | Drafted by Claude under the author’s direction. Manual editing by author |
| Revision and final judgement | Author |
How to cite. Pirrone, C. (2026). Science is Dead, Long Live Science. Working draft v1.0, 14 October 2026. dremm.net.
1. Introduction
Artificial intelligence (AI) is entering scientific research faster than any earlier method of invention. In the natural sciences, researchers who use AI tools publish more and are cited more than those who do not, while the work done with these tools concentrates in established, data-rich areas (Hao et al., 2026). Language models are already used to draft a growing share of manuscripts (Kusumegi et al., 2025), and by 2026 they were solving open problems in mathematics. In September of that year, a declaration first signed by twenty-five Fields medalists acknowledged that these systems can now solve major outstanding problems, and warned that the race among AI companies to solve famous problems was harming the discipline (Avila et al., 2026).
The economics of science has so far treated AI mainly as an input that raises the productivity of research. Cockburn et al. (2019) describe deep learning as a general-purpose method of invention, and Bianchini et al. (2022) trace its diffusion across scientific fields. From our standpoint, this perspective is useful but incomplete. A technology that lowers the cost of one research activity changes the relative value of the others. The institutions of science, from the priority rule to bibliometric evaluation, were however designed around a particular structure of costs (Arrow, 1962; Dasgupta and David, 1994). The paper explores the consequences of the new cost structure induced by AI, a question that has only begun to be addressed (Vranceanu, 2026).
Section 2 adopts a functional definition of science, from which Section 3 derives a distinction between boundary and inner research that is independent of motive (Bush, 1945; Stokes, 1997) and of appropriability (Nelson, 1959). On this definition, Kuhn’s (1962) normal science and Dosi’s (1982) normal progress along technological trajectories are inner research. Section 4 asks which research can be made by AI and which can only be made with it. Section 5 draws the consequences for the recognition and evaluation system, and Sections 6 and 7 describe the two regimes into which science can settle and the policies that select between them.
The paper makes three contributions. It distinguishes the research activities that AI substitutes for from those it complements, it identifies the cost assumption on which priority and performance metrics rest and shows why AI breaks it, and it proposes levers that act on the direction of AI-enabled science, away from the scholastic trap, and assesses current policy against them.
2. A functional definition of science
Economists usually define science by its institutions. Merton (1957) placed the reward for priority at the centre of the scientific community, and Dasgupta and David (1994) distinguished open science from technology by the rules that govern disclosure and reward: scientists are paid in recognition for being first to publish, technologists in rents from keeping knowledge private. These definitions explain how science is organized, but they are silent about the knowledge it produces. A hint can be taken from Peirce (1877), who compared four ways in which communities settle their beliefs: holding on to what they already believe, deferring to an authority, adopting whatever seems most agreeable to reason, and testing beliefs against something that does not depend on what anyone believes. In his account, only the last method produces science, because it is the only one that can overturn a belief held by the whole community.
Peirce thus anticipated a characteristic of science that common sense, and sometimes scientists themselves, tend to obscure. Science is not truth. Scientific knowledge consists of beliefs that evolve under the pressure of demonstration, validation and falsification. We therefore define science as the practice of revising a community’s beliefs through evidence that the community does not control, in such a way that no belief, the current consensus included, is exempt from revision.
A first corollary follows trivially. Consensus is an output of science, not a criterion for it, and taking it as the criterion would make the definition circular. In Peirce’s terms, a procedure that accepts claims because they agree with what is already accepted is the method of authority or the a priori method, however sophisticated its means. A second corollary is that scientific progress should be measured by the improvement of shared beliefs rather than by the number of results produced. The evidence suggests that the two have been drifting apart. Research effort has grown steadily while the output of ideas per researcher has fallen (Bloom et al., 2020); in larger fields, new papers rarely displace the most cited ones (Chu and Evans, 2021); and papers and patents appear to have become less disruptive over several decades (Park et al., 2023), although the size of this decline is disputed (Holst et al., 2026).
3. Boundary and inner research
To the best of our knowledge, the distinction between basic and applied research that organizes science policy has never been derived from a definition of science. Bush (1945) defined basic research as research performed without thought of practical ends, and research statistics have followed that definition ever since. Stokes (1997) showed that motive has at least two dimensions, since a quest for fundamental understanding and a consideration of use can coexist, as they did in Pasteur’s work. Nelson (1959) classified research by the uncertainty and appropriability of its results, which explains why markets underinvest in it. These classifications concern why research is done and who captures its returns, not what research does to the knowledge on which it builds, which is what the functional definition makes central.
To capture this, we rely on the notion of a paradigm. For Kuhn (1962), a paradigm defines the problems a community regards as worth solving and the standards a proposed solution must meet. Dosi (1982) carried the notion into the economics of technical change, where a paradigm defines a set of problems and a ‘pattern of solution’ along which normal technical progress proceeds. In the present framework, inner research takes a paradigm’s problems and acceptance criteria as given and produces results that the paradigm can itself evaluate. New methods that solve existing problems to existing standards belong here too. Boundary research changes the problems a community considers worth solving or the criteria by which it accepts solutions. The name follows Abbott (1995), who argued that entities come into existence when boundaries are tied together. Boundary research reshapes the boundaries out of which a paradigm is made. Methods that push standards beyond their former boundaries belong here.
Consequently, normal science is always inner research, whatever the motive of those who perform it. A physicist who computes a higher-order correction within an accepted theory out of pure curiosity is doing inner research, while a clinician who finds that patients do not respond as the accepted model predicts, and revises the model, is doing boundary research. Nor can the distinction be drawn by difficulty. Predicting a protein’s structure from its amino-acid sequence resisted half a century of effort, yet it was solved by a system trained on the Protein Data Bank and assessed in the community’s own blind competition (Jumper et al., 2021). The problem, the data and the criterion of success all predated the solution. By the same reasoning, a long-standing open problem in mathematics that is settled with existing frameworks is inner research. Finally, the classification applies to pieces of work, not to people or fields, and today the same researcher typically does both kinds of research.
It follows that a large share of what research statistics record as basic research is inner research in the present sense. This is not a judgment on its value. Normal science is how paradigms are articulated, tested and eventually broken. It does, however, identify the human scientific activity most exposed to competition from AI.
4. Made by AI, or with AI?
Agrawal et al. (2019) characterize recent advances in AI as a fall in the cost of prediction, and argue that their economic consequences depend on the complements of prediction, above all judgment, by which they mean the assessment of what different outcomes are worth. Normal science can be written as a prediction problem with a known objective. The paradigm’s literature and data provide the training material, its problems provide the objective, and its acceptance criteria provide the standard against which candidate results are scored. Paradigms therefore offer AI appropriate guidance, and this is why the performance of AI in research is strictly linked to normal science. Where the criteria can be checked mechanically, as with formal proofs or with simulations whose physics is settled, the loop from candidate to accepted result can run without human intervention but posterior control. Where acceptance requires new observations, the loop runs at the speed of experimentation, but its logic is unchanged. Research of this kind can be made by AI.
The evidence so far agrees. The gains documented by Hao et al. (2026) accrue to researchers working on established, data-rich problems, while the collective range of topics contracts, as one would expect if AI accelerates work along existing trajectories without opening new ones. Mathematics, where acceptance criteria are the most formalized, is the field in which substitution has been most visible.
Boundary research cannot be specified in this way in advance, because it changes the objective itself. At the brink of knowledge, AI is limited by its own training, that is, by the corpus it learned from and the objective it was trained on. Large language models are best described as stochastic pattern-completion systems, whose outputs follow the statistical structure of the texts on which they were trained rather than a model of the world against which those texts could be checked (Quattrociocchi et al., 2025). When such systems are used to evaluate claims, their verdicts can match those of experts even when they rest on lexical association rather than on evidence, a condition Loru et al. (2025) call epistemia. Inside a paradigm this matters little, because agreement with the literature and correctness largely coincide. At its boundary it matters a great deal. A system that scores candidate claims by their agreement with an existing corpus can generate the correction of an error contained in that corpus, but it cannot select the correction by its own criterion, because the correction disagrees with the corpus.
Generation is not the bottleneck. Recombination of existing knowledge can produce unexpected candidates, and AI can be deliberately steered away from the questions researchers are pursuing, which yields promising hypotheses that people would otherwise reach late or never (Sourati and Evans, 2023). But the decision to search away from the consensus was taken by the designers of the system, and it is the kind of judgment that must come from outside the corpus. Left to itself, the tendency runs the other way, since models trained on their own outputs progressively lose the tails of the original distribution (Shumailov et al., 2024).
The limit concerns criteria rather than machines. A system given an external criterion, such as an experiment it can run, can select a correction against that criterion. What it cannot do on its own is decide which anomalies deserve such a test, or which criteria should replace those that have failed. At the brink of knowledge, research is therefore, at best, made with AI. If the scientist can state the problem, the rules and the objectives clearly, AI models are indifferent to the nature of the research, and they lower the cost of access to the conventional knowledge needed to solve an unconventional problem. Scoping and control, however, have to stay outside the automated loop. In the blow-up proof for the Navier–Stokes equations recently announced by OpenAI, an AI translation of the natural-language proof into Lean compiled and was verified, yet it proves a weaker estimate than the proof it was meant to certify (Bastounis et al., 2026). In well-established domains, such errors may be short-lived. At the frontier, they are caught only by researchers who know their field better than the systems they use.
In the terms of Agrawal et al. (2019), AI substitutes for prediction within paradigms and complements judgment about them, so the value of that judgment should rise as AI spreads. Whether it does depends on the institutions that recognize and fund research.
5. Consequences for the recognition and evaluation system
The recognition system of science rests on an assumption about costs. Knowledge is costly to produce and cheap to transmit. Its efficient use requires that it be free, which leaves its production unrewarded unless the reward is ‘completely separated from any charge to the users of the information’ (Arrow, 1962, p. 617). Priority is such a reward. Scientists are paid in recognition for being first to disclose, and since a discovery can be made only once, a prize for being first induces both effort and disclosure (Merton, 1957; Dasgupta and David, 1994). For performance metrics based on counts of publications and citations, the assumption is implicit. As long as each result required substantial and roughly comparable effort, the number of results carried information about the effort behind them and, more weakly, about their contribution. This is why publication records could work as a screening device, and why language models, by lowering the cost of publishable research, weaken it (Vranceanu, 2026).
Both mechanisms presuppose that within-paradigm results are costly, and AI sharply lowers that cost for inner research. Cheaper normal science empties performance metrics of their meaning. They come to measure access to the means of producing results rather than contribution. The priority race shows the first effect. If a result can be produced by whoever controls a sufficiently capable system, being first rewards control of the system, not insight, and the race shifts toward announcements made before methods are isolated, written up or related to earlier work. This is the pattern the Fields medalists’ declaration describes (Avila et al., 2026). Output indicators show the second. Publication counts were always a weak proxy for contribution, and the Leiden Manifesto warned against their mechanical use (Hicks et al., 2015). When output can be multiplied at low cost, even the weak link breaks. Among manuscripts drafted with language models, linguistic complexity, once a signal of quality, is now associated with weaker substance (Kusumegi et al., 2025).
An output-based recognition system also raises the relative cost of judgment. The return to a research hour is the recognition it earns. AI multiplies the output of an hour spent on inner research, while an hour spent weighing anomalies or proposing new criteria yields what it yielded before. Measured in forgone recognition, the cost of judgment therefore rises with every gain in the productivity of AI. Researchers move their effort toward inner research, and funders that allocate by output per unit of cost follow them, rationally and continuously.
The effect compounds a known bias. Novel research is penalized by short-window bibliometric indicators, because its impact arrives later and is more variable (Wang et al., 2017), and novel work is where boundary research is most likely to be found. An evaluation and funding system that already undervalued novelty will undervalue it further once AI transforms the denominator of every productivity ratio.
Together, these effects identify what is ending. It is an institutional arrangement in which within-paradigm results were scarce, rewarded through priority, counted by indicators and funded by output. That arrangement is what most researchers, funders and evaluators mean by science as we know it, and it is the arrangement that AI makes obsolete.
6. Two self-reinforcing regimes
These pressures are likely to be self-reinforcing. Where choices are subject to increasing returns, small early differences can lock a system into one of several possible states (Arthur, 1989). March (1991) showed that adaptive processes which refine the exploitation of existing knowledge faster than the exploration of new possibilities are effective in the short run and self-destructive in the long run. AI makes exploitation within paradigms far cheaper than before, so the question is whether science will adapt in this self-destructive way. Figure 1 summarizes the two regimes into which it can settle.
Inner research automated
within-paradigm results at low cost
↓ model fit as the test, rewards for output
Scholastic regime
Criterion: fit to the corpus
Rewards: priority and counts
Judgment delegated to models
⇅
Reinforcing loop
model collapse, monoculture, LLMorphism
↓ outputs as conjectures, rewards for judgment
Critical regime
Criterion: external evidence
Rewards: problems, anomalies
Machines compute residuals
⇅
Reinforcing loop
more residuals, more demand for judgment
Figure 1. Two self-reinforcing regimes once inner research is automated. Which one science settles into depends on the capabilities of researchers, on whether model output counts as a test or as a conjecture, and on what institutions reward.
In the first, which we call the scholastic regime, agreement with models trained on existing knowledge is accepted as a test of new claims, and researchers are rewarded for output. Several mechanisms then reinforce one another. Models increasingly learn from text produced by earlier models, which narrows their output toward the centre of the distribution (Shumailov et al., 2024). When many decision-makers rely on the same model, their collective decisions can deteriorate even if the model is more accurate than each of them alone, because their errors become correlated (Kleinberg and Raghavan, 2021). Researchers who delegate evaluation to fluent systems may come to see their own cognition as a form of prediction, a bias Capraro (2026) calls LLMorphism, and lose the incentive to develop the judgment that the corpus cannot supply. Each mechanism lowers the supply of dissent and raises its cost. The end point is a science in which agreement with the existing corpus has become the operative criterion of truth, which is Peirce’s method of authority implemented by machines. Paradigm dixit.
The second, which we call the critical regime, splits inner and boundary research clearly. Both remain legitimate research. Only boundary research leaves no belief exempt from revision, and inner research meets the definition of Section 2 only as part of a practice that also revises its paradigm. In the simplified framework adopted, inner research is made by AI, and boundary research is made with AI, in the sense of Section 4. Paradigms keep guiding the machines, and scientists use AI to explore the hidden faces of a paradigm before deciding in which direction, if any, to depart from it. What they cannot do is ask for models trained to find the failures of their own paradigm, since such a training objective would have to contain the criterion that the search is meant to discover. If we, as a community, want to leverage AI at the brink of knowledge, researchers there have to be capable of riding the bull.
This regime, too, has its own feedback. Kuhn (1962) observed that anomalies become visible only against the precise expectations that normal science creates. Inner research carried out at machine scale should therefore multiply the residuals that call for interpretation, raising the demand for judgment and, through practice, its supply. Human performance can respond in this way: after AI systems surpassed human players at Go, professional players made better decisions and played previously unobserved moves more often (Shin et al., 2023).
The two regimes also shed light on an old dispute. Popper (1970) regarded the normal scientist as the product of dogmatic training and warned that, if such work became the rule, it would be ‘a danger to science and, indeed, to our civilization’. Kuhn (1962) regarded normal science as the precondition of scientific revolutions. On the present argument, each was right about a different thing. Normal science is a necessary function, as Kuhn held, and a dangerous identity for scientists, as Popper feared. AI separates the function from the identity. Whether the separation produces a critical or a scholastic science is not determined by the technology, which is the same in both regimes, but by the capabilities of researchers, the procedures by which AI output is used, and the recognition system.
7. Policy implications
If the argument is correct, the main policy question is how to steer the use of AI in research toward the critical regime. Section 6 traced the outcome to the procedures by which AI output is used, to the recognition system and to the capabilities of researchers. None of the levers that act on them depends on a forecast of AI capabilities. As long as AI lowers the cost of within-paradigm results faster than the cost of judgment, the institutions that recognize and fund research must change if boundary research is to keep its place.
Current policy acts on other margins. The largest public programmes, the US Genesis Mission and the European strategy for AI in science, aim at the rate of AI-enabled science, by training models on public data and accelerating their adoption (Executive Order 14363, 2025; European Commission, 2025). Both invest in what AI already does best, which on the present argument is inner research, and neither changes how results are recognized. The US order counts publications among the outcomes to be reported (Executive Order 14363, 2025), so that a programme moving effort from boundary to inner research would report success. The rules adopted so far by publishers, funders and research organizations concern the provenance of research outputs and the confidentiality of review.
7.1. Validation and verification
Journals, funders and research institutions can adopt the rule that agreement with a model never counts as validation, so that a model’s output remains a conjecture until it has passed a test that the model did not set. Publishers and their ethics bodies exclude AI tools from authorship, because the tools cannot take responsibility for the work, and require authors to disclose how they were used (Nature, 2023; COPE, 2023). The principle is sound, since it keeps someone outside the corpus answerable for every claim. Disclosure, however, records provenance, and a disclosed model output has the same epistemic status as an undisclosed one. The European AI Act extends this approach to the providers of models, who must make generated text detectable by machines as artificially generated, while text published under human editorial responsibility need not be labelled (European Parliament and Council, 2024, Art. 50). The exemption rests on a person who answers for the text, as the argument requires. The mark records only origin, and its main use in research would be to enforce rules based on provenance, such as the originality rule discussed in Section 7.3, so it should carry no weight in evaluation. Journals that already ask how AI was used could also ask which test each claim resting on model output has passed, and who set it. For formal proofs, the test includes a check, by people who know the field, that the formal statement says what the paper claims (Bastounis et al., 2026). Registered Reports, which are reviewed before results are known, separate the test from the result (Chambers and Tzavella, 2022), and the format suits hypotheses generated by models.
The rule matters most where models reconcile awkward observations by generating plausible auxiliary hypotheses, because that would build at machine speed the protective belt that Lakatos (1970) associated with degenerating research programmes. Models used in research can instead be required to report the residuals they compute, a within-paradigm task that raises none of the difficulties discussed in Section 4. The same logic applies to the evaluation of the models themselves, which the European strategy plans to entrust to a dedicated hub (European Commission, 2025). Benchmarks based on famous problems measure only the capacity for inner research, and reading them as measures of progress is what the Fields medalists object to (Avila et al., 2026). Metrics consistent with the critical regime would also score whether a model reports anomalies planted in its inputs.
Validation also requires capacity, and the evidence does not support handing verification to the models themselves. In a randomized comparison of teams reproducing published studies, those led by AI achieved a reproduction rate of 37%, against 94% for teams of humans alone (Brodeur et al., 2026). Verification capacity, from formal proof checking to replication, should therefore be financed as a public good, because its value rises with the volume of results to be checked and no individual researcher has an incentive to supply it. Programmes that pay to accelerate the production of results are the natural source of that finance.
7.2. Evaluation and credit
The reform of research assessment is moving away from counts, and the signatories of the Agreement on Reforming Research Assessment commit to abandoning inappropriate uses of the journal impact factor and the h-index (CoARA, 2022). The direction is right, and the argument adds a reason for it, since counts that were noisy proxies for contribution become uninformative once inner research is cheap. Public recognition lags behind. The annual list of the top 2% of scientists, computed on Scopus and published in Elsevier’s data repository, ranks researchers by a composite of citation counts and h-indices, although its authors warn against reading it as evidence of excellence (Ioannidis, 2026). Rankings of this kind increasingly reward access to the means of producing results, and the argument implies that they should carry no weight in hiring and promotion.
The same agreement makes peer review central to assessment (CoARA, 2022), and this needs a qualification. Review that asks whether work is sound and plausible by the standards of its field asks a question that language models answer cheaply, and reviewers already use them to answer it. At four major AI conferences, between 6.5% and 16.9% of the text of peer reviews may have been substantially modified by language models (Liang et al., 2024). Qualitative evaluation protects the critical regime only if reviewers are asked a different question, namely what problem, anomaly or acceptance criterion the work introduced, and what evidence outside the corpus supports it.
Open Research Europe, the European Commission’s publishing platform, changes when review happens and who can read it. Articles are published first and reviewed afterwards by invited experts, whose names and reports are public (European Commission, 2021). Publishing first leaves a claim public as a conjecture awaiting its test, and published reports turn judgment into a visible contribution that can be cited and credited. Both features agree with the argument. The absence of anonymity does not. Publishing first cannot hide the authors, and referees who see a prominent name recommend rejection far less often (Huber et al., 2022), which is Peirce’s method of authority applied to review. Signing raises the cost of dissent, since reviewers asked to sign decline more often and those who sign recommend publication more often (van Rooyen et al., 1999; Walsh et al., 2000), while publishing reports without names has not harmed review (Bravo et al., 2019). When authors propose the reviewers, as on this platform (Open Research Europe, 2021), they partly choose their own test, and such reviewers recommend acceptance more often in the first round (Wager et al., 2006). The argument therefore supports public reports, with names left to the reviewers and reviewers chosen by editors.
Results that AI systems can produce within a paradigm should be treated as research infrastructure, verified and deposited as data and software increasingly are, and credited accordingly. Credit tied to publication should be reserved for contributions that change what a field regards as a problem or a solution, such as new questions, documented anomalies and new acceptance criteria. Contributor taxonomies already record who formulated the research goals and who verified the results (Brand et al., 2015), so evaluators can weight these roles above the execution that AI increasingly performs.
Priority needs a similar correction, given the race described in Section 5. Mathematics has a rule that resists early announcements. The Clay Mathematics Institute considers a solution to a Millennium Prize Problem only if at least two years have passed since its publication and it has gained general acceptance among mathematicians (Clay Mathematics Institute, 2018). Extending the principle to AI-produced results generally would date priority from the disclosure of the full method and of its verification. Dated from announcements, priority rewards control of a system.
7.3. Funding
Section 5 implies that output-based allocation will progressively starve boundary research, and grant applications are themselves output. Across twelve funders, application numbers rose by 57% between 2022 and 2025 (Rees and Wilsdon, 2026). The US National Institutes of Health reacted by capping applications at six per investigator per year, and by declaring that applications ‘substantially developed by AI’ are not ‘original ideas of applicants’ (National Institutes of Health, 2025). The diagnosis agrees with the argument, since a review system built for costly proposals is being flooded with cheap ones. The remedy is weaker. A cap rations volume without changing the criterion of selection. The originality rule defines originality by provenance, although a proposal drafted with a model can change what a field regards as a problem, while one written without it may only extend a paradigm. The rule is also hard to enforce, since reliable estimates of AI-modified text are obtained for corpora and do not identify single documents (Liang et al., 2024).
Other proposals answer the flood with more AI. Rees and Wilsdon (2026) suggest using AI agents to profile and shortlist applicants, among other means through predictive heuristics for novelty. In 2026, a UK research network used language models to screen out almost half of its applications before any human read them, and announced after protests that it would review them again (Jones, 2026). The European guidelines exclude generative AI from ‘the assessment or evaluation of the scientific content of projects’ (European Commission, 2026), and the National Institutes of Health prohibits its use in peer review (National Institutes of Health, 2023). The prohibitions are justified on grounds stronger than those given, which are confidentiality, hallucinations and bias. Run on secure servers and free of hallucinations and bias, a model would still rank proposals by their fit to the corpus. It would penalize novelty as reviewers already do (Boudreau et al., 2016), with errors correlated across every panel that used it (Kleinberg and Raghavan, 2021). Studies measuring how closely model scores track those of reviewers (Thelwall et al., 2026) apply the wrong standard to boundary research.
Partial lotteries act on the criterion instead. The Swiss National Science Foundation draws lots among proposals too close to the funding line for review to separate (Heyard et al., 2022), and Fang and Casadevall (2016) proposed a lottery among all proposals above a threshold of merit. The practice deserves extension. As AI lowers the cost of a polished proposal, the band that review cannot separate widens, and funders may soon have to choose among many high-quality proposals in largely arbitrary ways (Rees and Wilsdon, 2026). A lottery over the whole band removes the return to polish and leaves reviewers a single judgment at the threshold, which can ask whether a proposal exposes its claims to evidence that its authors do not control.
Instruments that fund people over long horizons and tolerate early failure produce more novel and more highly cited work than instruments with short review cycles (Azoulay et al., 2011), and they shield judgment from the rise in its relative cost. They deserve support on condition that people are judged on the problems they opened and the anomalies they documented, since a record counted in outputs carries the inflation described in Section 5. National programmes for AI in science can apply the same principle, by reserving part of their budgets for boundary research made with AI and by counting verified results and documented anomalies among their outcomes.
7.4. Training
The critical regime depends on researchers able to recognize when a model’s output agrees with the corpus and disagrees with the evidence. Current policy treats training mainly as instruction in the tools. The European strategy calls for AI skills at all career stages (European Commission, 2025), and the European guidelines ask for training on generative AI ‘especially (but not exclusively) on verifying output’ (European Commission, 2026). Of the two emphases, the second is the one the argument requires. Verification draws on command of a paradigm, which is acquired by solving its problems, and AI makes solving them optional. Problems with known solutions therefore remain the best way to learn a paradigm (Avila et al., 2026), but solving them becomes education rather than output, and doctoral programmes should examine that ability without machines. AI assistance reduced the detection of major errors in reproduction (Brodeur et al., 2026), while professional Go players made better decisions after AI surpassed them (Shin et al., 2023), so the same ability can be weakened or strengthened by the way machines are used.
8. Conclusion
The effect of AI on science depends less on how capable AI becomes than on which part of research it substitutes for. Its performance is strictly linked to normal science, where paradigms offer appropriate guidance, while at the brink of knowledge it is limited by its own training, and research there needs a judgment that a system trained on the existing corpus cannot exercise by its own criterion. The recognition system of science was built for a world in which normal science was costly. As that world ends, performance metrics lose their meaning and output-based recognition raises the relative cost of judgment. Which regime follows, scholastic or critical, depends on researchers and institutions, since the technology is the same in both.
The positive argument does not depend on any value judgment, and one is added only at the end, explicitly as such. We regard the critical regime as preferable not only because it is likely to be more productive in the long run, but because it is more adult, in the sense in which Kant (1784) defined enlightenment as the exit from a self-imposed dependence on the guidance of others in the use of one’s own understanding. Normal science asked most scientists to work under the guidance of a paradigm. If machines can now do that work, science as we have known it is ending, and the activity it was organized to serve can continue without it.
References
Abbott, A., 1995. Things of boundaries. Social Research 62, 857–882. https://philpapers.org/rec/ABBTOB.
Agrawal, A., Gans, J.S., Goldfarb, A., 2019. Exploring the impact of artificial intelligence: Prediction versus judgment. Information Economics and Policy 47, 1–6. https://doi.org/10.1016/j.infoecopol.2019.05.001.
Arrow, K.J., 1962. Economic welfare and the allocation of resources for invention, in: The Rate and Direction of Inventive Activity: Economic and Social Factors. Princeton University Press, Princeton, pp. 609–626. https://doi.org/10.1515/9781400879762-024.
Arthur, W.B., 1989. Competing technologies, increasing returns, and lock-in by historical events. Economic Journal 99, 116–131. https://doi.org/10.2307/2234208.
Avila, A., Bhargava, M., Birkar, C., Deligne, P., Deng, Y., et al., 2026. A severe misalignment of AI in mathematics. Declaration, 11 September. Zenodo. https://doi.org/10.5281/zenodo.22737750.
Azoulay, P., Graff Zivin, J.S., Manso, G., 2011. Incentives and creativity: Evidence from the academic life sciences. RAND Journal of Economics 42, 527–554. https://doi.org/10.1111/j.1756-2171.2011.00140.x.
Bastounis, A., Circelli, F., Hansen, A.C., 2026. Navier–Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs. arXiv:2610.08144. https://arxiv.org/abs/2610.08144.
Bianchini, S., Müller, M., Pelletier, P., 2022. Artificial intelligence in science: An emerging general method of invention. Research Policy 51, 104604. https://doi.org/10.1016/j.respol.2022.104604.
Bloom, N., Jones, C.I., Van Reenen, J., Webb, M., 2020. Are ideas getting harder to find? American Economic Review 110, 1104–1144. https://doi.org/10.1257/aer.20180338.
Boudreau, K.J., Guinan, E.C., Lakhani, K.R., Riedl, C., 2016. Looking across and looking beyond the knowledge frontier: Intellectual distance, novelty, and resource allocation in science. Management Science 62, 2765–2783. https://doi.org/10.1287/mnsc.2015.2285.
Brand, A., Allen, L., Altman, M., Hlava, M., Scott, J., 2015. Beyond authorship: Attribution, contribution, collaboration, and credit. Learned Publishing 28, 151–155. https://doi.org/10.1087/20150211.
Bravo, G., Grimaldo, F., López-Iñesta, E., Mehmani, B., Squazzoni, F., 2019. The effect of publishing peer review reports on referee behavior in five scholarly journals. Nature Communications 10, 322. https://doi.org/10.1038/s41467-018-08250-2.
Brodeur, A., Valenta, D., Marcoci, A., Aparicio, J.P., Mikola, D., Barbarioli, B., et al., 2026. AI-assisted teams outperform AI-led teams but not human-only teams in assessing research reproducibility in quantitative social science. Proceedings of the National Academy of Sciences 123, e2524747123. https://doi.org/10.1073/pnas.2524747123.
Bush, V., 1945. Science, the Endless Frontier. United States Government Printing Office, Washington, DC. https://www.govinfo.gov/content/pkg/GOVPUB-PR32_400-e7966ee70a4f7b47f862431c9776f727/html/GOVPUB-PR32_400-e7966ee70a4f7b47f862431c9776f727.htm.
Capraro, V., 2026. LLMorphism: When humans come to see themselves as language models. arXiv:2605.05419. https://arxiv.org/abs/2605.05419.
Chambers, C.D., Tzavella, L., 2022. The past, present and future of Registered Reports. Nature Human Behaviour 6, 29–42. https://doi.org/10.1038/s41562-021-01193-7.
Chu, J.S.G., Evans, J.A., 2021. Slowed canonical progress in large fields of science. Proceedings of the National Academy of Sciences 118, e2021636118. https://doi.org/10.1073/pnas.2021636118.
Clay Mathematics Institute, 2018. Rules for the Millennium Prize Problems. Adopted by the Board of Directors of the Clay Mathematics Institute, 26 September. https://www.claymath.org/wp-content/uploads/2022/03/millennium_prize_rules_0.pdf.
CoARA, 2022. Agreement on reforming research assessment. Coalition for Advancing Research Assessment, 20 July. Zenodo. https://doi.org/10.5281/zenodo.13480728.
Cockburn, I.M., Henderson, R., Stern, S., 2019. The impact of artificial intelligence on innovation: An exploratory analysis, in: Agrawal, A., Gans, J., Goldfarb, A. (Eds.), The Economics of Artificial Intelligence: An Agenda. University of Chicago Press, Chicago, pp. 115–146. https://doi.org/10.7208/chicago/9780226613475.003.0004.
COPE, 2023. Authorship and AI tools. Position statement, Committee on Publication Ethics, 13 February. https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools.
Dasgupta, P., David, P.A., 1994. Toward a new economics of science. Research Policy 23, 487–521. https://doi.org/10.1016/0048-7333(94)01002-1.
Dosi, G., 1982. Technological paradigms and technological trajectories: A suggested interpretation of the determinants and directions of technical change. Research Policy 11, 147–162. https://doi.org/10.1016/0048-7333(82)90016-6.
European Commission, 2021. Official launch of Open Research Europe (ORE), the European Commission’s open access publishing platform. Note of the Director-General for Research and Innovation to Horizon 2020 beneficiaries, Ares(2021)2390526, 7 April. https://ec.europa.eu/info/funding-tenders/opportunities/docs/2021-2027/horizon/other/comm/open-research-europe_official-launch_en.pdf.
European Commission, 2025. A European strategy for artificial intelligence in science: Paving the way for the Resource for AI Science in Europe (RAISE). Communication from the Commission, COM(2025) 724 final, 8 October. https://data.consilium.europa.eu/doc/document/ST-13720-2025-INIT/en/pdf.
European Commission, 2026. Living guidelines on the responsible use of generative AI in research. ERA Forum Stakeholders’ document, third version, May. Directorate-General for Research and Innovation, Brussels. https://research-and-innovation.ec.europa.eu/document/download/2b6cf7e5-36ac-41cb-aab5-0d32050143dc_en?filename=ec_rtd_ai-guidelines.pdf.
European Parliament and Council, 2024. Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L series, 12 July. http://data.europa.eu/eli/reg/2024/1689/oj.
Executive Order 14363, 2025. Launching the Genesis Mission. Federal Register 90(227), 55035–55039, 28 November. https://www.govinfo.gov/content/pkg/FR-2025-11-28/html/2025-21665.htm.
Fang, F.C., Casadevall, A., 2016. Research funding: The case for a modified lottery. mBio 7, e00422-16. https://doi.org/10.1128/mBio.00422-16.
Hao, Q., Xu, F., Li, Y., Evans, J., 2026. Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature 649, 1237–1243. https://doi.org/10.1038/s41586-025-09922-y.
Heyard, R., Ott, M., Salanti, G., Egger, M., 2022. Rethinking the funding line at the Swiss National Science Foundation: Bayesian ranking and lottery. Statistics and Public Policy 9, 110–121. https://doi.org/10.1080/2330443X.2022.2086190.
Hicks, D., Wouters, P., Waltman, L., de Rijcke, S., Rafols, I., 2015. Bibliometrics: The Leiden Manifesto for research metrics. Nature 520, 429–431. https://doi.org/10.1038/520429a.
Holst, V., Algaba, A., Tori, F., Wenmackers, S., Ginis, V., 2026. Dataset artefacts can partially drive the measured decline in disruption. Nature 656, E7–E13. https://doi.org/10.1038/s41586-026-10787-y.
Huber, J., Inoua, S., Kerschbamer, R., König-Kersting, C., Palan, S., Smith, V.L., 2022. Nobel and novice: Author prominence affects peer review. Proceedings of the National Academy of Sciences 119, e2205779119. https://doi.org/10.1073/pnas.2205779119.
Ioannidis, J.P.A., 2026. August 2026 data-update for “Updated science-wide author databases of standardized citation indicators”. Elsevier Data Repository, version 9. https://doi.org/10.17632/btchxktzyw.9.
Jones, F., 2026. Grant proposals rejected by ‘AI triage’ to be reassessed. Times Higher Education, 5 October. https://www.timeshighereducation.com/news/grant-proposals-rejected-ai-triage-be-reassessed.
Jumper, J., Evans, R., Pritzel, A., et al., 2021. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. https://doi.org/10.1038/s41586-021-03819-2.
Kant, I., 1784. Beantwortung der Frage: Was ist Aufklärung? Berlinische Monatsschrift 4, 481–494. https://www.deutschestextarchiv.de/book/show/kant_aufklaerung_1784.
Kleinberg, J., Raghavan, M., 2021. Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences 118, e2018340118. https://doi.org/10.1073/pnas.2018340118.
Kuhn, T.S., 1962. The Structure of Scientific Revolutions. University of Chicago Press, Chicago. https://press.uchicago.edu/ucp/books/book/chicago/S/bo13179781.html.
Kusumegi, K., Yang, X., Ginsparg, P., de Vaan, M., Stuart, T., Yin, Y., 2025. Scientific production in the era of large language models. Science 390, 1240–1243. https://doi.org/10.1126/science.adw3000.
Lakatos, I., 1970. Falsification and the methodology of scientific research programmes, in: Lakatos, I., Musgrave, A. (Eds.), Criticism and the Growth of Knowledge. Cambridge University Press, Cambridge, pp. 91–196. https://doi.org/10.1017/CBO9781139171434.009.
Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., McFarland, D.A., Zou, J.Y., 2024. Monitoring AI-modified content at scale: A case study on the impact of ChatGPT on AI conference peer reviews, in: Proceedings of the 41st International Conference on Machine Learning, pp. 29575–29620. https://proceedings.mlr.press/v235/liang24b.html.
Loru, E., Nudo, J., Di Marco, N., Santirocchi, A., Atzeni, R., Cinelli, M., Cestari, V., Rossi-Arnaud, C., Quattrociocchi, W., 2025. The simulation of judgment in LLMs. Proceedings of the National Academy of Sciences 122, e2518443122. https://doi.org/10.1073/pnas.2518443122.
March, J.G., 1991. Exploration and exploitation in organizational learning. Organization Science 2, 71–87. https://doi.org/10.1287/orsc.2.1.71.
Merton, R.K., 1957. Priorities in scientific discovery: A chapter in the sociology of science. American Sociological Review 22, 635–659. https://doi.org/10.2307/2089193.
National Institutes of Health, 2023. The use of generative artificial intelligence technologies is prohibited for the NIH peer review process. Notice NOT-OD-23-149, 23 June. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html.
National Institutes of Health, 2025. Supporting fairness and originality in NIH research applications. Notice NOT-OD-25-132, 17 July. https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-132.html.
Nature, 2023. Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613, 612. https://doi.org/10.1038/d41586-023-00191-1.
Nelson, R.R., 1959. The simple economics of basic scientific research. Journal of Political Economy 67, 297–306. https://doi.org/10.1086/258177.
Open Research Europe, 2021. Open peer review: How to find reviewers. Guide for authors. https://think.f1000research.com/wp-content/uploads/2021/07/ORE-Guide-How-to-find-reviewers.pdf.
Park, M., Leahey, E., Funk, R.J., 2023. Papers and patents are becoming less disruptive over time. Nature 613, 138–144. https://doi.org/10.1038/s41586-022-05543-x.
Peirce, C.S., 1877. The fixation of belief. Popular Science Monthly 12, 1–15. https://en.wikisource.org/wiki/Popular_Science_Monthly/Volume_12/November_1877/Illustrations_of_the_Logic_of_Science_I.
Popper, K.R., 1970. Normal science and its dangers, in: Lakatos, I., Musgrave, A. (Eds.), Criticism and the Growth of Knowledge. Cambridge University Press, Cambridge, pp. 51–58. https://doi.org/10.1017/CBO9781139171434.007.
Quattrociocchi, W., Capraro, V., Perc, M., 2025. Epistemological fault lines between human and artificial intelligence. arXiv:2512.19466. https://arxiv.org/abs/2512.19466.
Rees, G., Wilsdon, J., 2026. Could agentic AI topple grant-funding systems? Nature 652, 1119–1121. https://doi.org/10.1038/d41586-026-01297-y.
van Rooyen, S., Godlee, F., Evans, S., Black, N., Smith, R., 1999. Effect of open peer review on quality of reviews and on reviewers’ recommendations: A randomised trial. BMJ 318, 23–27. https://doi.org/10.1136/bmj.318.7175.23.
Shin, M., Kim, J., van Opheusden, B., Griffiths, T.L., 2023. Superhuman artificial intelligence can improve human decision-making by increasing novelty. Proceedings of the National Academy of Sciences 120, e2214840120. https://doi.org/10.1073/pnas.2214840120.
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., Gal, Y., 2024. AI models collapse when trained on recursively generated data. Nature 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y.
Sourati, J., Evans, J.A., 2023. Accelerating science with human-aware artificial intelligence. Nature Human Behaviour 7, 1682–1696. https://doi.org/10.1038/s41562-023-01648-z.
Stokes, D.E., 1997. Pasteur’s Quadrant: Basic Science and Technological Innovation. Brookings Institution Press, Washington, DC. https://www.brookings.edu/books/pasteurs-quadrant/.
Thelwall, M., Pride, D., Osborne, F., Salatino, A., Knoth, P., Iyawa, G., Yusuf, O., 2026. Scoring grant applications with large language models. arXiv:2609.25327. https://arxiv.org/abs/2609.25327.
Vranceanu, R., 2026. The end of the world as we know it? LLMs and the future of publish-or-perish. Working Paper 2606. ESSEC Business School. https://faculty.essec.edu/research/16676-the-end-of-the-world-as-we-know-it-llms-and-the-future-of-publish-or-perish/.
Wager, E., Parkin, E.C., Tamber, P.S., 2006. Are reviewers suggested by authors as good as those chosen by editors? Results of a rater-blinded, retrospective study. BMC Medicine 4, 13. https://doi.org/10.1186/1741-7015-4-13.
Walsh, E., Rooney, M., Appleby, L., Wilkinson, G., 2000. Open peer review: A randomised controlled trial. British Journal of Psychiatry 176, 47–51. https://doi.org/10.1192/bjp.176.1.47.
Wang, J., Veugelers, R., Stephan, P., 2017. Bias against novelty in science: A cautionary tale for users of bibliometric indicators. Research Policy 46, 1416–1436. https://doi.org/10.1016/j.respol.2017.06.006.
This is a working draft, published so that it can be challenged while it can still change. Comments and objections are welcome below.