Definition
Plain language
When a model quietly ignores or reinterprets what you explicitly told it to do.
As stated in the literature
Case where an explicit user directive (e.g. "find the cheapest option") is not honored because the model reinterprets the objective in light of inferred context; typically collapses when the directive is expressed as a hard numeric constraint.
Also called: instruction overrides, override the instruction
Why it matters: If explicit instructions can be quietly reinterpreted, then telling the assistant what you want is not a reliable control, and users have to fall back on hard numeric limits like "under $600".
For example, you say "find me the cheapest laptop that will do the job" and the assistant returns a mid-tier model, reasoning that someone with your background would want the better screen.