A recent article discusses the application of AI to particle physics and repeats the familiar suggestion that what is needed is an 'AlphaGo for physics,' understood as a machine capable of generating genuinely novel insights. AI systems are indeed effective at tasks such as equation solving and computational optimization. However, foundational physics is not reducible to a sequence of well-posed problems governed by fixed rules, as is the case in domains such as protein folding. The central challenge in foundational physics is not merely to solve problems within a given formalism, but to identify the appropriate formalism itself, and to recognize when the prevailing rules are inadequate. This requires the capacity to interrogate and revise the underlying metaphysical assumptions.

The process of identifying the correct rules in physics is not equivalent to pattern recognition over data. It also involves recognizing when existing conceptual frameworks fail to account for empirical phenomena, and systematically isolating the implicit assumptions or metaphysical presuppositions embedded within those frameworks. This process often requires discarding elements that, while superficially well-integrated, lack empirical justification. Such work is not purely algorithmic; it involves a form of qualitative judgment that cannot be reduced to data-driven inference. I doubt if AI can ever perform such qualitative reasoning.

Einstein's 1905 work on relativity illustrates this point. The essential advance was not a matter of computational proficiency, but of formulating new foundational postulates and discarding entrenched but unnecessary assumptions, such as the aether. An AI system, if it had existed at the time, might have been able to perform the requisite symbolic manipulations once the problem was formally specified, but it would not have generated the conceptual shift itself. The distinction is between executing a prescribed procedure and originating the framework within which such procedures acquire meaning.

In addition to the distinction between qualitative and data-driven reasoning, physics involves both the exploration of questions within a defined research program and the creation or redefinition of that program. The research program determines which questions are pursued, what counts as an explanans and an explanandum, what counts as a solution, and which mathematical structures are relevant. The discussion in the article appears to overlook not only the importance of qualitative reasoning, but also the possibility of redefining existing research programs. The capacity to redefine a research program is, as far as I can determine, not something that current AI systems possess.

The article reflects a methodological orientation in which the discovery of new physics is equated with the identification of new particles, typically by positing speculative micro-dynamics and attempting to extend the Standard Model through additional Lagrangian terms. There is, however, no guarantee that nature conforms to this approach, and it is possible that such strategies may not yield further insight. In this context, the application of AI is limited to operating within the existing methodological framework, optimizing within a set of assumptions whose validity may itself be in question.