Who Owns Code Generated by Claude Code?

You open a new file, type a prompt, and Claude Code produces 200 lines of working TypeScript. You review it, ship it, and bill the client. Somewhere in the back of your head, a question you have not fully answered: who actually owns that code? It is not a hypothetical. It is a question with real answers, and most of those answers are uncomfortable.

The legal landscape, step by step

Here is how the current framework applies to code you generate with Claude Code.

1. Check Anthropic's terms of service. Under Anthropic's usage policy, you retain ownership of outputs you create through the API and consumer products. Anthropic explicitly disclaims ownership of the output. This sounds reassuring. It is not the whole story.
2. Understand what Anthropic can disclaim. Anthropic can tell you they do not want the copyright. They cannot grant you copyright they do not have. Copyright only attaches to works created by humans. If no human author sufficiently contributed to the expression, there may be no copyright to transfer or retain.
3. Know what the Copyright Office says. The U.S. Copyright Office's guidance, updated in February 2023 and again in 2024, states that works produced entirely by AI without human creative control are not eligible for copyright registration. They have rejected several applications on exactly this basis.
4. Map your human contribution. This is the operative question. Did you write the prompt and accept the output verbatim? Did you select from multiple outputs, edit the result, combine it with hand-written code, or structure the architecture yourself? More documented human choice means more defensible authorship.
5. Document as you go. If you are shipping AI-generated code into a commercial product, start keeping records. Prompts, edit history, the decisions you made about structure and logic. This documentation is what a copyright claim would rest on if challenged.

The real cost of this uncertainty

The legal ambiguity is not free. It has a cost, and it accrues in specific places. Client contracts. If you are a freelancer or agency delivering software built with Claude Code, your contract almost certainly includes an IP assignment clause. You are warranting that you own the code you are handing over. If that warranty is wrong, you have a liability problem. Reviewing and potentially amending those clauses takes time and legal fees. Due diligence for funded companies. Investors ask about IP chains before closing rounds. A codebase where significant portions were generated by AI, with no documentation of human authorship, is a red flag in due diligence. Not a deal-killer in every case, but something that will require explanation and potentially escrow or indemnification provisions. Open source licensing complications. If AI-generated code has no copyright owner, it cannot be licensed under GPL or MIT or Apache. It may be in the public domain by default. That sounds fine until you realize it also means you cannot enforce license compliance, and you cannot prevent competitors from taking your open source project and going proprietary with it. Time cost of the workaround. The practical defense is documentation. Keeping prompt logs, maintaining edit history, writing the scaffolding yourself and using Claude for implementation details rather than architecture. That is not zero-cost. For a small team shipping fast, it is a non-trivial workflow change.

What the pricing page does not cover

The Legal Layer analysis makes a point that gets overlooked in most tool comparisons: the subscription cost of Claude Code is not the actual cost of using Claude Code in a commercial context. The Max plan runs $100 per month. That gets you expanded usage limits and access to the full model. What it does not buy you is legal clarity.

Cost category	Estimated range	Who pays it
Claude Code subscription	$20-$100/month	Individual or team
IP clause review (one-time)	$500-$2,000	Freelancers, agencies
Contract amendment per client	$200-$800	Anyone delivering software
Due diligence response (funding round)	$3,000-$15,000	Startups raising capital
Ongoing documentation workflow	1-3 hrs/week	Any commercial user

The subscription price is the floor, not the ceiling. Teams that treat it as the full cost are the ones that get surprised during an acquisition or a client dispute. The comparison to GitHub Copilot is instructive. Copilot has an IP indemnification clause for Enterprise customers, introduced in late 2023, that covers certain infringement claims if the generated code matches training data. Anthropic does not offer equivalent indemnification. That asymmetry matters if you are making a procurement decision for a team. You can see how these tools differ more directly in our Cursor vs GitHub Copilot comparison, which covers the IP treatment in detail.

How this resolves over the next six months

Here is a specific prediction: by the end of Q1 2026, at least one major AI coding tool will introduce a contractual IP indemnification tier, similar to what GitHub did with Copilot Enterprise, and it will be priced at a meaningful premium above the standard subscription. The pressure driving this is not philosophical. It is commercial. Enterprise procurement teams are starting to ask hard questions about IP warranty chains before approving AI coding tools at the organizational level. Legal departments at companies with pending exits or fundraising rounds are the ones surfacing this issue. Tool providers that want to close enterprise deals need a credible answer. Anthropic will face this directly. Claude Code is positioned as a serious development tool, not a consumer toy. The enterprise market it is targeting will not accept "we disclaim ownership, good luck" as a sufficient IP policy indefinitely. Either Anthropic adds an indemnification clause, or a competitor does first and uses it as a differentiator. The prediction is falsifiable: if no major AI coding tool has introduced a paid IP indemnification tier by March 2026, the enterprise procurement friction I am describing is less severe than the current signals suggest.

The mechanism underneath copyright and AI output

Copyright law in the U.S. has a human authorship requirement that traces back to a Supreme Court case from 1884, Burrow-Giles Lithographic Co. v. Sarony, which held that a photograph could be copyrighted because it reflected the photographer's original creative choices. The principle has held: copyright requires a human author making creative decisions expressed in the work. When you prompt Claude Code, the model generates output using transformer-based prediction over a probability distribution shaped by training data. The model does not have intentions or make creative choices in any legally cognizable sense. It produces text that statistically fits the context. The question for courts - and no court has fully answered this for code specifically - is how much human contribution is enough. A prompt like "write a REST API for user authentication" with zero editing probably does not clear the bar. A session where you wrote the interface definitions, specified the error handling logic, rejected three outputs and edited the fourth substantially, and integrated the result into architecture you designed is a different story. Here is a simplified illustration of how that documentation might look in practice:

# session-log-2025-06-12.md
## Prompt
"Implement the UserRepository class per the interface in types/user.ts.
Use pg-promise. Follow the error handling pattern in OrderRepository."
## Human contributions
- Wrote types/user.ts interface (100% human)
- Specified pg-promise (human decision)
- Rejected first output: missing transaction handling
- Edited final output: rewrote error mapping section (lines 87-112)
- Integrated into existing service layer (human)
## Files modified by human after generation
- src/repositories/user.ts (lines 87-112 rewritten)

This is not legal advice. It is a documentation pattern that makes human authorship legible if it is ever questioned. The difference between "I used Claude" and "I used Claude and here is what I contributed" is the difference between an uncertain IP position and a defensible one. For teams comparing options, the Cursor vs Claude comparison covers how each tool handles the generation workflow differently, which affects how easy this kind of documentation is to maintain.

Before you close the tab

• Confirm your client contracts do not contain IP warranties that outrun what you can actually defend
• Check whether your company's IP assignment agreements cover AI-assisted work or exclude it
• Verify you have a prompt log or edit history for any AI-generated code in production
• Confirm the sections you edited manually are distinguishable from the raw generated output
• If you are in a funded company or approaching a raise, flag the AI-generated codebase question to your legal counsel before due diligence starts, not during it
• Check whether the AI coding tools you use offer any contractual IP coverage, and at what tier