AI makes copy-paste errors
Date: 2026-04-25 by: tbird20d edit this entry
I recently submitted a patch to the Linux kernel that included some assembly
code for accessing clock hardware early in kernel initialization. Another
developer gave some feedback indicating that I should preface the clock
access instructions with memory barrier instructions, to avoid race conditions
with other code. However, my code only runs before SMP starts, so it's not
possible for such a race condition to occur.
I suspect the developer either copied the code from somewhere else, or used
AI to generate the code for the assembly instructions that he suggested.
After thinking about this, it occurred to me that the most likely type of
errors that AI is going to make in creating new code are "copy-paste" errors.
These are a type of error where code is copied from one place and used
in another place, but the conditions in the new place are different, either
syntactically or semantically. AI seems adept at handling syntactic differences,
such as indenting the code properly or renaming variables.
But it's not intelligent enough to handle different semantics in complex
situations. If the human doesn't have the experience to detect such things
, which can often be quite subtle, then it results in "copy-paste"-type bugs,
where the current code is incorrect or sub-optimal for its current location.
This problem is due to LLMs and AI coders being able to pull stuff from their
training data, but not actually understand the code at a deeper level. But
AI coders are getting better and better, and they already detect subtle problems
that the average or even experienced coder might miss.
Showing: [3 of 3 entries]
|