Can opaque models guide discovery without needing interpretation?

The paper's claim is that the pessimism in philosophy about opaque deep learning comes from examining the wrong stage of research. Opacity's epistemological concerns "will arise chiefly when network outputs are treated as scientific claims that stand in need of justification." Used as parts of a discovery process, outputs "guide attention and scientific intuition toward more promising hypotheses but do not, themselves, stand in need of justification," and "the mere inductive support DLMs provide is epistemically sufficient to guide pursuit." Two cases are offered: deep learning guiding mathematical intuition about relations between classes of knot properties in low-dimensional topology, and a fully opaque model whose predictions led to a revised theory of aftershock dynamics in geophysics.

The mechanism is a division of labor between generating a hypothesis and justifying it. The network sits beside abduction and problem-solving heuristics and faces only "preliminary appraisal"; justification falls on the product. The revised earthquake theory is judged by the discipline's own tests: it is "consistent with first principles," "aids in the explanation and understanding of aftershock dynamics," and "outperforms extant theory in prediction." Because those checks apply to the theory, the author calls it "epistemically irrelevant" whether the network represents the geophysical quantities it flagged. The author grants that the saliency analysis might count as an interpretive step and calls that objection "well taken," but argues it does not change the outcome. The understanding the excerpt credits belongs to the theory, not the network, and verification is likewise applied to the theory, not to the network's outputs.

This is the sharpest contrast with the nearest notes on what opacity hides. The Do language models understand in fundamentally different ways? note treats understanding as an internal achievement; this paper says scientific payoff can arrive without it. The findings on heuristics and masked structure are not overturned but sidestepped. A model that predicts accurately without a world model (Do foundation models learn world models or task-specific shortcuts?) fits the discovery role, since the bar is inductive support, not a general law. Biased representation analysis (Do standard analysis methods hide nonlinear features in neural networks?) and fractured internals behind identical performance (Can identical outputs hide broken internal representations?) matter most when outputs are taken as findings. On the paper's account, such failures would cost a wasted hypothesis rather than a false justified claim, an inference the excerpt does not test.

The excerpt does not establish how strong either case is. Each is summarized in a few sentences, with no model, data or accuracy figures, and the Section 4 detail the argument points to is absent. It does not say whether the knot relationship was later proven. The claim that discovery outputs need only inductive support is asserted rather than measured. The paper's own limit is explicit: problems arise when outputs are treated "as findings in their own right," which "only network transparency can provide." The supportable implication is narrow. Opaque models can steer hypotheses where independent checks follow, and the excerpt does not show that opacity is harmless when results are taken at face value.