AI Outputs: Understanding Copyright and Ownership

If you’ve generated an image, a paragraph, or a chunk of code with an AI tool this year, you’ve probably wondered whether it’s actually yours. The honest answer is: it depends, almost entirely, on how much of a human you put into it. That’s not a dodge — it’s the actual legal test the U.S. Copyright Office (USCO) and the courts have converged on, and it has real teeth.

This guide covers two separate legal fights that people often mix together:

1. Can you own the copyright in something an AI generates for you? (the output question)
2. Did the AI company have the right to train on copyrighted material in the first place? (the input question)

Both are live, both have money and legal precedent attached to them now, and both changed meaningfully in 2025 and 2026.

Part 1: Can you copyright an AI-generated output?

The short version, as of mid-2026: pure AI output — a prompt in, an image or paragraph out, nothing else — cannot be copyrighted in the United States. The more human creative work you layer on top, the stronger your claim gets.

The rule, from the source

The USCO’s position has been consistent since its first formal guidance in March 2023: copyright requires human authorship, and a work generated by AI in response to a prompt — no matter how detailed or iterated — doesn’t count as that human’s authorship, because the AI system, not the human, is the one making the expressive decisions (composition, wording, color, structure). The Office reaffirmed and elaborated on this in Part 2 of its Copyright and Artificial Intelligence report, published in January 2025, which walked through four categories of human/AI collaboration and where the line falls in each.

Three registration decisions illustrate where that line actually sits:

• Zarya of the Dawn (Feb. 2023): a graphic novel with Midjourney-generated illustrations. The Office denied protection for the images themselves but allowed copyright in the human-written text and in the human’s selection and arrangement of images and text into a book.
• Théâtre D’opéra Spatial (Sept. 2023): an AI-generated artwork the applicant then modified. Denied — because the applicant hadn’t properly disclosed and disclaimed which parts were AI-generated, as USCO policy requires.
• SURYAST (Dec. 2023): a human photo run through an AI style-transfer filter based on Van Gogh’s Starry Night. Denied, because the AI — not the human — decided how to blend the two images.

A similar case, Sahni, was rejected on the same logic in December 2023: the human’s original photograph and the AI’s stylistic transformation of it couldn’t be separated, so nothing was registrable.

What actually does qualify

Copyright protection attaches to the human-authored portion of a mixed work, not the AI-generated portion. In practice, this covers:

• Substantially rewriting or editing AI-drafted text in your own words
• Creatively selecting, arranging, or compositing multiple AI outputs into something original
• Making meaningful post-generation edits (hand-painting over an image, restructuring a draft, adding your own original passages)

One frequently misunderstood point: your prompt itself can be copyrighted as a piece of writing, if it’s creative enough — but that protects the prompt, not the image or text it produces. Those are legally separate works.

The Supreme Court weighed in — sort of

On March 2, 2026, the Supreme Court declined to hear Thaler v. Perlmutter, the case brought by AI researcher Stephen Thaler, who argued his AI system “DABUS” should be recognized as an author and that he, as its owner, should hold the resulting copyright. Declining cert leaves the lower court’s ruling in place: AI cannot be a legal “author,” and fully autonomous AI output without meaningful human creative input isn’t copyrightable. This didn’t overturn or expand USCO’s existing guidance — it confirmed it. What it explicitly did not resolve is the harder, more common case: AI-assisted work with real human creative input, which is still evaluated case by case.

The practical takeaway

If you want your AI-assisted work to hold up:

• Document your creative choices — which outputs you rejected, what you edited, what you added yourself. This is your evidence of authorship if it’s ever challenged.
• Disclose AI content when you register a copyright. Under USCO’s March 2023 policy, failing to disclose can invalidate your registration entirely under 17 U.S.C. § 411(b) and cost you the ability to sue for infringement later — a much worse outcome than just not owning the AI-generated slice.
• Do real post-generation work. A detailed prompt, by itself, isn’t authorship under current guidance — no matter how many iterations you ran.
• Treat “who owns this” as a spectrum, not a switch. A blog post you drafted with AI help and then substantially rewrote is on solid ground. A raw, unedited image-generator output is not protected at all — which also means, legally, anyone else can use it too.

Part 2: Did the AI get trained legally in the first place?

This is a separate battle, and it’s the one with actual dollar figures attached right now.

The industry’s core defense: fair use

AI companies have argued that training a model on copyrighted books, articles, and images is “transformative” fair use — the model learns statistical patterns, they say, rather than storing and redistributing the original text. Courts have been willing to accept part of that argument, but not all of it, and the details matter enormously.

Bartz v. Anthropic: the case that set the price

In June 2025, Judge William Alsup ruled that training an LLM on legally acquired books was fair use — but drew a hard line at how Anthropic acquired some of its training data. Anthropic had downloaded more than seven million pirated books; the court held that acquiring and storing pirated copies wasn’t fair use, even though the intended downstream use (training) might have been transformative. There was, in the court’s words, no “get out of jail free” card for piracy just because the eventual use was legal.

Rather than go to trial on statutory damages that could theoretically have run into the hundreds of billions of dollars, Anthropic settled for $1.5 billion in September 2025 — the largest copyright settlement in U.S. history to date. The deal, given final approval in 2026, pays out roughly $3,000 per book to an estimated 500,000 titles, and requires destruction of the pirated dataset. Crucially, the settlement covers past acquisition and storage of training data — it does not concede that AI outputs themselves infringe, and it doesn’t establish a licensing requirement for training generally.

Other cases testing nearby edges

Where this leaves AI companies — and their outputs

Several major labs have shifted strategy in response to this legal pressure: OpenAI has signed licensing deals with the Associated Press and other publishers; Google has licensing arrangements with Reddit and various news organizations. The clear signal from the Bartz settlement and the Ross ruling is that “train on everything scraped from the internet, sort out the legal risk later” is no longer a viable default posture for a company planning to operate at scale.

For everyday users, this litigation mostly matters indirectly: it shapes what training data future models are allowed to use, and it’s part of why some tools now restrict or filter outputs that closely resemble known copyrighted works (song lyrics being the clearest current example, thanks to the Concord Music suit).

Frequently asked questions