AI and Intellectual Property: A 2026 Guide

Generative AI can draft a contract, paint a mural, clone a voice, or write a song in seconds. That speed has outrun the law, and the result is a lot of confusion about who owns what. This guide walks through the questions people actually search for, in plain English, and links to deeper guides on each topic. It is general education, not legal advice, and it does not create an attorney-client relationship. For anything specific, talk to an attorney licensed in your jurisdiction.

Can AI-generated work be protected by copyright?

In the United States, copyright protects original works “of authorship,” and both the U.S. Copyright Office and the courts read “authorship” to mean a human author. That single requirement drives almost everything else.

The clearest ruling came in Thaler v. Perlmutter. Computer scientist Stephen Thaler tried to register an image, “A Recent Entrance to Paradise,” that he said was created autonomously by an AI system he called the Creativity Machine, listing the machine as the sole author. In March 2025 the U.S. Court of Appeals for the D.C. Circuit affirmed that human authorship is a “bedrock requirement” and that a machine cannot be an author under the Copyright Act. In March 2026 the Supreme Court declined to hear the case, leaving that rule in place.

The Copyright Office reached the same place in its 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability. Its key conclusions:

• Purely AI-generated material is not copyrightable. A work that a machine produces on its own, with no meaningful human creative control, cannot be registered.
• Prompts alone are usually not enough. Even long, detailed prompts generally do not, by themselves, make the resulting output copyrightable, because the human is not controlling the specific expressive choices the AI makes.
• AI as a tool is fine. Using AI to brainstorm, edit, or assist does not poison an otherwise human-authored work.
• Mixed works get mixed treatment. Where a human meaningfully arranges, selects, or modifies AI output, or weaves it together with human-made content, the human-authored parts can be protected even though the raw AI output is not.

Real decisions show the line. The Office allowed a registration for the comic book Zarya of the Dawn but limited it to the human author’s text and the creative arrangement of the images, not the individual Midjourney-generated pictures. And it refused to register Théâtre D’opéra Spatial, the AI image that won a Colorado State Fair art prize, because the human’s role in generating the picture was too limited even after hundreds of prompts.

The practical takeaway: keep evidence of your own creative contribution, and do not assume a registration certificate will cover the machine-made parts. For a deeper walkthrough, see Can you copyright AI-generated art?.

Who owns the output of an AI tool?

“Can it be copyrighted?” and “who is allowed to use it?” are different questions, and people constantly mix them up.

Even when the purely AI-generated output is not protected by copyright (meaning no one has an exclusive copyright in it), the AI company can still control your access to it through a contract: its terms of service. Those terms typically decide:

• whether you may use the output commercially at all, sometimes only on paid tiers;
• whether the company assigns output rights to you or merely grants a license;
• whether the company keeps the right to use your prompts and outputs to improve its models;
• what indemnification (legal cover) the company offers if a third party claims the output infringes their rights.

Because uncopyrightable output may not be exclusively “owned” by anyone, two different users could receive very similar results, and you may have limited power to stop someone else from copying yours. That is a business risk worth understanding before you build a brand around AI output. Read the terms for each tool, save a dated copy, and match the plan tier to how you actually intend to use the work. More detail lives in Who owns the output of an AI tool?.

The training-data copyright fights

The loudest legal battle is about the inputs: was it lawful for AI companies to copy huge amounts of text, images, and code to train their models? The honest answer in 2026 is that the law is unsettled, and early rulings cut in different directions.

• Bartz v. Anthropic (N.D. Cal., June 2025). Judge William Alsup found that training a large language model on books was “exceedingly transformative” and a fair use, but separately held that downloading and keeping a library of pirated books was not fair use. The parties later reached a large settlement over the pirated copies.
• Kadrey v. Meta (N.D. Cal., June 2025). Judge Vince Chhabria granted summary judgment to Meta on the records before him, but pointedly warned that the authors had simply failed to prove market harm, and that in many future cases plaintiffs who build a better record on harm could win. So this was not a green light.
• Thomson Reuters v. Ross (D. Del., February 2025). The court rejected fair use where a startup used Westlaw editorial headnotes to build a competing legal-research tool, partly because of harm to a potential AI-training-data market. An appeal is pending.
• Andersen v. Stability AI; Getty Images v. Stability AI. The artists’ case in Andersen is moving toward a 2026 trial. Getty’s U.S. case was sent to private dispute resolution, while a parallel UK case largely failed on technical copyright grounds in late 2025.
• The New York Times v. OpenAI. Still in active discovery in 2026, with high-profile fights over access to ChatGPT logs. No fair-use verdict yet.

Two themes emerge: courts treat transformative training more favorably than verbatim copying or building a direct competitor, and using pirated sources is a serious problem regardless of the eventual fair-use ruling. None of this is the final word. For the current state of play, see Is it legal to train AI on your work? and our copyright case archive.

Using AI content commercially

If you want to put AI-assisted images, music, video, or text into a product, an ad, or a logo, stack three checks on top of each other.

1. The tool’s terms. Confirm the license actually allows commercial use for your plan, and whether any attribution is required.
2. Copyright ownership. Remember that the purely AI-generated portion may not be protectable, which affects your ability to stop copycats and may affect deals where a buyer expects you to own clean rights. Add your own human creative work to strengthen what you can claim.
3. Third-party rights. AI output can accidentally reproduce someone else’s protected material, a recognizable trademark, or a real person’s likeness. You can be liable for that even if you did not intend it. Brands and logos are an especially common trap.

For a checklist-style version, including the trademark and right-of-publicity traps, see Using AI images and music commercially.

AI, voice, and likeness (deepfakes)

When AI imitates a specific person, copyright usually is not the main tool, the right of publicity is. That is the right to control the commercial use of your name, voice, image, and likeness. Many states protect it by statute, common law, or both, and a deliberate AI imitation of someone’s voice or face used to sell or promote something can violate it.

California is a useful example. Its Civil Code section 3344 protects name, voice, signature, photograph, and likeness, with statutory minimum damages, and section 3344.1 extends protection after death. The state also added AI-specific rules (discussed below). Other states differ widely in what they protect and for how long. Defamation, false-endorsement, and even fraud laws can also come into play with deepfakes.

Two big caveats: protections vary enormously by state, and genuine news, commentary, parody, and other protected speech are treated differently from commercial exploitation. For the mechanics, see AI voice cloning and deepfakes, our right-of-publicity case archive, and the California right-of-publicity guide.

The new digital-replica laws

Lawmakers have responded to deepfakes with a fast-growing set of digital-replica laws.

At the state level, several states have enacted statutes targeting unauthorized AI replicas. California’s AB 1836 (effective 2025) restricts using AI-generated digital replicas of deceased performers in audiovisual works and sound recordings without consent of the estate, and AB 2602 addresses replica clauses in performers’ contracts. Tennessee’s ELVIS Act expanded that state’s older right-of-publicity law to expressly cover an individual’s voice and AI imitations. Expect more states to follow, with meaningful differences among them.

At the federal level, the most-discussed proposal is the NO FAKES Act, which would create a federal right against unauthorized digital replicas of a person’s voice or likeness. It is important to be precise here: as of June 2026, the NO FAKES Act has been introduced and revised in Congress but has not been enacted. It is a bill, not a law, and it has drawn free-speech objections that could still change or stall it. Do not rely on it as if it were in force. For where things actually stand, see The NO FAKES Act and digital-replica laws.

What businesses should put in contracts

Because so much here is unsettled, contracts are where you manage the risk. A few clauses worth raising with an attorney licensed in your jurisdiction:

• Disclosure of AI use. Require contractors and employees to disclose when deliverables were created with AI, so you know what may not be protectable.
• Ownership and assignment. Assign whatever rights can be assigned, and acknowledge that purely AI-generated material may not carry an enforceable copyright.
• Reps and warranties. Have vendors represent that their AI tools and training did not infringe third-party rights and that outputs do not include someone’s likeness or another party’s protected content.
• Indemnification. Decide who absorbs the loss if an AI output triggers an infringement or right-of-publicity claim.
• Confidentiality and trade secrets. Prohibit pasting sensitive or confidential information into public AI tools, which can waive trade-secret protection.
• Right-of-publicity consents. Get written consent before using anyone’s voice or likeness in AI-generated content.

A starting framework lives in AI clauses for your contracts.

The bottom line

AI does not rewrite intellectual property law so much as collide with it: U.S. copyright still needs a human author, tool terms of service usually govern who can use the output, the training-data fair-use question is genuinely unsettled, and deepfakes are handled through the right of publicity plus a rising tide of state digital-replica laws (with the federal NO FAKES Act still only a proposal). Document your human creativity, read the fine print, and put the risk in your contracts.

This article is general legal education, not legal advice, and does not create an attorney-client relationship. For guidance on your specific situation, consult an attorney licensed in your jurisdiction.