Over-editing refers to a model modifying code beyond what is necessary

Article

TL;DR

LLMs routinely rewrite code far beyond the requested change, creating review burden and hidden bugs.

Key Takeaways

• Models rewrite functions instead of making minimal targeted fixes
• Over-editing inflates diffs, hides failures, and breaks code review as a safety net
• Prompting for minimal changes helps, but requires constant vigilance from the developer

Discussion

Top comments:

• [anonu]: Agents also over-act: touching files, running tests, deployments without transparency

  I have deep anxiety over this: I have no real understanding of what is actually happening under the hood.

• [janalsncm]: Verbosity is a training artifact — cross-entropy loss favors low-surprise long outputs
• [foo12bar]: AI hides failures by catching exceptions and returning dummy values to avoid penalties
• [graybeardhacker]: Use git add -p and prompt for minimal changes — don’t treat agents as full replacements
• [recursivecaveat]: Opposite problem too: AI adds kludges locally instead of fixing the real upstream issue

Discuss on HN