Guest and Martin question the logic of inferring brains from ANNs

A 2023 paper in Computational Brain & Behavior by Olivia Guest and Andrea E. Martin is doing quiet rounds again, and it is worth sitting with because it targets a habit that has crept into a lot of AI-and-the-brain talk. The habit is this: an artificial neural network does well on a task humans do well on, and that resemblance gets treated as evidence the model tells us something about the brain. Guest and Martin argue that inference does not hold up when you write it out carefully.

The move they make in the paper is to formalise how researchers actually argue. They propose what they call a metatheoretical calculus, and they express claims from the published record about models' successes and failures in first-order logic. As the Max Planck Institute's summary puts it, their formalisation describes the decision-making processes scientists use to adjudicate over theories. Writing the reasoning out that way, they say, uncovers potential deep issues in how theory is related to phenomena.

Why this matters outside cognitive science: a lot of the public narrative around large models leans on the same slippage. A system matches a human score on a benchmark, and the language quickly slides from 'it can do the task' to 'it does the task the way we do'. Guest and Martin's point, applied to that setting, is that similarity of output does not license similarity of mechanism, and if your research programme or your product story depends on the second claim, you owe an argument for it, not a leaderboard.

The honest caveat is that this is a philosophy-of-science paper, not an empirical takedown of any specific model, and the retrieved summaries do not name which published claims the authors most want to unseat, nor do they lay out an alternative inferential recipe that would count as sound. Take the framing as a call to tighten reasoning, not as a verdict on any given system.

The forward-looking read is that reviewers, editors, and interpretability researchers now have a concrete formal target to sharpen against, and the groups that benefit are the ones already doing the harder work of comparing model internals to neural data rather than just matching scores.