The answer is messier than most developers realize. When you write a prompt and Claude Code generates the implementation, who holds copyright to that output? Your employer? You? Anthropic? Nobody?
That last option is more likely than most people expect.
What Copyright Law Says About AI Output
U.S. copyright law requires human authorship. The Copyright Office has been consistent on this: purely AI-generated work doesn't qualify for protection. In 2023, it rejected a copyright registration for a comic book where the images were made with Midjourney. In 2024, it clarified that AI outputs are only protectable when a human author makes "sufficient creative choices" - selecting, arranging, or modifying the AI's output in a meaningful way. Typing a prompt and accepting the first result generally doesn't meet that bar.
This creates a practical problem for developers using AI coding tools. If you write a prompt and ship the generated code without significant modification, that code may sit in a copyright gray zone - and your company can't necessarily stop a competitor from copying it.
Anthropics terms of service assign outputs to the user; they're not claiming ownership of what Claude Code writes. That's the straightforward part. The harder question is whether that user-owned output is actually protectable under copyright law, which is something Anthropic's terms can't answer. Terms of service can assign contractual rights between parties. They can't create copyright protection where copyright law says none exists.
A legal analysis by LegalLayer walks through the current U.S. framework specifically for Claude Code, covering what the authorship threshold means in practice.
The Practical Stakes for Companies
For individual developers building side projects, this is mostly a background concern. For companies shipping commercial software, the exposure is real:
- Proprietary software built heavily on unmodified AI output may have weaker IP protection than assumed
- Open source contributions generated by AI tools raise license compliance questions - can you grant a license to something you don't legally own?
- Employee IP agreements that assign all code to the employer may need updating to address AI-generated outputs specifically
The safest practical position right now: treat AI-generated code as a first draft. Developers who review, refactor, and meaningfully shape what an AI produces have a stronger copyright argument than those who generate-and-ship. How much modification is "sufficient" is still an open question courts haven't fully answered.
The Copyright Office is expected to publish further guidance on AI authorship in 2026. Until that guidance lands, any company building commercially significant software primarily through AI generation should get a legal opinion rather than assume the same protections apply as to hand-written code.