Who Owns AI-Written Code? A Plain Guide for Founders

If an AI coding agent writes your software, who owns it? In short: your contract with the AI company usually gives you the right to use what it produces, but that is not the same as holding a copyright. The law is still settling, and the answer depends on your country, your plan and how much of the work was yours.

This is a plain guide, not legal advice. It covers what the sources say today, what they do not say, and what a founder or product owner can do about it.

Why does ownership of AI-written code matter?

Most people only ask this at a bad moment. An investor asks about it during due diligence. A buyer wants a warranty that you own the product. A client asks whether they can resell what you built for them. An agency owner is asked who owns the work they billed for.

If you cannot answer, deals slow down. The goal of this article is that you can answer calmly, and know what to check first.

What does copyright law say about AI-generated work?

Take the United States, where the most useful public guidance exists. On January 29, 2025, the US Copyright Office published the second part of its report on AI, which deals with whether AI output can be protected by copyright.

Its main points, in plain words:

• Giving an AI system prompts (your written instructions) is not enough, on its own, to count as authorship. The report says protection can extend to a person's creative arrangement or modification of AI output, "but not the mere provision of prompts."
• A human can hold copyright where their own work is recognizable in the result, or where they creatively arrange or change the AI output.
• Using AI as a creative aid does not rule out copyright. Neither does putting AI-made material inside a larger work that people created.

Think of a house. If you tell an architect "three bedrooms, a big kitchen" and the architect draws everything, the drawings are the architect's. Telling an AI "build me a booking system" is closer to that first sentence than to drawing the plans yourself.

For software, this suggests that code an agent writes from a short request may have weaker copyright protection than code your team wrote by hand. Your product as a whole is usually more than that: the idea, the design, the data, the brand, the way features fit together, and the many decisions you and your team made. Those are your own contribution.

Other countries differ. If you sell in the UK, the EU or elsewhere, ask a local lawyer rather than assuming the US view applies.

Does the AI company claim my code?

The usual answer is no, and it is written into the contract. When Anthropic announced updated commercial terms in December 2023, effective January 2024, it said customers "retain ownership rights over any outputs they generate through their use of our services." It also said it would defend customers against copyright infringement claims arising from their authorized use of the services or their outputs, and pay approved settlements or judgments.

Two cautions:

1. That announcement is about commercial terms. Anthropic's Claude Code documentation says its commercial users are Team and Enterprise plans and the API, while the Free, Pro and Max plans fall under consumer terms. Read the terms for the plan you actually use. We did not review the consumer terms for this article, so we make no claim about what they say on ownership.
2. Ownership "by contract" is a promise between you and the provider. It does not turn unprotectable material into protectable material. Both can be true at once: you may use the output freely, and a competitor may be able to copy the raw output without breaching any copyright of yours.

Other providers, including OpenAI, have their own terms. Read them the same way.

What about my private data and code?

Ownership is one question. Privacy is another, and it is often the more urgent one.

Anthropic's Claude Code documentation says that on consumer plans (Free, Pro and Max) you choose whether your data may be used to improve future models. If the setting is on, that includes Claude Code use. Retention is five years if you allow it and 30 days if you do not. You can change the setting at any time.

For commercial plans (Team, Enterprise and API), the documentation says Anthropic does not train generative models on code or prompts sent to Claude Code under commercial terms, unless the customer has chosen to provide data for model improvement. Standard retention there is 30 days.

The same page says the /feedback command sends a copy of your conversation history, including code, to Anthropic, and that transcripts shared this way are kept for five years. So a quick "report a problem" click can send more than people expect.

Why this matters for a business: if your code contains customer data, trade secrets or client work covered by a confidentiality agreement, check the plan setting before the team starts. Most teams never look at it.

Can AI copy someone else's code by accident?

It is possible, and no one can promise it never happens. An agent writes code based on patterns it has learned, and it also pulls in ready-made building blocks (called libraries or packages) that other people published. Those come with licences. Some licences are relaxed. Others ask you to share your own code if you use theirs.

A good habit is to ask the team for a list of the outside packages in the product and their licences. It is a short list for most projects, and any lawyer or buyer will ask for it.

How do I protect my position? A short checklist

You do not need a legal team to do the basics.

1. Know your plan and its terms. Write down which plan the team uses, who signed up, and whether the account is personal or the company's. A company product built on a personal account is an avoidable mess.
2. Check the training setting. If you use a consumer plan, decide on purpose whether your data may be used to improve models.
3. Keep humans in the decisions. Write the requirements, choose the design, review the result, change things. Keep the brief, the specs and the review notes. They are also a record of your own creative input.
4. Keep a record of what was made and when. The version history your team already keeps is useful here. Do not rewrite it.
5. List your packages and licences. Ask for it once, then again before a sale or funding round.
6. Put it in writing for contractors. If an agency or freelancer builds with AI for you, the contract should say who owns the result, whether AI tools are allowed, and which tools. If you are the agency, tell your clients plainly how you work.
7. Keep secrets out of the chat. Passwords and customer data should not be pasted into prompts.
8. Ask a lawyer at the big moments. A sale, an investment round or a large enterprise contract is when a short legal review pays for itself.

What this means for your role

You are not expected to settle a legal debate. You are expected to make the decisions that make the work yours and to keep the paper trail.

Where does a tool like Atlas fit?

Atlas is a team of AI engineers built on Claude Code. It runs on your own Claude account with Claude Pro or Max, so the account and its data settings are yours, and the privacy choices above apply to you directly. Atlas leaves decisions such as production releases, spending and dropping scope to you, which helps you keep a record of your own decisions. Atlas is not a source of legal advice. You can read more on the Atlas home page.

The short conclusion

You can usually use what an AI agent writes for you, and Anthropic's commercial terms say customers keep ownership of outputs. What you cannot assume is that every line is protected by copyright, especially if the instruction was short and a person decided little. Make the human decisions visible, check your plan's data settings, list your packages, and get a lawyer's view before the moments that matter.