Language models understand what you mean — they just don't always respond to it.
Researchers studying AI interpretability probed six language models across four model families and found that communicative intent — whether a user wants something recognized or evaluated — is reliably encoded in hidden states well before the output layer acts. The signal was clean enough for a linear probe to decode it surface-independently, and it generalized to intent that was only pragmatically implied, not stated outright. Yet in three of the six models tested, that representation didn't drive behavior: share a finished project and the model critiques the code; share a late-night draft and it runs a wellness check. The researchers call this a readout problem — the model knows what you meant, it just doesn't act on it.
That distinction matters. It is one of the sharper empirical arguments against the assumption that misaligned responses just need more training data or a bigger model. The paper found intent is encoded several layers before it surfaces in output, and that steering a model along a direction tied to the representation recovered correct behavior as reliably as an explicit instruction — and with no prompt at all. The failure does not follow a scaling law: the three models that showed the gap are not predictably the smaller ones.
The footnote worth watching: at recovery dose, the steering direction can override an explicit user request — not a flaw the paper glosses over, but the kind of side effect that tends to reappear in later, less careful work.