Can artificial intelligence ever truly understand science?

The authors argue that AI can take part in "android-assisted scientific understanding" in three ways: as a "computational microscope" that reveals properties of a system that are "otherwise difficult or even impossible to probe", as a muse supplying "new concepts and ideas that are subsequently understood and generalized by human scientists", and as an agent that gains understanding itself. In the first two, humans do the understanding. For the third they write that "we have not found any evidence of computers acting as true agents of understanding in science yet." The framing starts from an oracle that "correctly predicts" every experiment and still leaves scientists unsatisfied, because they "want to comprehend how the oracle conceived these predictions."

The criterion comes from the contextual theory of Henk de Regt and Dennis Dieks: a phenomenon can be understood if scientists "can recognise qualitatively characteristic consequences of T without performing exact calculations", where T is an intelligible theory of that phenomenon. On this view understanding sits in the scientist's ability to reason from a theory, not in the accuracy of predictions. That is what sorts the dimensions. In the first two, "the android enables humans to gain new scientific understanding"; in the third, "the machine gains understanding itself." The authors also say they surveyed dozens of scientists and collected dozens of anecdotes, but the excerpt does not reproduce them.

Set against the nearest notes, this is a graded account of understanding that differs from the mechanistic one. The mechanistic-interpretability note grades an LLM's internal organization into three tiers; this source grades the machine's role relative to a human scientist, so the two answer different questions and can be read side by side. The oracle is the case the heuristics note describes: a transformer that predicts orbital trajectories accurately without a Newtonian model is the kind of oracle the authors say a scientist would not accept. The potemkin note is a contrast. It shows that a correct explanation is not enough for understanding, while this source supplies a positive test, recognizing consequences without calculation, that the potemkin pattern can be checked against. On the source's terms, the output of the automated Mechanist instrument belongs to the first dimension, since "humans then lift these insights to scientific understanding."

The excerpt establishes nothing as verified. It uses "correct" for the oracle's predictions and "understanding" for grasping how they arise, and it describes no check by a proof assistant, an evaluator or a panel of experts. The three dimensions are argued from the philosophy of science and from anecdotes that are not shown, and the agent dimension is, in the authors' own framing, a roadmap for which they have no evidence. The implication is modest. The taxonomy is best read as a research agenda and as a way to ask, of any claimed AI result, which dimension it belongs to and who does the understanding, not as a finding that any existing system understands.