AI Regulation Tracker  /  Agency and courts

US Copyright Office Flags AI Training as Infringement Risk; Human-Authorship Rule Holds

The Copyright Office's pre-publication Part 3 report treats training generative AI on copyrighted works as raising prima facie infringement questions absent a license or fair use. Separately, the Supreme Court let stand the rule that purely AI-made work cannot be registered.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

The US Copyright Office has put developers and creative firms on notice that training generative AI on copyrighted works is not a settled question. In the pre-publication version of "Copyright and Artificial Intelligence, Part 3: Generative AI Training," the Office frames the use of protected material to train a model as conduct that can raise prima facie infringement questions unless the developer holds a license or can establish fair use. The report is not yet final. The Office released the pre-publication Part 3 report (May 9, 2025) and has said a final version will follow; the report itself is a 2025 document, and the 2026 news peg here is the separate Supreme Court cert denial.

That framing matters because it puts the burden where infringement law usually puts it. Reproduction of a copyrighted work during training is treated as a use that must be justified, not as an automatically permitted step. Fair use remains available as a defense, but it is fact-specific and decided case by case, not a blanket exemption for model training. For any firm building or fine-tuning models on scraped or licensed corpora, that is the difference between a compliance assumption and a live legal exposure.

What the report says, and what it does not

The report is guidance from the Copyright Office, not a statute or a court judgment. It does not declare that all AI training infringes, and it does not resolve the pending litigation testing fair use in specific factual settings. What it does is lay out the Office's analysis of when unlicensed training implicates the exclusive rights of copyright owners and why a license or a fair use showing is the path to safety. Treat it as the Office's considered position, informative to courts and to your own risk assessment, rather than as a binding rule that ends the debate.

Because it is a pre-publication version, the document can still be cited for the Office's reasoning, but firms should track the final release expected in 2026 before relying on any particular phrasing as settled.

The human-authorship line held

On the registration side, the older rule got firmer. According to reporting, the Supreme Court denied certiorari on March 2, 2026 in the line of cases over whether output with no human author can be registered. The denial leaves in place the rule, developed through the Copyright Office and the lower courts, that a work created entirely by a machine with no human authorship is not eligible for copyright registration. A cert denial is not a ruling on the merits, and it sets no new nationwide precedent of its own, but its practical effect is to leave the existing no-registration rule undisturbed.

For working professionals, the two developments point the same direction. Human contribution is what the copyright system rewards. Prompt-only output, where a person types instructions and the model produces the work, sits on the wrong side of that line.

What creative and agency shops must actually do

If your firm registers works that were made with AI tools, the operational duty is disclosure. You must identify the AI-generated portions of a work when you register it, and you may claim copyright only in the material a human authored, such as the selection, arrangement, and human-created elements. You cannot claim copyright in output that came from prompts alone. Overclaiming on a registration is not a technicality. A registration that misstates authorship can be challenged, which undermines the very protection you filed to secure.

Developers face the mirror-image duty. The Part 3 framing means that unlicensed training on copyrighted works is a question you should be able to answer, either with a license or with a fair use position you are prepared to defend. Silence is not a strategy the Office endorses.

The near-term watch items

Two dates anchor the near term. The final version of Part 3 is expected in 2026, and firms building AI compliance around the report should confirm the final text when it publishes. The cert denial, reported as March 2, 2026, closes off one route of appeal but leaves the broader fair use fights to continue in the trial courts. Neither event is the last word, and both reward firms that document their licensing and authorship decisions now rather than after a dispute.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.