Coding AI's Blind Spot: It Fixes Bugs Only When Told Where

A team of computer scientists studied something surprisingly basic: when an AI edits computer code, who decides where the changes go? They tested four approaches — rewriting a whole file from scratch, searching for a phrase and swapping it, locating the right spot and filling in the blank, and making tiny word-level tweaks. It turns out the approach matters as much as the AI model itself.

The key finding is what they call a "composition gap." When researchers handed the model a map showing exactly which lines needed changing, it handled several edits at once quite well. But when the model had to find those lines on its own, most of that ability vanished. In plain terms, the hard part of editing isn't writing new code — it's knowing which code to touch.

There's a trade-off hiding in that problem. Play it safe by flagging a wide stretch of code for rewriting, and the model may damage parts that were already correct. Mark too narrow an area, and a required change gets missed entirely. Good editing needs both full coverage and tight boundaries, and those two goals pull against each other.

The effect isn't limited to code. A smaller test on editing Wikipedia-style sentences showed the same pattern, suggesting this is a general weakness in how these tools work, not just a programming quirk. For anyone using AI assistants at work, the takeaway is practical: today's AI is a fast typist, not a reliable editor. It performs best when a person says exactly what to change — so expect humans to stay in the loop for a while yet.