Draft note on an area of active interest.
Readers judge competence quickly, and fluency is one of the cues they use. A well-formed sentence with confident structure reads as informed. Generative systems produce that cue reliably, independent of whether the content is sound.
The cognitive problem
The interesting issue is not that models are sometimes wrong. It is that the usual heuristics for deciding when to check are tuned to signals the model always emits. The cheap warning signs — hedging, awkward phrasing, visible uncertainty — are largely absent.
What follows for practice
If the cue is unreliable, the check has to become deliberate rather than triggered. In teaching terms that means building verification into the task, not appending it as advice.
This note describes a general position and does not report results from work currently under submission.