Ok, short answer first: no, the provider does not own the software. Not Anthropic, not OpenAI, not Google, not Microsoft, not Cursor. And now the long one, because that is the uncomfortable one. The question came up at the table this week, meant quite seriously, using Claude Code as the example, but it applies to every paid model: if the tool writes a hundred percent of the code, then the tool is the creator, so it belongs to it? That sounds logical. It is wrong, but the right answer pleases most people even less. I am not a lawyer. But I have read what lawyers and authorities say about it, and I build every day with exactly these tools. So here is what I understood, and what I do about it.
The question
Two years ago the question was theoretical. An assistant suggested a few lines, the human took them or not, the human was the author. Today someone sits in front of an agent, says "build me a CRM", goes to get coffee, and comes back to an application with twenty thousand lines he has never read. Who is the author now? The one who said the sentence? The company running the model? The model? Or nobody?
Three roles you have to keep apart
The mistake in "it belongs to the tool" is that three different things land in one pot.
The technical creator. That is the model. It produced the characters. That is a fact, not a right. A hammer is also the technical creator of the dent, and still the table does not belong to it.
The author in the legal sense. That is, in Switzerland as in the US, a human, and only when an intellectual contribution of their own is in it. Swiss copyright law speaks of an intellectual creation with individual character, and only a natural person can be an author. A machine is not one.
The rights holder under the contract. That is whoever gets, by contract, what there is to get. And here every provider I know says the same: they assign the rights in the outputs to the user, to the extent such rights exist. The last five words are the whole point.
So the tool is role one, has no claim to role two, and the provider gives away role three voluntarily. So far the good news: no provider owns your CRM. The bad news comes in a moment. First a look into the contracts, because I wanted to know myself.
What the providers say: all the same
I looked at the terms of the five tools that I and the people around me work with. The wording differs, the content is the same, and none of them promises that any right arises in the outputs at all.
| Provider | What the contract says | What it does not promise |
|---|---|---|
| Anthropic (Claude, Claude Code) | Assigns to the user all rights in the outputs, "if any". Under the commercial terms the output belongs to the customer. | That the rights exist. |
| OpenAI (GPT, Codex) | Assigns to the user all rights in the output, subject to compliance; enterprises own inputs and outputs "where allowed by law". | That the output does not infringe third-party rights. |
| Google (Gemini) | Claims no ownership of generated content. But expressly reserves the right to generate the same or similar content for others. | That your code is only yours. |
| Microsoft (GitHub Copilot) | Claims no rights in suggestions; you retain ownership of and responsibility for your code. Copyright indemnification only in the Business and Enterprise plans. | That the suggestions contain no third-party code. A lawsuit by open-source developers has been running for years. |
| Cursor (Anysphere) | Assigns to the user all rights in suggestions, "if any". The business agreement adds: suggestions may be the same as those given to other customers. | Exclusivity. |
Two sentences from this table I find remarkably honest. Google: we reserve the right to generate the same content for others. Cursor: suggestions may be the same as those for other customers. That is not small-print malice. That is the truth about what a model is: a machine that gives the same answer to the same question, no matter who asks. And that leads straight to the bad news.
The uncomfortable answer: nobody
"To the extent such rights exist" means: the provider transfers to you what it has. If it has nothing, it transfers nothing. And whether any copyright arises at all in purely machine-generated code is the question being answered right now in two jurisdictions, and both times with no.
The US Copyright Office stated in January 2025, in the second part of its report on AI: purely AI-generated material is not protectable, and a prompt alone, even a detailed one, is not enough to make the user the author. A prompt is an instruction that conveys an idea, and ideas are not protected, only their execution. The Office even says why: with a model of billions of parameters the human cannot control precisely enough what comes out. The court in the Thaler case confirmed that. Protection exists where a human has recognisably contributed: own code that remains visible in the result, creative selection and arrangement, own changes to the result.
Switzerland has no ruling yet, but the scholarship is fairly united, and the University of Zurich white paper of June 2025 puts it in one sentence: with AI, an intellectual creation exists only if the human can anticipate the result from his input so precisely that he steers it deliberately. Mere smoothing is not enough. Pure machine output is in the public domain.
Three scenarios
So how much human is in it is the whole question. Three cases, as I see them every day.
You write, the model helps. Autocompletion, a suggestion here, a function there, and you decide line by line. Your contribution is recognisable, the protection is as before. That is the case the law was made for.
You design, the model writes. You set the architecture, the modules, the interfaces, the rules, and the agent fills them. Here it gets interesting and unclear. The architecture is your work. Whether the code generated from it is yours too depends on how precisely you anticipated and steered it, and on whether you can prove that. This is where the arguments of the next years will be.
You say "build me a CRM". The agent does everything. By today's reading the code belongs to nobody. Not to you, not to the provider, not to the model. You can run it. That is all.
Where the danger really lies
Most people who ask the question are afraid someone will take their code away. That is the wrong fear. The right one is more boring and more expensive.
Imagine two companies that both use the same agent and both type in roughly the same thing, because they have the same problem. Both get very similar software. Neither can forbid the other anything, because neither owns anything. That is not theft. That is the end of a business model that rested on code being scarce and belonging to a human.
And the second danger, which is in every legal source and which I can confirm from experience: the model learned from code that belongs to someone, often open source with licences that inherit. If it reproduces a piece of that and you build it into your product, you have a problem you do not see until someone else sees it.
What a software company still sells
Here the legal question ends and the interesting one begins. If code costs nothing and belongs to nobody, what does a software company still sell? I have three answers, and none of them is code.
First, the architecture and the decisions. Why it is built this way, what was discarded, what the rules are. That is human work, it is protectable, and it is what an agent would not have without you. Second, the knowledge of the system: who understands it, who can explain it, who knows what will tip in two years. I called that the film no tool sees, in "Software is a film". Third, the ability to keep developing the system without nobody understanding it after eighteen months. That is the answer to what I wrote about volume and speed.
The distinction we have to learn is: who wrote the code, versus who developed the software. Those are not the same. An architect who makes the essential decisions does not have to type every line, that was always so. Conversely, nobody becomes the creator of an application because he commissioned an agent.
What I do about it
I cannot change the law, and I cannot wait until a Swiss court has decided. What I can do is build so that the human share stays visible. Not because I want to win a lawsuit. Because it is the better way to build software anyway.
Record the history of how it was made. Which requirement, which architecture decision, which rule, which correction by a human, in which order. That is exactly the chain recorded at NeoCoder anyway, because we need it for learning: state, action, result, next state. On the side it is the best proof of who created what here. A prompt proves nothing. A hundred documented decisions, more likely.
The architecture belongs to the human, and it says so. Modules, boundaries, interfaces, rules come from the architect, before the agent, in writing, versioned. The agent works inside them. Then the creative work is not in the generated code but in what shaped it, and that is the part that is protectable.
Less code, more of your own. What the agent builds in two thousand lines is nobody's. What a human cuts to two hundred and understands in the process is more likely his. Here too less is the goal, and here too for a reason that has nothing to do with lawyers.
Check licences before it hurts. A scan for known open-source code in the generated output, on every build, not at the sale. That is a script, not a project.
Evidence
No legal statement here without a source. These are the texts I rely on, all read in October 2026. And again: I am not a lawyer. Whoever needs to know for sure asks one.
US Copyright Office, "Copyright and Artificial Intelligence, Part 2: Copyrightability", 29 January 2025. Purely AI-generated material is not protectable; prompts alone do not make the user the author; protection for recognisable human contributions, selection, arrangement, modifications. Summarised among others at Reed Smith and Akin Gump.
Thaler v. Perlmutter, US Court of Appeals D.C. Circuit 2025. A work generated entirely by a machine has no human author and thus no protection. Placed in context at terms.law.
Provider terms. Anthropic: overview and wording at gwpl on GitHub, commercial version in the Anthropic archive. OpenAI: Enterprise Privacy and analysis at terms.law. Google: Gemini API Additional Terms. GitHub Copilot: analysis at terms.law and Comply Code. Cursor: Terms of Service and Master Services Agreement. Always read the current version before relying on a clause, they change.
Florent Thouvenin, ITSL University of Zurich, white paper "KI und Urheberrecht", June 2025. Swiss law: intellectual creation only if the human deliberately steers the output through his input; the author is a natural person (URG Art. 2 and 6). itsl.uzh.ch. Practical on the open-source licence risk: justis.ch. Swiss courts have not ruled so far.
What stays
The question "who owns the code" will in a few years sound like "who owns the electricity from the socket". Nobody, and it does not matter. What does matter: who owns the knowledge of how the software is built, why it is built that way, and how to keep it alive. That still belongs to humans. And that is, if you go about it right, also what a software company will earn its money with in the future.
No provider owns the software, they all give away the rights they have. They just have none for purely machine-generated code, and neither do you. Nobody has any. So what counts is no longer who wrote the lines but who developed the software: architecture, decisions, understanding, care. I build so that this stays visible. Not for the judge. For the system.
Written by me. The thoughts, the values, the learnings, the mistakes: all mine. Grammar and spelling are corrected by a twin model of my own, trained on my texts. Sometimes that turns into a mix. Of course I read everything again, but nobody is infallible. Whatever stays in is mine too.