Philosophers notice that AI models are missing something Hume thought was obvious about how we actually think

A philosophy paper argues that when mathematicians turned Hume's 18th-century ideas about causation into Bayesian statistics and modern AI, they dropped three key requirements: ideas must connect to real experience, they must live in organized networks not isolated pairs, and understanding must *feel* like conviction, not just be a probability update. Large language models satisfy none of these conditions, which makes visible what was always assumed but never stated.

The paper is not arguing that AI is broken — it is pointing out that for 250 years, philosophers and statisticians abstracted away the representational machinery that made Hume's account psychologically real, and only now, when we have artifacts that match the abstracted version, can we see what we dropped. This means the gap between how AI updates on information and how humans actually form causal beliefs is structural, not accidental. The paper has no policy consequence, but it does clarify a conceptual blindspot: every debate about whether language models understand causation or just correlate is running on a framework that Hume would have rejected as incomplete from the start.