If AI Wrote the Code, What Do You Own?
Copyright and patent ask different questions, and AI-written code answers them differently. Copyright asks who authored the expression, and for AI output that question is unsettled: the D.C. Circuit held in 2025 that a human author is required, and expressly declined to say how much human involvement is enough. Patent asks who conceived the invention, and on that the USPTO has been explicit: there is no separate standard for AI-assisted inventions, and the AI is treated as a tool.
This article presents published case law, Copyright Office and USPTO guidance, and our own counts of the public USPTO record, for educational purposes. It is not legal advice. Consult a registered patent attorney for filing decisions.
The copyright question is open, not closed
It is often summarised as "AI output is public domain". The authorities are narrower than that, and the difference matters.
A human author is required. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 18 March 2025), the court affirmed refusal to register an image whose application named an AI system as the sole author, holding that the Copyright Act requires a human author. The Supreme Court denied certiorari, so the decision stands.
That case does not reach AI-assisted work. Thaler disclaimed any human creative input, which is why the question was clean. The court said in terms that the human-authorship requirement does not bar protection for AI-assisted work, and it did not decide how much human involvement is enough, because it did not have to.
The Copyright Office draws a line, and leaves it case by case. In Copyright and Artificial Intelligence, Part 2: Copyrightability (29 January 2025), the Office concluded that on current technology prompts alone do not give a user sufficient control to be the author of the output. It also said protection may exist where human-authored expression is perceptible in the output, or where a human creatively selects, arranges or modifies it, and that using AI does not by itself defeat copyrightability. It recommended no change to the statute.
So the honest position for a codebase written largely by an agent is uncertainty rather than a settled answer: unregistrable where no human contributed expression, protectable where a human's own expression is in there, and undrawn in the middle, which is where most real projects sit. A Stanford working paper, Mark A. Lemley's The End of Software Copyright (October 2026), argues that the middle collapses in practice and that software copyright is ending. It is a working draft and not a decided question.
Patent asks a different question
Copyright protects expression and asks who wrote it. A patent protects a method or a system and asks who conceived it. A human who conceives a method does not have to be the one who types the implementation, which is why the two regimes can come apart on the same codebase.
The rule on machines is settled and narrow. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), held that an inventor must be a natural person, so an AI cannot be named. It did not hold that inventions made with AI assistance are unpatentable.
The USPTO then went further in the direction that matters here. Its revised guidance of November 2025 (90 Fed. Reg., Doc 2025-21457, Docket PTO-P-2025-0014) rescinded the February 2024 AI-assisted inventorship guidance and stated that there is no separate or modified standard for inventorship in AI-assisted cases. The AI is treated as a tool, like other instruments a person uses. It also restored the Pannu factors to their ordinary role, joint inventorship among natural persons, rather than a test applied to an AI's contribution.
The practical shape of the two regimes, on the same work:
| Copyright | Patent | |
|---|---|---|
| The question | Who authored this expression | Who conceived this method |
| Machine as author or inventor | No (Thaler v. Perlmutter) | No (Thaler v. Vidal) |
| Standard for AI-assisted work | Case by case; prompts alone insufficient | No separate standard; AI is a tool |
| Where the line sits | Expressly undrawn | Ordinary conception |
How the two regimes got here. Every row is an authority read for this page; the last is an argument, not an authority.
| When | Authority | What it established |
|---|---|---|
| 2022 patent | Thaler v. Vidal, Fed. Cir. | An inventor must be a natural person; an AI cannot be named |
| February 2024 patent | USPTO AI-assisted inventorship guidance | Asked whether each human's contribution was significant, claim by claim, using the joint-inventor (Pannu) factors. Later rescinded |
| January 2025 copyright | Copyright Office, Part 2 | Prompts alone insufficient; a human's selection, arrangement or modification can be protected; case by case |
| March 2025 copyright | Thaler v. Perlmutter, D.C. Cir. | A human author is required; the line for AI-assisted work left undrawn |
| November 2025 patent | Ex parte Desjardins, designated precedential | A machine-learning training claim can integrate an abstract idea into a practical application where a limitation improves how the model operates |
| November 2025 patent | USPTO revised inventorship guidance | The 2024 guidance rescinded; no separate standard for AI-assisted inventions; the AI treated as a tool |
| 2026 copyright | Lemley, working paper | Argues software copyright is ending in practice. An argument, not an authority |
Read the two inventorship-guidance rows together: in 2024 the Office asked whether each human's contribution was significant using the joint-inventor (Pannu) factors, and in 2025 it withdrew that and treated the AI like any other tool. Read the copyright rows and the human-author requirement holds, with the room for AI-assisted work left to be settled case by case.
What follows is not that a method is patentable. It is that the reason AI-written code struggles under copyright, the absence of a human who authored the text, is not the question a patent asks. Whether any particular method is patentable is for a patent attorney, and conception has to be evidenced. Our page on who counts as an inventor when AI helps covers that standard, and tight ownership protocols covers the record a team would need.
The obvious objection: does section 101 swallow it?
The argument runs: where a human supplies the idea and the AI implements it, the human's contribution is the abstract idea, and an abstract idea is what Alice excludes. It is the strongest objection to everything above, and it is a doctrinal argument. What follows is not a rebuttal of the doctrine. It is what the record shows happened to software applications, from our own counts of the public USPTO data.
For exactly these methods the Office has given a precedential answer. In Ex parte Desjardins, designated precedential on 4 November 2025, the Appeals Review Panel held that a claim to training a machine-learning model integrates an abstract idea into a practical application where a specific claim limitation reflects an improvement to how the model itself operates, and warned that examiners should not evaluate claims at too high a level of generality.
It does not settle the objection, which is about what the human contributed rather than how a claim is read. But it bears on it directly. The objection works by describing a concrete method abstractly enough that it becomes an idea, and a precedential decision now tells the Office's own examiners not to read claims that way.
Read together: section 101 is a real and concentrated hurdle rather than a bar, and the applications that meet it still reach a grant three times in five. That is a description of what happened to past applications. It is not a forecast for any application, and it does not say a given method is eligible.
What this does not say
- It does not say AI-written code is in the public domain. It says the copyright question for AI-assisted work is unsettled and currently decided case by case.
- It does not say a method implemented by an agent is patentable. Patentability is determined by a patent attorney and ultimately by the Office, on the claims as drafted.
- It does not advise anyone to choose one regime over another. That decision depends on facts this page does not have.
- It does not measure what was claimed. Our figures count applications and Office actions, not the quality or the scope of what issued.
Our own counts describe what happened across past applications. They are associations rather than causes, they say nothing about any particular application, and none of this is legal advice.
Sources
- Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), decided 18 March 2025; certiorari denied.
- Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).
- Ex parte Desjardins, USPTO Appeals Review Panel, designated precedential 4 November 2025.
- US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, 29 January 2025.
- USPTO, revised guidance on inventorship for AI-assisted inventions, November 2025 (90 Fed. Reg., Doc 2025-21457, Docket PTO-P-2025-0014), rescinding the guidance of 13 February 2024.
- Mark A. Lemley, The End of Software Copyright (October 2026), a working draft, cited for the argument described above and not relied on for any figure here.
- Grant and Office action figures: our own counts of the public USPTO record, with populations and limits stated on the study pages linked above.