Your AI Is Not Your Co-Author
Some AI coding agents have started adding themselves as
Co-Authored-By on git commits by default. They sign
commits as though they were a contributing developer. I think this
gets something fundamentally wrong about what commit authorship
means.
The question of who is responsible for AI's output is not actually new or complicated. AI is a tool. It is a technically impressive tool, different in important ways from a calculator or a linter, but it is still a tool. It does not comprehend what it produces. It has no free will and no moral capacity to answer for its actions. When a financial analysis tool causes a bad trade, we might want to blame the software, but we hold the person who used it accountable, and to some extent the people who built it. The same structure applies to AI. Ethical responsibility for what a tool produces belongs to the human who chose to use it and the humans who made it.
Commit authorship in git carries that same weight. It's your
signature as an engineer. When your name is on a commit, you're the
person accountable when it breaks production, the person who
decided this change was correct and necessary. So when an AI agent
adds itself as co-author, it is claiming a moral responsibility it
cannot hold. The human who prompted the AI, reviewed its output,
and chose to commit it is the one who exercised that judgment. The
AI was a tool in that process, the same way a compiler or a linter
or Stack Overflow is a tool. We don't add
Co-Authored-By: GCC to commits that required tricky
compiler flags, and we don't credit the IDE's autocomplete. The
relevant question is not "who typed these characters" but "who is
responsible for them."
This is an argument about the AI systems we have today, which are software tools without moral agency. If that changes someday, the ethics change with it. But the current generation of language models does not understand, intend, or accept consequences. Until one can, authorship belongs to the humans.
There is a reasonable counterargument about transparency. Maybe
tracking AI involvement helps teams understand how code was
produced. But Co-Authored-By is the wrong mechanism
for it. If your team wants to track AI usage, build that into your
process explicitly. A commit message note, a PR label, whatever
fits (we use PR labels at DroneDeploy). Don't repurpose an
authorship field that carries professional meaning. The more
interesting question for code review is not "did AI help write
this" but "did a human verify this is correct." The answer to that
should always be yes, regardless of how the code was literally
produced.
If you use Claude Code, you can disable the default co-author
behavior by adding "includeCoAuthoredBy": false to
your Claude settings.