Imagine you hand a code review assistant a program and ask, "Anything wrong here?" A good assistant would say "nope, looks fine" when the code is genuinely correct. This paper investigates what actually happens when you let large language models (LLMs) play that role on a loop — and the answer is unsettling.

The researchers ran a simple experiment: feed programs to several popular LLMs and repeatedly ask them to find and fix bugs. They tried this on programs that were already correct and on programs with real bugs. Two findings jump out:

• LLMs hallucinate bugs in perfectly good code. When given bug-free programs, the models still confidently claim to spot problems and "fix" them — often introducing regressions in the process.
• They break more than they mend. Across the runs, the rate at which LLMs damaged correct programs was higher than the rate at which they successfully repaired buggy ones. In other words, if you blindly accept every fix an LLM suggests over multiple rounds, your codebase drifts toward worse, not better.

Why does this happen? LLMs are trained to be helpful, and "helpful" in a bug-fixing prompt means producing a fix. They rarely have the confidence — or the incentive — to say "nothing to do here." So each iteration nudges the code somewhere, and those nudges accumulate. The paper calls out the long-term dynamics: even if a single round has modest error rates, iterating turns those errors into steady erosion.

The title captures the practical takeaway perfectly: if it's not buggy, don't fix it. That sounds obvious, but it's exactly the discipline that "agent-in-the-loop" tooling lacks by default. Automated repair pipelines that keep asking "any more bugs?" until the model shrugs will never shrug — the model will keep inventing things to change.

The authors' recommendation is straightforward: treat LLM bug-fixers as one-shot suggesters that require a real oracle (tests, human review, or a stronger verifier) to accept or reject each change, and be extremely wary of open-ended iterative loops. Confidence calibration — teaching models to say "no bug found" — is flagged as an important missing piece.