GPT-4 can tell the difference between "tigers are striped" and "cars have radios" — and that turns out to be a harder problem than it sounds.
Researchers tested whether language models could distinguish between principled properties, traits that are true because something belongs to a category, and statistical ones, traits that just happen to be common among members. The gap matters: a tiger without stripes is still a tiger; a car without a radio is still a car. Humans track this distinction almost automatically. Every language model tested struggled with it, controlling for how often a property appeared in text — until GPT-4, which handled it correctly.
The finding cuts against a long-standing assumption in cognitive science that this principled-vs-statistical distinction is innate and cannot be acquired from experience alone. If GPT-4 learned it from text, that suggests the signal is present in language and that sufficiently large models can extract sophisticated causal structure without grounding in the physical world. It also raises a quieter point: what other supposedly unlearnable conceptual distinctions might fall the same way as models scale?
That said, one model passing one benchmark is not a theory of mind — it is a data point, and the field has learned to be cautious about what benchmark performance actually tells us about internal representations.