Introduction↩︎

Artificial intelligence is a genuine game changer. Across medicine, engineering, education, and creative work, it is disrupting established practices at a pace and scale that few technologies in history have matched. As Pope Leo XIV’s encyclical Magnifica Humanitas [1] reminds us, this disruption is not inevitable in its consequences — it is a tool whose character depends entirely on the ends it serves and the conditions under which it is deployed.

Using AI to accelerate scientific discovery is, we believe, one of the most consequential and promising directions available to us. AlphaFold [2] transformed molecular biology in a matter of months. ML models have identified new antibiotics [3], [4], predicted material properties [5], and detected patterns in astronomical datasets that would have taken human researchers decades to find. LLMs have independently solved long-standing open problems [6] and disproved decades-old conjectures [7]. This is not hype — it is a demonstrated and expanding capability.

By AI-driven science, we mean more specifically the deployment of AI systems — in particular large language models and autonomous agents — as active participants in the scientific research cycle: formulating hypotheses, designing experiments, analyzing results, and proposing next steps, with varying degrees of human supervision. This is distinct from AI as a computational instrument, where the system executes a well-defined task within boundaries set by human researchers. The difference is not one of degree but of kind: in the first case, AI accelerates what humans do; in the second, AI begins to do what humans do. Pursued at scale, this transition amounts to what we call the industrialization of research — a shift from a craft model of scientific production, in which knowledge and judgment are embedded in the researcher, to a pipeline model, in which these steps are decomposed, automated, and supervised.

The industrialization of research, as AI-driven science proposes it, is not without promise — but it raises systemic questions that must be addressed in parallel with the deployment, not after it. Large-scale initiatives such as the Genesis Mission, which aims to double scientific productivity within a decade by placing AI agents at the center of the full research cycle, make these questions urgent. This essay addresses them, in the spirit of the Leiden Declaration on Artificial Intelligence and Mathematics [8], focusing on the consequences for scientific practice.

1 The Industrialization of Research↩︎

What AI-driven science proposes is, in the most precise sense of the term, the industrialization of research. The analogy with the industrial revolutions of the nineteenth and twentieth centuries is not rhetorical — it is structural¹. In both cases, a craft model of production gives way to a pipeline model: tasks that were previously integrated in a single practitioner or a small team are decomposed, standardized, and delegated to machines, while the human role shifts from execution to supervision. The craftsman who mastered the full arc of a trade — material, tool, judgment, and result — is replaced by the operator who monitors a system whose inner workings exceed any single person’s grasp.

In the craft model of science, a researcher identifies a question, designs an approach, executes experiments, interprets results, draws conclusions and present the work among its peers and to the society, exercising judgment at every step. Knowledge, method, and intuition are embedded in the person and transmitted through apprenticeship. In the pipeline model that AI-driven science proposes, these steps are decomposed and partially automated: agents handle hypothesis generation, experimental design, data collection, analysis, and validation, while human researchers supervise, validate, and orient the system. The result is a qualitative change in the relationship between the researcher and the scientific process — and, potentially, in the nature of the knowledge produced.

The parallel with industrial history is very instructive. The industrial revolution produced real and substantial benefits — gains in productivity, speed, and scale that no craft system could have matched — alongside real and substantial costs that were systematically ignored at the moment of deployment: environmental externalities, concentration of capital, loss of craft knowledge, and asymmetric access between those who owned the infrastructure and those who supplied it with labor. The same structure of benefits and costs appears to be reproducing itself in the industrialization of research, as Table 1 illustrates. What makes the current transition particularly significant is not that it breaks the parallel, but that it deepens it: the industrial revolution automated physical labor while leaving intellectual labor intact; the industrialization of research extends that same logic to the mind itself. Science — one of the most distinctly human of cognitive practices — becomes the next domain to be organized around machines, workflows, and supervisors rather than craftsmen.

Table 1 draws the parallel with industrial history, but not all dimensions are treated equally in this essay, which focuses specifically on the impact of AI-driven science on how research is conducted, transmitted, and evaluated. The advantages of AI-driven science and the constitutive reasons that motivate its development are examined in Section 2. The more complex consequences — for competence transmission, scientific comprehension, paradigm shifts, agenda capture, and access equity — are addressed in Sections 4 to 8. Two dimensions specific to AI-driven science that have no direct industrial precedent receive dedicated treatment: the collapse of research evaluation under AI-generated volume (Section 9), and the compounding of errors in closed-loop pipelines (Section 10). The broader societal and environmental impacts of large-scale AI deployment — energy costs, democratic governance, concentration of economic power — are noted where they bear directly on scientific practice but deserve separate treatment beyond the scope of this essay.

2 Why AI-Driven Science?↩︎

Before examining the fundamental questions that large-scale AI-driven science raises, it is worth asking a prior question: why pursue it in the first place? The answer is not self-evident, and it matters — because the strength of the case for AI-driven science is precisely what makes the concerns that follow worth taking seriously.

The term “AI for science” covers two qualitatively different things, and conflating them creates confusion.

The first is AI as an algorithmic tool — a system designed and trained by humans to solve a specific, well-defined class of problems. AlphaFold predicts protein structures from amino acid sequences [2]. Graph neural networks reconstruct particle collisions at the LHC [10]. GNoME identifies stable crystal candidates from materials databases [5]. In each case, the architecture, the training objective, and the problem framing are human choices; the AI executes within those boundaries with extraordinary efficiency. This is a authentic and important contribution — but it is, in nature, an advanced instrument, not an independent scientific actor.

The second is AI as a (semi-)autonomous research assistant — a general-purpose system given an open problem, which explores, reformulates, and arrives at a solution through its own iterative reasoning, with minimal human guidance. Two recent results illustrate this mode. In February 2026, Donald Knuth described how Claude Opus 4.6 solved in a few hours and some thirty iterative explorations a combinatorics problem he had been working on for several weeks [6]. In May 2026, an internal OpenAI model disproved an 80-year-old conjecture by Erdős on unit distances in the plane — one of the best-known open problems in combinatorial geometry — by producing a family of counterexamples using sophisticated tools from algebraic number theory that no one had previously thought to apply.

This essay is concerned primarily with the second mode, and with what happens when it is deployed at the scale of national and international science infrastructure. The motivations are real and deserve to be stated clearly. As shown above, difficult scientific results can already be achieved semi-autonomously by LLMs — and this is no accident. These results follow from intrinsic properties that make LLMs particularly well-suited to scientific research: they are trained on a corpus that spans the entirety of human scientific knowledge and the full conceptual vocabulary of all disciplines simultaneously, making them natural instruments for interdisciplinary research because the connections between fields are already embedded in their representations. Beyond interdisciplinarity, AI systems can explore hypothesis spaces in parallel at speeds no team can match; they can read and ingest new results efficiently and rapidly; they are free from the confirmation biases that lead entire communities to miss promising directions; they can accelerate the full cycle from hypothesis to result in ways that compound across iterations; and their capacity for parallel deployment allows a scale of scientific production that no human workforce could sustain.

These are serious advantages. They justify serious investment, and they explain why initiatives like the Genesis Mission are being pursued with urgency. The question is not whether these advantages are real. The question is what else comes with them.

3 The Genesis Mission: A Case Study in AI-Driven Science↩︎

Announced by the United States Department of Energy in late 2025, the Genesis Mission is one of the most ambitious AI-driven scientific initiatives to date — a national effort to build an integrated AI-for-science infrastructure connecting supercomputers, AI systems, and quantum technologies across the DOE national laboratories [11]. Rick Stevens, Associate Laboratory Director at Argonne and one of the mission’s principal architects [12], has compared its ambition to the Apollo program and its urgency to the Manhattan Project [13]. The stated goal is to double scientific productivity within ten years.

What makes Genesis qualitatively different from previous investments in scientific computing is the role envisioned for AI. These agents are not passive tools — they are active participants in the research cycle, formulating hypotheses, designing experiments, interpreting results, and proposing next steps in closed-loop automated workflows. Genesis proposes to consolidate and scale that model across all domains — spanning energy, materials, biotechnology, nuclear security, quantum computing, and grid modernization [14].

The scale of the undertaking is matched by its strategic intent. Genesis is not primarily a scientific decision — it is a geopolitical one, shaped by the explicit objective of maintaining US technological leadership in the context of intensifying global competition, particularly with China [15], [16], while collaborating with selected partners such as Japan [16]. This means the time horizons and priorities driving the mission are not those of the scientific community, but of national security and industrial competitiveness. These tensions are not unique to Genesis — they emerge whenever AI is deployed at the heart of the scientific enterprise. The concerns raised in the sections that follow therefore extend beyond the Genesis Mission: they concern what AI-driven science does to knowledge, to the people who produce it, and to the institutions through which it is shared and evaluated. Genesis is the most visible and ambitious instantiation of AI-driven science today — but the questions it raises are structural.

4 Competence, Training, and the Long-Term Investment Model↩︎

Human-led research rests on a specific investment logic. We train young people through years of undergraduate study,graduate programs, and doctoral research. We select among them — for talent, motivation, and fit with institutional priorities. The investment is substantial and slow to pay off, but it compounds over time: a researcher trained in their twenties will still be producing knowledge, mentoring others, and shaping the direction of a field thirty years later. The return on that investment is not just the papers they publish, but the living competence they carry and transmit.

This model has an important inherent feature: the transmission of knowledge is embedded in the people themselves. A PhD supervisor does not simply transfer information to a student — they transmit intuitions, habits of mind, a sense of what questions are worth asking, a feel for when results are trustworthy, and a set of community norms about what counts as good science. This kind of transmission cannot be fully documented or formalized. It exists in the relationship between minds engaged in a common practice over time.

If scientific work is increasingly delegated to AI systems, this investment model changes fundamentally. The primary capital is no longer human — it is machines, infrastructure, and model weights. And these assets have a radically different economic profile. Hardware becomes obsolete in three to five years. Models must be retrained as knowledge advances. The pressure to amortize these investments quickly pushes toward intensive deployment, not patient cultivation. Once a system is obsolete, it is discarded. It does not, by itself, sustain an intellectual community, mentor junior scientists, or transmit the tacit dimensions of scientific practice. One might argue that AI systems help train future AI systems — and this is true in a narrow technical sense — but this is a much weaker and more attenuated form of transmission than the human cycle of student \(\to\) researcher \(\to\) mentor \(\to\) student.

The risk, then, is not merely that fewer human scientists will be trained — it is that the incentive to train them will gradually erode. If an AI system can do in hours what a PhD student does in three years, the rational institutional response may be to hire fewer PhD students and deploy more compute. In the short term, output increases. In the medium term, the pipeline of trained scientists shrinks. In the long term, the capacity to understand, evaluate, and correct the AI systems themselves — which requires deep domain expertise — begins to atrophy.

This concern does not argue against AI-assisted science as such. But it argues for actively thinking through the consequences of this shift before the system has been fully deployed, rather than discovering them afterward when the pipeline of trained scientists has already been allowed to thin.

5 The Problem of Scientific Comprehension↩︎

There is a deeper issue that concerns not the transmission of competence but the comprehension of results. Science, at its best, does not only solve problems — it produces understanding. It gives us models of the world that are interpretable by human minds: theories that explain why things behave as they do, not just how to predict or control them. This distinction matters enormously.

Consider an analogy from chess. The best chess engines in the world — Stockfish, AlphaZero — play at a level that no human can match. But even the strongest grandmasters cannot always explain why an engine chooses one move over another in complex positions. The engine’s “reasoning” is opaque, distributed across millions of parameters in ways that do not map onto the conceptual vocabulary of human chess understanding. We trust the move because the engine wins, not because we understand it.

If scientific theories are increasingly generated by AI systems, we may find ourselves in an analogous situation: in possession of models that predict experimental outcomes with extraordinary accuracy, but that no human mind can truly comprehend. We would have science in the instrumental sense — a powerful tool for solving problems — but we might lose science in the epistemic sense: a human-intelligible account of why the world is the way it is.

This would be a profound change in the nature of the scientific enterprise, and it is not obviously benign. Human-interpretable theories have value beyond their predictive power. They allow scientists to identify errors, generate novel hypotheses by analogy, explain results to non-specialists, and connect findings across disciplines. If theories become opaque, these capacities are compromised. We would then have entered a form of intellectual dependency from which there is no obvious exit: we might need AI not only to generate theories, but to explain them.

One might object that theories have always been technically complex, and that popularization has always been an approximation. This is not entirely wrong. But the risk here is of a different kind: not that AI models are uninterpretable (they are, but that is a known limitation), but that the theories and conceptual frameworks they generate may themselves become inaccessible — correct, perhaps, in the sense that they produce accurate predictions, but not grounded in the kind of human-legible reasoning that allows scientists to argue about them, extend them, or know when they break. A theory that no human derived, built on intermediate representations no human chose, validated by systems no human fully oversaw, may be true in a narrow empirical sense while remaining fundamentally outside the reach of human scientific culture.

6 Incremental Progress Versus Paradigm Shifts↩︎

A deeper concern is epistemological, and perhaps the most difficult to resolve. It is the question of whether AI is capable of real breakthrough science — the kind of discovery that does not just extend what is known, but reframes the entire conceptual landscape.

The history of science is punctuated by moments of this kind: Newton’s laws, Darwin’s theory of evolution, Maxwell’s equations, Mendel’s genetics, Einstein’s relativity, Watson and Crick’s double helix, plate tectonics. What these discoveries have in common is that they were not, in any straightforward sense, the inevitable output of accumulating data. They required a leap — an act of conceptual reframing that reorganized existing observations under a new and unexpected framework. The data was often available before the theory; what was missing was the insight.

It remains an open and genuinely difficult question whether AI systems are capable of this kind of leap. Current large language models are extraordinarily good at pattern recognition, interpolation, and sophisticated recombination of existing knowledge. They can find correlations in large datasets, generate plausible hypotheses within established frameworks, and optimize known solutions. But it is much less clear that they can do what Einstein did in 1905: start from a profound dissatisfaction with the internal consistency of existing theory, and arrive at a framework so different that it requires abandoning deeply held intuitions about space and time.

A useful thought experiment — a kind of scientific Turing test — would be: give an AI all scientific data available up to 1905, and ask whether it can discover special relativity. This is not a frivolous exercise. It asks whether AI can reproduce the actual structure of a paradigm shift: not a good prediction, not a clever optimization, but a conceptual revolution. The core objection is this: LLMs are stochastic interpolation machines — however sophisticated, they recombine what they have been trained on, and special relativity required abandoning, not recombining, the conceptual framework available in 1905. Whether current systems are capable of that kind of leap remains an open question, and the evidence so far does not give strong grounds for confidence.

This uncertainty matters for how we structure AI-driven science. If AI is primarily useful for incremental progress — for exploring design spaces within established frameworks, for automating the repetitive aspects of data analysis, for finding patterns that human researchers would eventually have found but more slowly — then the AI-for-science vision makes sense as an accelerator of normal science. But if we expect AI to substitute for human judgment in identifying which frameworks to pursue, which anomalies are worth taking seriously, and when a field needs to abandon its assumptions and start over, we may be attributing to these systems capacities they do not yet possess.

The risk, in that case, is not that AI-driven science produces nothing valuable — it will produce a great deal of value — but that it produces a flood of incremental results while true paradigm shifts become rarer, precisely because the systems driving research are optimized to work within existing frameworks rather than to challenge them.

7 The Capture of the Scientific Agenda↩︎

Among all the concerns raised in this essay, one targets something with no equivalent in previous technological transformations of science: the question of who defines the questions.

Science benefits from a property that is easy to overlook precisely because it has always been there — its agenda is distributed and heterogeneous. Funding agencies, governments, and disciplinary fashions all exert real pressure, but underneath these forces thousands of researchers across hundreds of institutions are following their own curiosity, pursuing angles no committee has endorsed. The result is a global cognitive ecosystem whose diversity is messy and inefficient by many measures, but capable of producing surprises no planned system could have anticipated. Penicillin was not in a roadmap. Neither was the cosmic microwave background, the discovery of prions, or the link between H. pylori and peptic ulcers. These findings emerged from researchers who were, in one sense or another, working on the wrong question at the wrong time.

Genesis challenges this structure directly. An infrastructure of this scale — tens of billions of dollars, 17 national laboratories, partnerships with Nvidia and OpenAI, an explicit geopolitical objective — does not emerge from the distributed preferences of the scientific community. It emerges from political and industrial decisions made by a small number of actors with specific interests and specific blind spots. The agenda Genesis encodes is therefore not the agenda of science; it is the agenda of the institutions that built it, at the moment they built it.

The consequences are already visible. Genesis is oriented toward energy, materials, biology, and defense-relevant physics. Climate science is not among them — not by accident. Genesis was designed under a federal administration that has systematically downgraded climate research, withdrawn from international climate commitments, and reoriented the DOE away from climate-related work. The most powerful scientific infrastructure ever built is thus being deployed with a deliberate blind spot on one of the most urgent questions facing humanity. The irony is structural: a system projected to consume the energy equivalent of dozens of nuclear reactors will accelerate scientific production across many domains, but not the one most directly relevant to the consequences of its own existence.

The deeper risk is not this single omission but the gravitational logic it exemplifies. Once large-scale AI science infrastructure becomes the dominant mode of scientific production, the questions it does not ask become progressively harder to ask — not by prohibition, but because research outside its agenda will be slower, less visible, and less fundable. Capture operates at two levels simultaneously: at the political level, priorities shift with administrations, but the physical infrastructure they build persists and constrains what comes after; at the industrial level, private partners have a direct interest in orienting science toward what is computationally tractable, economically valorizable, and strategically legible. Neither requires a conspiracy. Both compress the freedom to ask inconvenient questions — and that freedom, imperfect and embattled as it has always been, is precisely what has made science capable of surprising the people who fund it.

8 The Structural Bifurcation of Science↩︎

The standard framing of unequal access to AI infrastructure is a “multi-speed science” — some institutions move faster, others slower, but all in the same direction. What large-scale AI science initiatives are likely to produce is not a speed gap but an organizational bifurcation: two forms of scientific practice that become progressively less commensurable, losing the shared language that makes critical evaluation possible between them.

The asymmetry operates in both directions at once, and one direction has received almost no attention. Researchers outside advanced AI infrastructure will continue to publish — and their results will be immediately ingested by the systems they cannot access. Large-scale AI science platforms are designed to absorb the entire corpus of published science. Researchers outside such systems contribute automatically to the productivity of systems that offer nothing in return. Their papers are ingested, their datasets absorbed, their findings used to train the next generation of models — while they remain unable to access, replicate, or even fully read the science those models produce in return. This is a form of epistemic extraction with no real precedent: previous resource asymmetries gave richer institutions better tools, but they did not involve the systematic harvesting of intellectual output from less-resourced institutions to power the engines of more-resourced ones.

The downstream consequences compound every concern raised earlier in this essay. Peer review loses its foundation when authors and reviewers no longer work in the same epistemic environment. Brain drain accelerates toward the handful of institutions with advanced AI infrastructure — precisely where the human researcher is most at risk of being replaced. And the career logic of science — build expertise, publish, contribute to a cumulative conversation — breaks down when one side of the conversation operates at a speed and scale the other cannot follow.

What emerges is not a global acceleration of science but its fragmentation into two largely non-communicating practices: one fast, opaque, and institutionally concentrated; the other human-scaled and increasingly marginalized. The danger is not that the second disappears, but that it loses the ability to critically engage with, challenge, or correct the first — and with it, science loses the distributed skepticism that is its only reliable defense against its own errors.

9 The Collapse of Evaluation↩︎

The concern that follows is more institutional, but no less serious. The scientific community regulates the quality of its output through a system of peer review, editorial judgment, replication, and collective criticism. This system is imperfect — it is slow, prone to bias, and has never fully solved the problem of reproducibility. But it rests on a crucial assumption: that there are enough qualified human experts to read, evaluate, and challenge the work being produced.

If AI systems generate scientific papers at scale — and this is not a speculative scenario, it is already happening — this assumption collapses. No community of human reviewers can evaluate ten times the current volume of publications. The response might be to use AI to review AI-generated papers, which simply displaces the problem: who evaluates the reviewers? We could end up with a system in which papers are generated by AI, reviewed by AI, accepted or rejected by AI, and cited by AI — a closed loop that is productive in a quantitative sense but increasingly disconnected from any real community of human scientific judgment.

History offers little comfort here. Science has always published more than it should, and the volume of publications has grown continuously for over a century. There is no mechanism by which a surge in output naturally leads to a corresponding improvement in quality. The more likely outcome is more noise — more incremental results, more papers that advance no one’s understanding, more difficulty identifying the work that actually matters.

A related challenge concerns the evaluation of researchers themselves. Academic careers rest on both quantitative signals (publication counts, citation indices) and qualitative judgment (reading the work, interviewing the candidate, assessing whether the researcher can explain and defend their ideas). AI-driven science destabilizes both: fifty AI-assisted papers a year tells you nothing about scientific quality, and “walk me through how you arrived at this result” becomes truly harder to answer — and to evaluate — when the result was produced by an agent the researcher supervised rather than derived themselves. The boundary between human and machine contribution is blurred in ways that good-faith panels will struggle to navigate — and this applies to promotion as much as to publication.

Hiring raises a distinct but related problem, and it cuts in both directions. A researcher who excels at directing large-scale AI workflows may be lost without that infrastructure — or with a different version of it in five years. The reverse is equally true. This creates a specific challenge for early-career hiring: how do we evaluate a PhD student or postdoc whose scientific ability was formed in an AI-driven research environment? And conversely, how do we assess a traditionally trained candidate’s potential in a world they have never worked in? We do not yet have answers to these questions — and the decisions being made in hiring committees today will shape the scientific workforce for the next thirty years.

This does not mean the system will simply break. Scientists and institutions will adapt. But we should be honest about the scale of the challenge. The entire apparatus for evaluating research and researchers — metrics, journals, grant review, promotion panels — was designed for a world in which scientific output was produced at human speed by human minds. That world is ending. What replaces it needs to be designed deliberately, not improvised under pressure.

10 The Compounding Error Problem↩︎

The scientific method is, at its core, a machine for catching mistakes. Its self-correcting property rests on one intrinsic requirement: that the agents producing results and the agents checking them are largely independent — different equipment, different training, different institutional incentives. Errors in a human pipeline are largely uncorrelated, and their tension surfaces flags.

A closed-loop AI pipeline breaks this independence at every stage simultaneously. Hypothesis generation, experimental design, data collection, analysis, and validation are all handled by agents sharing the same foundation models, the same training data, the same systematic blind spots. A bias that enters at stage one does not encounter a genuinely independent check at stage three — it encounters a system that shares its assumptions. The errors do not cancel; they compound silently across the full cycle.

This is the AI analogue of the replication crisis, but fundamentally worse. In psychology or medicine, shared methodological assumptions eventually met the resistance of reality and independent replication. In a large-scale AI science pipeline, both corrective mechanisms are weakened simultaneously. Volume makes independent human replication impossible for more than a fraction of results. And AI-to-AI replication, if it shares the same underlying architecture, produces not an independent test but an echo — the form of verification without its substance.

The further danger is reingestion. AI systems are retrained as new results accumulate. A flawed result absorbed into the next generation’s training corpus does not remain an isolated error — it becomes a prior, shaping subsequent hypotheses and experimental designs. The error is institutionalized rather than corrected, and the pipeline drifts silently away from empirical reality with no single step appearing to cause it.

The deepest version of this problem is the most basic: who reads the raw output of physical reality? In all science, the ultimate check is not peer review but contact with nature — the spectrum, the microscope image, the detector signal, the patient outcome that refuses to match the prediction. The more the pipeline is automated, the more mediated this contact becomes. The question of who verifies the verifiers, in a system optimized above all for speed, is not a technical detail. It is the central epistemological challenge of AI-driven science, and it has not yet received an adequate answer.

Conclusion: Transformation, Not Abolition — But Not Without Deliberation↩︎

What AI-driven science proposes is, in the most precise sense of the term, the industrialization of research — with all that industrialization has historically brought and may yet bring: extraordinary gains in productivity, reach, and cross-domain recombination, alongside concentration of capital, standardization of output, and the gradual erosion of forms of knowledge that do not fit the new mode of production. Previous industrial revolutions were not stopped — but their consequences were shaped by the choices made during their deployment, and those choices are what this essay has tried to map.

The concerns raised here do not lead to a single conclusion. Some will read them as reasons to question AI-driven science discovery altogether — of which Genesis is one prominent example. The position taken here is different: the central question is not whether to pursue AI-driven science, but under what conditions it will deliver on its promises. What happens to the intergenerational transmission of scientific competence that no infrastructure can replace? What kind of understanding do we preserve when theories become too opaque for human minds to interpret? Can AI produce the paradigm-shifting discoveries that have defined science at its most transformative — or only accelerate the incremental? Who controls the questions being asked — as the absence of climate science from Genesis’s priorities already shows? And what happens to the researchers whose work is absorbed by systems they cannot access, contributing to a productivity they will never benefit from? How do we evaluate quality — and researchers — when quantity explodes and AI reviews AI? Who catches the errors when the pipeline that produces results and the pipeline that validates them share the same systematic biases? These are not peripheral details. They are the difference between an infrastructure that transforms science and one that merely accelerates its worst tendencies.

These are not rhetorical questions designed to slow down progress. They are genuine open questions, and the answers will shape the scientific culture that emerges on the other side of this transition. A science that is faster and more productive, but less comprehensible, less well-evaluated, less capable of training the next generation, and structurally fragmented between a well-resourced AI tier and everyone else — is not straightforwardly better than what we have now. Getting this right requires not just engineering ambition, but a serious, ongoing conversation about what science is for — and what we do not want to lose in the process of accelerating it.

Acknowledgment↩︎

The ideas, arguments, and positions developed in this essay are entirely my own. I used Claude Sonnet 4.6 (Anthropic) as a writing assistant to help structure, draft, and refine the text. I remain solely responsible for the content, including any errors, omissions, or debatable claims it may contain.

References↩︎