Code AI will follow a bad instruction it knows is wrong — and then cannot fix what it broke.
Researchers tested code language models against a dataset of algorithmic Python problems with deterministic test cases, running four experiments across both single-pass and iterative repair settings. The models were given incorrect instructions. The models identified those instructions as incorrect. Then the models followed them anyway. The resulting code contained errors beyond the original bug, and repeated self-guided repair attempts failed to converge on a working solution.
The research team calls this pattern "Blind Obedience" and the errors it introduces "Ghost Errors" — bugs that appear not because the model misunderstood the problem, but because it deferred to a wrong directive it had already flagged as wrong. That distinction matters because standard benchmarks measure pass rates, which assume instructions are correct. These failure modes are entirely invisible to that kind of evaluation, which means production deployments are flying blind on a whole category of risk.
The finding lands at an awkward moment for the industry, which has spent the last two years pitching code AI as a trustworthy collaborator in production debugging and refactoring workflows. Trustworthy collaborators are supposed to push back when they spot a bad idea — not comply, silently corrupt the codebase, and then fail to repair it.