The pressure becomes visible in three flagship systems. AlphaFold predicts protein structures within a field that already possesses established structural representations and experimental comparison practices. GNoME reorganizes an enormous space of candidate inorganic crystal structures and ranks objects that may later compete for experimental attention and funding. FunSearch combines a language model with an evaluator and an evolutionary search loop to produce executable programs that improve solutions to mathematical and algorithmic problems. The authors argue that the importance of these systems lies not only in model performance but in how each changes the moment at which scientists must decide what deserves testing, explanation, investment and recognition.

Discussions of these difficulties have often centered on model opacity, the problem that researchers may be unable to inspect a learned representation or trace how a particular output was produced. The authors accept that opacity matters, but contend that opening a model locally does not by itself determine what scientific status its output should receive. A feature attribution may identify influential inputs without showing that they correspond to a causal mechanism. A computationally stable candidate structure may be neither novel, synthesizable nor useful. An executable program may improve a mathematical bound without supplying the proof or principle that makes the improvement intelligible. The unresolved issue, they argue, lies between model output and scientific uptake, and the persuasive force of successful results can make it worse. Citing work by Messeri and Crockett on illusions of understanding, they warn that predictive coverage, visual fluency or numerical scale can be mistaken for evidence that a field has acquired explanatory control.

The paper is careful to deny that discovery before explanation is new. Empirical laws were applied before micro-level derivations existed, instruments stabilized phenomena that theory had not yet absorbed, and technical interventions worked before their mechanisms were agreed. The authors situate their argument within several existing literatures: data-anomaly typologies, such as Foorthuis’s classification of unusual patterns by data form and attribute relations; breakthrough typologies, including the question-driven versus object-driven distinction developed by Wuestman, Hoekman and Frenken in a qualitative analysis of 335 breakthrough papers; research on serendipity, notably Yaqub’s account of how unexpected findings become productive; and anomalistics, which examines methodological controls and alternative explanations for contested effects. These frameworks classify the form or genesis of a result, the authors argue, but they do not determine when a generated candidate has acquired adequate domain validation, become usable for inference, made its failures informative, and entered relations of communal criticism and responsibility. Those are post-generation questions.

To address that gap, the paper develops scientific maturity as a graded reporting framework built on four dimensions. Domain anchoring asks what constrains an output independently of the process that generated it, whether through experimental structures, synthesis and characterization, or formal proof and verification. Representational integration asks whether researchers can use the output to reason about a target, supporting inferences about binding sites, chemical structure types or proof strategies rather than merely repeating a model’s score. Failure articulation asks what a negative result changes, distinguishing genuine scientific correction from routine model tuning. Communal accountability asks whether uncertainty, provenance and responsibility can be followed across the distributed network of model builders, data curators, experimentalists and institutions. The dimensions are graded and field-relative, but the framework applies a non-compensatory rule: exceptional benchmark performance cannot substitute for a missing target relation, and widespread uptake cannot compensate for the inability to say what a failure means. Complete absence on any dimension blocks a strong claim of maturity.

Applied to AlphaFold, the framework shows a comparatively direct route to anchoring and integration, because a predicted three-dimensional structure enters a representational and experimental practice the field already knows how to handle. Yet the route remains incomplete. The authors cite Schuster’s argument that AlphaFold2’s complexity and opacity limit its capacity to contribute directly to scientific explanation of the protein-folding problem, even though its predictions can support object-level and, when embedded in further work, explanatory understanding of more specific questions in protein biology. Maturity, the authors stress, attaches to outputs in context rather than to a model once and for all. A high-confidence prediction for a well-represented protein family may enter downstream work with substantial support, while predictions involving disorder, dynamics, complexes or unfamiliar conditions may require far more extensive validation.

GNoME supplies the paper’s most contested case. Merchant and colleagues reported 2.2 million predicted crystal structures and identified 381,000 entries on a revised convex hull of computational stability. The authors argue that computational stability anchors only a limited candidate claim, while novelty, credible material identity and synthesizability require further evidence. Examining the claims from the perspective of experimental materials chemistry and crystallography, Cheetham and Seshadri reported scant evidence that sampled candidates simultaneously satisfied novelty, credibility and utility, identified many minor or implausible variants, and drew attention to organizational and symmetry problems in the released structures. Inspecting the first 250 entries in the GNoME Explorer compilation, they identified trivial substitutions within known structure families, chemically implausible compositions, and systematic lowering of symmetry produced by ordering elements that would more plausibly share crystallographic sites, alongside tens of thousands of entries involving highly radioactive elements whose practical status as materials is doubtful. The related A-Lab controversy sharpens the point: Leeman and colleagues re-examined 43 products reported as autonomously discovered materials and concluded that the analysis did not establish any as new materials, citing unreliable automated interpretation of powder diffraction and the neglect of compositional disorder.

FunSearch, by contrast, illustrates a more developed maturation pathway despite an opaque search process. Its outputs are executable, so researchers can inspect them, test them on cases, simplify them, compare them with existing constructions and seek a proof or general principle. Domain anchoring here is formal rather than empirical: a program’s performance can be verified within specified computational bounds, but a mathematical claim requires proof or another accepted form of formal warrant. The authors note that execution should not be confused with explanation, since a program may improve a bound while leaving researchers to discover why it works and whether its strategy transfers. They connect this to work by Davies and colleagues showing how learned patterns can direct mathematicians toward conjectures that are later interpreted and proved, and to Khosrowi and Finn’s question of whether generative AI can produce novel evidence. A FunSearch program can constitute new evidential material, they argue, but its authority derives from execution, proof, comparison and communal scrutiny, not from generation alone.

The authors draw practical implications for research governance. Performance reporting, they argue, should be paired with maturation reporting: a discovery system should state what target-level evidence constrains its outputs, what representations support domain inference, how negative results revise claims, and how responsibility is allocated across the pipeline. Candidate-generating systems may create a validation bottleneck rather than remove one, so funding and institutional design should support experimental confirmation, formal verification, domain interpretation, negative-result repositories and shared provenance alongside model development. The authors also propose an empirical research program: measuring maturation lag as the interval between public output claims and independent domain validation, tracing how representations are constructed around outputs, examining whether negative results alter scientific beliefs or disappear into model maintenance, and mapping who bears responsibility for data curation, candidate selection and correction. They acknowledge clear limits: three flagship systems from one organization cannot establish that the explanation crisis is routine across science, and the framework makes no claim about its prevalence. Its contribution is a four-question reporting schema that disciplines judgment rather than replacing it, keeping labels such as prediction, candidate, validated structure and discovery anchored to the evidence each actually carries.

Subject of Research: Epistemic framework for assessing AI-driven scientific discovery and the explanation crisis

Article Title: AI-driven scientific discovery requires scientific maturity beyond model opacity

Article References: Liu, Y., & Zhang, S. (2026). AI-driven scientific discovery requires scientific maturity beyond model opacity. Discover Artificial Intelligence, 6(1), Article 1633. https://doi.org/10.1007/s44163-026-02577-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02577-2

Keywords: AI-driven discovery, scientific maturity, model opacity, AlphaFold, GNoME, FunSearch, explanation crisis, domain anchoring, communal accountability, materials discovery, philosophy of science, epistemic integration