The Epistemic Risks of AI-Only Science
I work on machine learning for science, and I believe deeply in its transformative potential. From predicting protein structures (Abramson et al. 2024) to modeling climate systems (Bodnar et al. 2025; Allen et al. 2025), AI is accelerating research in ways that seemed implausible just a few years ago. The potential is genuinely remarkable, and I’m not here to argue otherwise.
But there’s something troubling about the current trajectory toward fully autonomous “AI scientists”—systems designed to independently formulate hypotheses, design experiments, and draw conclusions. While these systems will certainly produce impressive results, their widespread adoption might risk something more subtle and perhaps more dangerous: the gradual erosion of epistemic diversity that makes science robust.
On Value Lock-In and Scientific Monocultures
Value lock-in describes what happens when systems become so entrenched that they perpetuate specific ways of thinking, making alternatives prohibitively difficult to pursue.1
Science faces an analogous risk. Once AI scientists become the dominant research paradigm—and given their speed and cost advantages, this seems likely—they won’t simply reflect current scientific values. They’ll crystallize them. What appears today as one approach among many could become the only viable approach, not through conscious choice but through path dependence.
The concern isn’t that AI scientists will be poorly designed, but rather that they’ll be too successful at optimizing for current definitions of scientific progress. This creates the “lock-in effect”: the more we invest in AI-driven research infrastructure, the harder it becomes to pursue alternatives, regardless of their potential merit.
The Myth of Value-Neutral Science
Science has never been value-neutral, though we often pretend otherwise. Every research program embeds choices about what questions matter, what methods are legitimate, what constitutes adequate evidence (Huff 2017; Bronowski 1961).
Take computer vision’s relationship with ImageNet. This single benchmark, created by a particular research community with specific assumptions about visual recognition, shaped an entire field for over a decade. It privileged approaches that performed well on a large, static, pre-labeled photographs (Dotan and Milli 2019).
These kind of value-laden choices are the along the entire chain of (AI) research (Dehghani et al. 2021; Hooker 2020; Alampara, Schilling-Wilhelmi, and Jablonka 2025).
Concentration and Its Discontents
Contemporary AI scientists emerge from a remarkably homogeneous ecosystem: similar academic backgrounds, shared technical assumptions, a handful of dominant companies providing the underlying infrastructure. This concentration creates troubling dynamics (Paul 2019; Crawford 2021).
First, epistemic monoculture becomes increasingly likely. When AI systems trained on similar data with comparable objectives dominate research, they’ll systematically favor certain types of questions. Approaches that don’t translate well into current AI paradigms—perhaps because they rely on tacit knowledge, or require forms of reasoning that resist formalization—risk being dismissed as unscientific rather than simply incompatible with our current tools.
Second, we face the prospect of algorithmic gatekeeping. As AI scientists become more productive, human researchers will encounter mounting pressure to adopt these tools or become irrelevant (and performing “current science” as a form of “art”). A small number of AI platforms could effectively determine which ideas get explored and which get ignored—not through explicit censorship, but through the subtler mechanism of making alternatives economically unviable: Power tends to concentrate, and scientific institutions aren’t immune to this tendency.
Normal Science and Revolutionary Potential
According to Thomas Kuhn’s philosophy of science, science alternates between periods of “normal science”—where researchers work productively within established paradigms—and revolutionary episodes that fundamentally reframe entire fields (Kuhn 1962). Crucially, Kuhn observed that normal science “often suppresses fundamental novelties because those novelties are necessarily subversive of its basic commitments.”
AI scientists, optimized on existing scientific literature, might end up being just sophisticated normal science machines. They could excel at incremental advances within current paradigms but struggle with the radical reconceptualization that drives scientific revolutions. This is because current paradigms of machine learning systems easily question their foundational assumptions because those assumptions are embedded in their training data and optimization targets . It’s an inevitable consequence of how these systems work: They are trained to maximize the likelihood of the training data. But it suggests that a science dominated by AI scientists might become extraordinarily good at certain kinds of progress while systematically failing at others.
Toward Epistemic Pluralism
None of this constitutes an argument against AI in science, which would be both futile and counterproductive. Rather, it’s a case for what we might call epistemic pluralism: the deliberate maintenance of diverse approaches to scientific inquiry.
We need something like a portfolio approach to scientific methodology. Some research should leverage AI scientists for their remarkable speed and scale. Other investigations should preserve space for human-led inquiry. Still other work should explore hybrid approaches that combine artificial and human intelligence in novel configurations.
Monocultures are efficient under stable conditions but catastrophically vulnerable to unexpected challenges. Diverse ecosystems sacrifice some efficiency for resilience. Given the stakes involved in scientific knowledge production, resilience seems worth prioritizing.
The robustness of scientific knowledge depends not on any single approach, however sophisticated, but on the productive tension between multiple ways of understanding the world. Preserving that tension, even as AI transforms scientific practice, may be one of the most important challenges facing the scientific community today.
References
Footnotes
Consider American urban development: what began as apparently rational choices about highways and suburbs eventually made walkable neighborhoods economically unfeasible to build. The infrastructure didn’t merely reflect certain values—it enforced them, long after those values might have been questioned. Similarly, the economic investments in American slavery created powerful incentives to maintain the system—once millions of dollars were invested in enslaved people as “property,” with entire industries and political systems built around protecting those investments, the institution became extremely resistant to change even as moral opposition grew. (MacAskill 2022)↩︎