Can Generative AI Outputs Infringe U.S. Copyright?

Yes. Generative AI outputs can infringe U.S. copyright when they reproduce protected expression without permission or a valid defense. Human authorship determines whether an output can receive copyright protection, but it doesn’t determine whether that output violates someone else’s rights.

AI copyright infringement therefore creates two separate concerns: whether you can use generated content and whether you can stop others from copying it. The starting point is what the output reproduces, not simply which tool created it.

Key Takeaways

• An AI-generated image, song, script, or article can infringe even if the output itself lacks copyright protection.
• Copyright protects original expression, not ideas, facts, or artistic style alone.
• Training, output infringement, and ownership require separate legal analyses.
• Fair use depends on the particular use and evidence. A platform’s commercial-use permission doesn’t clear third-party copyrights.

When Does an AI Output Infringe Copyright?

Under Section 106 of the Copyright Act, copyright owners control reproduction, adaptation, distribution, and certain performances and displays. Generative AI doesn’t create an exception to those rights.

The output must copy protected expression

An infringement claim generally requires ownership of a valid copyright and copying of protected elements. Those elements can include original prose, musical composition, or visual expression.

However, resemblance alone doesn’t settle the question. Copyright doesn’t protect a general concept, historical fact, genre, or artistic style by itself. An output can resemble an artist’s approach without necessarily copying protected expression.

Conversely, changing the medium or replacing some details doesn’t automatically eliminate infringement. An unauthorized adaptation can implicate the copyright owner’s derivative-work right.

Access and substantial similarity matter

A plaintiff can establish copying through direct evidence or through circumstantial evidence involving access and similarities. Courts then assess whether the output copies protected expression under the applicable legal test.

Evidence that a model encountered a work can support the access inquiry. It doesn’t establish that every generated response infringes that work.

There is also no universal percentage of acceptable copying. Courts consider the nature and significance of the expression taken, not merely a word count or similarity score.

What AI Copyright Infringement Cases Have Decided

Court decisions must be read according to the question presented and the procedural stage. Allowing a lawsuit to proceed isn’t a final finding of infringement.

Tremblay v. OpenAI: output claims survived dismissal

In an October 27, 2025 ruling, the court allowed output-based direct-infringement claims in Tremblay v. OpenAI to proceed. Some alleged outputs could support a reasonable jury’s finding of substantial similarity.

That ruling recognized a legally plausible claim concerning generated content. It didn’t establish that OpenAI ultimately infringed, and it didn’t resolve whether the challenged outputs qualified as fair use.

Justice v. Suno: plausible allegations still matter

In a decision reported August 20, 2026, the court denied dismissal of an output-related claim in Justice v. Suno.

The court required plausible, non-speculative allegations that substantially similar outputs existed. However, it didn’t require plaintiffs to identify particular output files at the dismissal stage.

Again, permitting further litigation didn’t establish that any particular song infringed.

Bartz v. Anthropic: training was a different question

The June 23, 2025 Bartz v. Anthropic order held that the training use before the court qualified as fair use. It treated the acquisition and retention of pirated books as a separate issue.

The ruling didn’t decide whether particular Claude outputs infringed. Businesses negotiating AI training licenses should therefore distinguish permission to train from rights concerning generated content.

Copyright Ownership Is a Separate Question

A work can lack copyright protection for its generated elements while still reproducing someone else’s protected material. That distinction matters when commissioning content or promising clients exclusive rights.

Human authorship controls copyright protection

In Thaler v. Perlmutter, the D.C. Circuit affirmed the denial of registration on March 18, 2025. The applicant had identified the work as autonomously generated by AI.

The decision addressed authorship, not infringement. It also didn’t establish that using AI anywhere in a creative process defeats copyright protection.

Human-written material and sufficiently creative human modifications can remain protectable. The analysis concerns what the person actually contributed.

Prompts and platform terms don’t establish every right

The Copyright Office’s January 2025 report on AI copyrightability distinguishes prompting from human control over expressive elements. Prompts alone generally don’t supply sufficient authorship, although creative selection, arrangement, or modification may qualify.

Likewise, a platform’s assignment of output rights governs the contractual relationship. It can’t create federal copyright where human authorship is absent.

Permission to use an output under platform terms doesn’t establish permission to reproduce third-party material contained in that output.

Your agreement with a designer, writer, or production vendor should address both contractual ownership and the actual scope of copyright protection.

Does Fair Use Protect AI-Generated Content?

Fair use can excuse an otherwise infringing use, but AI generation doesn’t automatically establish that defense. Section 107’s four-factor test examines purpose and character, the nature of the source work, the amount taken, and market effect.

For an output claim, the analysis concerns the challenged output and its use. A training-related fair-use ruling doesn’t automatically authorize publishing a generated passage, image, or recording.

A court may consider whether the use adds a different purpose or character. It also examines whether the output substitutes for the original or harms relevant licensing markets.

Commercial use weighs into the analysis but doesn’t automatically defeat fair use. Similarly, nonprofit use doesn’t automatically qualify.

There is no fixed safe amount that a user can copy. Taking a short but important portion can matter more than taking a longer, less distinctive passage.

Crediting the original creator doesn’t replace this analysis. Attribution and permission answer different questions.

Who Can Be Liable for an Infringing Output?

A user who publishes, distributes, or commercially exploits infringing material can face a copyright claim. An agency or publisher doesn’t receive immunity merely because software supplied the content.

However, liability isn’t automatic for everyone connected to an output. Courts examine each defendant’s conduct, including who caused the copying and which exclusive rights the challenged activity implicated.

Claims against developers and platforms may involve direct infringement or secondary-liability theories. Those theories have different requirements, including questions about knowledge, contribution, control, and financial benefit.

Lack of deliberate copying also doesn’t automatically defeat a direct-infringement claim. Intent can affect remedies without deciding whether infringement occurred.

Contracts can allocate defense obligations, indemnity, and financial responsibility between parties. They don’t bind a copyright owner who wasn’t part of the agreement.

For businesses, this makes vendor terms important even when the final content looks original. Responsibility for clearance should be explicit before publication.

Reducing Risk Before Publication or Enforcement

A practical review should address source rights, generated content, and the human contributions to the finished work. Keep those questions separate throughout production.

For businesses and creators using AI

Retain prompts, source uploads, relevant platform terms, and versions showing human edits. These records help establish what entered the process and how the final work developed.

Review recognizable passages, melodies, characters, and visual elements before release. Similarity-detection tools can flag concerns, but they don’t decide substantial similarity or fair use.

Contracts should distinguish permitted assistance, such as transcription or proofreading, from generated creative material. Avoid warranties stating that no AI was used when the workflow included approved AI tools.

Also address confidential uploads, training permissions, third-party assets, and responsibility for claims. Broad ownership language doesn’t resolve these issues.

For copyright owners challenging an output

Preserve the disputed output, its publication location, and the circumstances of access. Retain your original files, publication history, registration information, and relevant licenses.

A useful comparison identifies the protected expression allegedly copied. General complaints that an output resembles your style provide a weaker foundation.

Our copyright registration and enforcement services address ownership, registration, licensing, and infringement disputes. At ChaseLawyers®, we help creators and businesses assess AI-related claims, document human contributions, and negotiate agreements that assign responsibility clearly.

Registration applications should also identify human-authored contributions and exclude more-than-minimal AI-generated material where required. Overstating authorship can put a registration at risk.

Frequently Asked Questions

Does crediting the creator or adding an AI label prevent infringement?

No. Credit identifies a source, while an AI label describes the production process. Neither grants permission to copy protected expression.

Disclosure may still matter under contracts, platform rules, or other laws. However, those obligations are separate from copyright clearance.

The relevant copyright questions remain whether protected expression was copied and whether permission or a defense authorizes the use.

Is cloning someone’s voice always copyright infringement?

No. A person’s voice or identity isn’t itself a copyrighted work. However, copying protected material from an existing recording can raise copyright issues.

Unauthorized impersonation may also support publicity-rights or false-endorsement claims, depending on the facts and applicable law. Our guidance on deepfake takedown rights addresses those separate legal routes.

Identifying the correct right matters because a copyright demand shouldn’t substitute for an identity-based claim.

Protect the Work and Clear the Output

Generative AI outputs can infringe U.S. copyright, even when their machine-generated elements cannot receive protection. The central question is whether a particular output copies protected expression without authorization or a valid defense.

Keep ownership, training, and output clearance separate. Before a release or enforcement demand, a focused legal review can identify the rights involved and the evidence needed to support your position.