mode collapse
When a model settles into one preferred way of saying things instead of spreading across the many ways a person might phrase it.
Ask ten people to describe the same photo and you get ten different sentences. A language model, asked the same thing repeatedly, tends to converge on one house style. That narrowing is mode collapse, and the opposite state, where the model spreads its choices more evenly across possible phrasings, is called mode coverage.
The interesting part is where it comes from. Bradley Emi of the detector company Pangram argues it is largely a side effect of post-training, the stage where a model learns rules about what to avoid and how to behave. Base models, the raw versions before that shaping, write with much more variety and are correspondingly harder to detect. If that holds, then every guardrail a lab adds makes its model’s writing a little more uniform, and a little easier to spot.