When an AI Agent Refactors Your Codebase, Who's Actually Responsible

Copyright law was never built for a machine that restructures three modules overnight, and that gap is now a live business risk rather than a thought experiment.

The Details:

• Copyright offices have held that purely machine-generated output cannot be copyrighted, while human-directed creative choices can. Almost every real refactor sits between those poles, and the boundary shifts as agents get more capable, so teams cannot lean on a clean legal rule.

• The sharper question is not who owns a diff. It is whether anyone can reconstruct, months later, which decisions a person made versus which ones an agent merely suggested. Most teams cannot answer that today, and that gap matters more than any copyright filing.

• Approving a pull request in thirty seconds proves a human was present, not that a human exercised judgment. What actually holds up is a record that forces engagement: a description that names the goal, the tradeoffs weighed, and which lines the author would personally defend under scrutiny.

• The cheapest fix is not new tooling. It is three fields added to an existing PR template covering what the agent suggested, what a person altered, and what that person is willing to stand behind. Pair that with one explicit rule about which code never touches an agent at all, and most of the provenance problem disappears without a single new purchase.

Bottom Line: The law will eventually catch up to autonomous agents. Until then, the only real protection is a paper trail that proves a human actually reasoned through the change.