IP Ownership of AI-Generated Code

The ownership question

When an agent generates code, who owns it?

The honest answer: nobody knows for certain. Ownership and copyrightability are distinct legal concepts, and courts are still working through both. The cases keep coming, the rulings sometimes contradict each other, and enterprises are left building frameworks for a moving target.

The US position: human authorship required

The U.S. Copyright Office has taken a clear stance: copyright protection requires human authorship. Works generated entirely by AI, without meaningful human creative contribution, cannot be registered.

The Copyright Office reports

The Copyright Office released a three-part report on "Copyright and Artificial Intelligence" across 2024-2025:

Part 2 addresses the copyrightability question directly.

What cannot be copyrighted:

• Works generated solely by AI lack the required human creative element
• Prompt selection alone does not constitute authorship, no matter how many prompts were used or how detailed they were
• The Copyright Office stated that "prompts may reflect a user's mental conception or idea, but they do not control the way that idea is expressed"

What can be copyrighted:

• Human modifications: substantial editing, arranging, or combining AI-generated materials in creative ways
• Expressive inputs: directing AI to modify existing human-created work where human elements remain "perceptible in the AI-generated output"
• Mixed works: human-authored portions of works containing AI-generated material

Registration requirements

When registering works that include AI-generated content, applicants must:

1. Disclose AI involvement in the application
2. Explain the human contribution
3. Identify which portions are human-authored vs. AI-generated
4. Disclaim AI-generated portions that lack human authorship

Failure to disclose AI involvement risks invalidating the registration.

Zarya of the Dawn

The 2023 Zarya of the Dawn decision set early precedent. Kristina Kashtanova created a graphic novel using Midjourney for images.

Protected:

• The human-written text
• The selection and arrangement of text and images together

Not protected:

• Individual images generated by Midjourney

The Office concluded that "it was Midjourney not Kashtanova that originated the 'traditional elements of authorship' in the images." Prompts function as instructions, not creative expressions. Active participation in prompting does not make users authors of AI output.

Thaler v. Perlmutter: AI cannot be an author

In March 2025, the D.C. Circuit Court of Appeals affirmed that AI cannot be an author under the Copyright Act. Dr. Stephen Thaler had sought to register a work listing an AI system as the author. The court rejected this, calling human authorship a "bedrock requirement" of copyright law.

The court did note that its ruling does not preclude protection for works created "by or with the assistance of" AI where human creative choices determine the expressive outcome. The distinction: AI as a tool is permissible; AI as an author is not.

Thaler petitioned the Supreme Court for certiorari in October 2025. The government's response was due January 26, 2026. The Supreme Court has not yet decided whether to hear the case.

The UK position: reports pending

The UK government is working through AI copyright reform via the Data (Use and Access) Act 2025, which received Royal Assent in June 2025. Section 137 mandates two reports due by March 18, 2026:

• An economic impact assessment of policy options
• A report on use of copyright works in AI development

Policy options under consideration

The UK government is evaluating four options:

Over 11,500 consultation responses showed overwhelming support for mandatory licensing (Option 1), though the government initially preferred Option 3. The tension between government preference and public response will shape the March 2026 reports.

AI-generated works in UK law

The UK government's preferred position is to remove copyright protection for computer-generated works without human authorship, aligning with US and EU approaches. No legislation has passed yet. The law remains unchanged pending the 2026 reports.

Getty v. Stability AI: model weights are not copies

In November 2025, the UK High Court ruled in Getty Images v. Stability AI that an AI model itself cannot constitute an "infringing copy" under UK copyright law.

The court found that Stable Diffusion's model weights (numerical parameters derived from training) "never contained a copy" of Getty's original works. Using copyrighted material during training does not make the resulting model an infringing copy.

This limits copyright exposure for AI companies that train models outside UK jurisdiction. Getty is appealing on secondary copyright infringement grounds.

The EU position: enforcement approaching

The EU AI Act entered into force in August 2024 with copyright provisions becoming enforceable in August 2025. A Copyright Directive review is mandated for June 2026, which could further address AI training on copyrighted works.

GEMA v. OpenAI: training liability

In November 2025, the Munich Regional Court ruled in GEMA v. OpenAI. GEMA, the German music rights organization, sued OpenAI for reproducing copyrighted song lyrics.

The court found:

• Memorization constitutes reproduction: The statistical probability values in model weights that enable later retrieval satisfy reproduction requirements under German copyright law
• Text and data mining exception rejected: Storing recognizable song lyrics exceeds the evaluative scope of the TDM exception
• OpenAI bears responsibility for selecting training data, designing architecture, and operating the system

The court ordered injunctive relief, mandatory disclosure, and damages. This was the first European decision holding that AI training using copyrighted works without authorization constitutes infringement.

OpenAI is appealing. A reference to the Court of Justice of the European Union is possible, which would create binding precedent across EU member states.

What this means for code

The same legal principles apply to code as to other creative works.

Purely AI-generated code: Not copyrightable under current US guidance. If an agent writes a function from scratch with no human creative input beyond the prompt, that function may lack copyright protection.

Human-modified AI code: Human contributions remain copyrightable. Substantial modifications, arrangements, or extensions by a developer can receive protection.

AI-assisted coding: Using AI tools for ideation, debugging, or enhancement does not disqualify human-written portions from protection. The AI functions as a tool, not an author.

Ownership vs. copyrightability

These are different concepts, and the difference matters.

Ownership: Who has the legal right to control and commercialize the code? Vendor terms typically assign ownership of outputs to the user. GitHub's FAQ states: "The code you write with GitHub Copilot's help belongs to you."

Copyrightability: Can the work receive copyright protection? This depends on human authorship, regardless of contractual ownership assignments.

Here's the catch: purely AI-generated content may lack copyright protection despite contractual ownership. You can "own" something that is not copyrightable but that ownership provides limited protection. Anyone could copy that code without infringing your copyright because there is no copyright to infringe.

Work-for-hire and AI

Can AI-assisted code qualify as work-for-hire? The Thaler v. Perlmutter decision addressed this.

Work-for-hire requires a binding legal contract (employment agreement or work-for-hire agreement). AI systems cannot enter into contracts. The court noted that work-for-hire language stating hiring parties are "considered" the author signals they are not actual authors distinguishing ownership from authorship.

For enterprises commissioning AI-assisted work: contracts should specify that employees or contractors provide sufficient human control, arrangement, or modification to ensure copyrightability. The threshold for "sufficient" remains unclear, but documenting human contribution matters.

Practical guidance for enterprises

Legal uncertainty does not mean paralysis. Enterprises can adopt defensive practices while the law develops.

Document human contribution. Track which code was AI-generated versus human-written. Preserve prompts, iterations, and refinements. Note where developers rejected AI suggestions and wrote code themselves. Document architectural decisions that shaped what the AI produced. This strengthens copyright claims for human contributions and demonstrates the development process if ownership is ever disputed.

Ensure meaningful human involvement. Substantial modification of AI output strengthens copyrightability claims. Have developers review, edit, and extend AI-generated code rather than accepting it verbatim. The more human creative input, the stronger the copyright position.

Leverage vendor indemnification. Microsoft's Copilot Copyright Commitment and Anthropic's enterprise indemnification provide some protection against IP claims arising from AI-generated outputs. These do not make AI-generated code copyrightable, but they shift infringement risk to the vendor within specified conditions. Indemnification typically excludes customer modifications, combinations with non-vendor technology, and beta features. Enable required safety features (duplication detection, filtering) to maintain coverage.

Monitor evolving law. The Supreme Court may take up Thaler v. Perlmutter. UK reports due March 2026 could reshape UK law. EU Copyright Directive review due June 2026. GEMA v. OpenAI appeal could establish EU-wide precedent. Policies adopted today may need revision as case law develops.

Assess copyright necessity. Not all code requires copyright protection. Open source projects may not need it. Internal tooling may not need it. Proprietary product code does. Match documentation rigor to actual protection needs.

Jurisdictional complexity

Enterprises operating globally face conflicting frameworks:

• US requires human authorship but allows AI as a tool
• UK is considering reforms with March 2026 reports pending
• EU is establishing training liability through cases like GEMA v. OpenAI
• Other jurisdictions vary widely

Code produced in one jurisdiction may have different legal status in another. For global enterprises, the most conservative applicable standard often becomes the de facto policy.

The ownership question will not be fully resolved soon. In the interim: document human contribution, leverage vendor protections, and monitor legal developments.