AI discovers "X"

One thing I suspect is that claims about AI “discovering” new science or algorithms involve far more human agency than most researchers recognize.

Take AlphaEvolve ¹ as an example. It is advertised as a coding agent capable of making novel scientific discoveries—for instance, finding a faster method to multiply 4×4 complex-valued matrices. This is presented as an autonomous breakthrough by AI. But in reality, the process depended heavily on human agency.

AlphaEvolve was not simply prompted as “please invent a better algorithm for matrix multiplication”. Instead, the problem was reformulated as a continuous optimization task, so that standard gradient-based methods could be applied. LLMs were then used to suggest hyperparameters (e.g., learning rate, weight decay), loss functions, and other heuristics. But this raises important questions that are often ignored: Why was this problem chosen? Why was it formulated in this particular way? The answer is that the authors drew on their understanding of LLMs and AlphaEvolve to identify problems where the system was most likely to succeed, and to frame those problems in a form it could handle. The paper even explicitly acknowledges that mathematicians (including Javier Gomez Serrano and Terence Tao) advised them on which problems to target and how best to formulate them:

“Most of these discoveries are on open problems suggested to us by external mathematicians Javier Gomez Serrano and Terence Tao, who also advised on how to best formulate them as inputs to AlphaEvolve.”

This invites comparison with earlier uses of computers in discovery. In 1976, Kenneth Appel and Wolfgang Haken used a computer to prove the four-colour conjecture; the result that any map can be coloured with only four colours such that no two adjacent regions share the same colour. Their program required hundreds of hours of computation, producing steps that no human could feasibly check in a lifetime. Yet the key to the proof was not an autonomous “discovery” by the computer. Rather, the human researchers played the key role by decomposing the problem into smaller pieces that a computer could handle, and then relying on brute-force computation.

So in both the cases (AlphaEvolve and the four-colour proof) the human role was central. The systems only produced results after researchers exercised agency in choosing which problems to tackle, how to represent them, and how to guide the computational process. Hence, computer-aided discovery is not new and LLM-based systems advance this trend in a continuous way. This leads to an amazing question: in current claims that “AI discovered X,” how should we understand the division of credit between humans and machines? Where along this continuum of computer-use in scientific-discovery do we draw the line of whether the discovery is made by the human or the computer?