References: Who owns AI's work

Source material

This lesson is an original adaptation. Its structure, why we protect ideas at all, the fair-use factors, the human-authorship principle, the three-candidate ownership question, and the copying-versus-style distinction, comes from the intellectual-property session of a Harvard Kennedy School course. The bucketing discipline, the current case statuses, the reader habits, and the prose are our own. Clawdemy is independent of Harvard, which has not reviewed or endorsed this track.

• The Science and Implications of Generative AI (HKS DPI-681M), Harvard Kennedy School, Spring 2024. Faculty: Sharad Goel, Dan Levy, and Teddy Svoronos. This lesson adapts Class 9 from the Spring 2024 course site, whose content is licensed under Creative Commons Attribution 4.0. Class 9 is the course’s intellectual-property session: it asks whether model training infringes copyright, whether AI outputs can be copyrighted, and whether AI outputs can infringe. Its four session videos are the source for our map of the debate.
• Official course lecture playlist on YouTube, Harvard Kennedy School. The full lectures, free to watch. We mean it when we encourage you to take the original course alongside this track: good teachers deserve more students.
• Provenance note: the framework and classroom examples in this lesson were drawn from the transcripts of the official Class 9 lecture videos in the playlist above, obtained from the official Harvard sources only. No third-party re-uploads or mirrors were used.

A note on how these cases are cited

This is the copyright lesson, so accuracy cuts both ways. Every legal claim below is placed in one of three buckets, and each carries the same limits stated in the lesson. A decided ruling is a fact of record from a single trial court, cited with its limits (trial-court level, fact-specific, appealable or already on appeal), never downgraded to merely “contested.” A settlement is reported as a fact, with no inference about who was right. A genuinely open question is marked “as of July 2026, unresolved.” This corner of the law moves fast, so a non-author reviewer re-verifies every case, date, and status live before publish and on each freshness sweep; this lesson is first in every sweep. None of this is legal advice.

Decided rulings, with limits

• Thomson Reuters v. Ross Intelligence (U.S. District Court for the District of Delaware, summary judgment February 11, 2025). The court held that a startup’s copying of a legal-research company’s material, used to build a competing search tool, was not fair use. Limits on record: the court expressly limited its reasoning to a non-generative AI tool, and the decision turned on its own facts. Current status: a federal appeals court, the Third Circuit, heard oral argument in 2026 and a decision is still to come. Sources: the opinion issued by the District of Delaware (docket 1:20-cv-00613); Authors Alliance analysis. The lesson keeps this to “a legal-research company” and “a startup” for readability; the material at issue was Westlaw editorial headnotes.
• Bartz v. Anthropic, the ruling (U.S. District Court for the Northern District of California, summary judgment June 2025). The court held that training an AI on books the company had lawfully bought was fair use and highly transformative, while downloading and keeping pirated copies of books was not fair use and was infringement. Limit on record: the fair-use-on-training holding applied to the named plaintiffs. Source: Authors Guild coverage. The lesson pairs this with the Kadrey ruling to show that trial courts have not agreed.
• Kadrey v. Meta Platforms (U.S. District Court for the Northern District of California, June 25, 2025). The court granted summary judgment for the AI company on the training question, but the judge stressed that he was ruling narrowly, that the plaintiffs had argued their case poorly, and that in most future cases he expected such training to be found infringing, not fair. This ruling has no final judgment and is not on appeal as of July 2026. Source: Norton Rose Fulbright analysis.
• Thaler v. Perlmutter (U.S. Court of Appeals for the District of Columbia Circuit, affirmed March 18, 2025; U.S. Supreme Court declined to hear the case on March 2, 2026). The courts held that, under the Copyright Act, an author must be a human, so an image its creator said a machine had made entirely on its own could not be registered. Limit on record: the case concerned a work claimed to have no human involvement at all, and it did not decide how much human input an AI-assisted work needs. Sources: the D.C. Circuit opinion (No. 23-5233), available through the D.C. Circuit; Skadden client alert. This is the same “creativity machine” case the course discusses, updated to its 2026 posture.

A settlement, reported with no merits inference

• Bartz v. Anthropic, the settlement. After the ruling above, Anthropic reached a deal to settle the piracy part of the case, agreeing to pay at least 1.5 billion dollars and to destroy its pirated copies. A class settlement like this needs a judge’s approval, and this one went before a court for that review in 2026. The release covers past conduct only, not model outputs or future conduct. Sources: Authors Guild, “What Authors Need to Know”, Copyright Alliance summary. Discipline that must survive: a settlement is a deal to end a case, not a court finding, and settling is not an admission of anything. No derivative hardens the figure past “at least 1.5 billion dollars,” and none infers wrongdoing from the settlement or implies a final approval that could not be confirmed as of July 2026.

Genuinely open questions

• The New York Times v. Microsoft and OpenAI (U.S. District Court for the Southern District of New York). The Times says the companies trained on millions of its articles without permission; the companies agree they used the articles and argue fair use. A judge let much of the case move forward in 2025 (a motion to dismiss was partly denied in April 2025), but there has been no ruling on the core question, no trial date, and no settlement as of July 2026. This was the course’s 2024 anchor example and it stays at status level in the lesson. Sources: case overview; AI Lawsuit Tracker.
• Allen v. Perlmutter (U.S. District Court for the District of Colorado). This is the AI-assisted line, exactly where the fight now sits. An artist made a prize-winning image with an AI tool after hundreds of prompts and revisions, was refused registration, and is challenging that refusal in court. Cross-motions were fully briefed in early 2026 before the judge, and there is no ruling as of July 2026. Sources: IPWatchdog coverage; background on the work, Théâtre D’opéra Spatial. The lesson omits the artist’s name and the work’s title for a cleaner read.
• Andersen v. Stability AI and Getty Images v. Stability AI (United States) (both U.S. District Court for the Northern District of California). Other suits, brought by visual artists and by a large image library, are working through the same California court. Both are in active pretrial with no merits ruling on copyright or fair use as of July 2026; the Getty motion to dismiss was decided in April 2026 (one ancillary claim was dismissed, and the copyright claim proceeds), and the Andersen matter has a trial date set for late 2026. Source: AI Lawsuit Tracker, Getty v. Stability.
• The general training question. Whether it is fair use to train AI on copyrighted work is, as a general matter, genuinely open. The individual rulings above are decided facts; the general question is open because those decisions are narrow, fact-specific, conflicting, and unreviewed by any higher court. The lesson never states, in either direction, that training is or is not fair use as a settled rule.

A non-US decision, quarantined

• Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch), UK High Court, judgment November 4, 2025. A related case in the United Kingdom reached a decision in late 2025 that mostly went against the image library. This ruling sits under a different country’s law, does not bind United States courts, and is itself being appealed. Source: Pinsent Masons analysis. The lesson carries this as one clearly labeled non-US line only.

Agency guidance, not binding law

• US Copyright Office, “Copyright and Artificial Intelligence, Part 2: Copyrightability” (published January 29, 2025). Framed throughout the lesson as the Office’s own registration practice, not a final word from the courts. Its position: typing prompts alone does not provide enough human control to make the user the author, because the same prompt can produce many different outputs; but a human’s creative selection, arrangement, and modification of AI output, and human-authored expression that shows up in the result, can be protected. Worked example: the Office registered the comic book Zarya of the Dawn as a compilation for the human selection and arrangement of AI images plus the human-written words, while the individual AI-generated images on their own were not protected. Sources: US Copyright Office AI initiative, and the Part 2 report. Limitations: this is agency guidance; the courts remain the arbiter.

The course’s own worked examples

• Authors Guild v. Google (Google Books) is the transformative-use precedent the course draws on when explaining the purpose factor: a court found that scanning books to make them searchable added something genuinely new. The lesson keeps the factor generic in the body and names the precedent here.
• The monkey-selfie case (Naruto v. Slater) is the course’s older worked example of the human-authorship rule: a photograph taken by an animal could not be copyrighted because there was no human author. The lesson trims it from the body for length; it makes the same point as the Thaler case.
• Heart on My Sleeve, the 2023 AI-made track that mimicked two famous musicians and was pulled from streaming services, is the course’s example for the copying-versus-style question. It is kept in the lesson as a dated 2023 historical marker.

On this site

• The risk map, lesson 6 of this track. It puts ownership on the four-category risk map as one square big enough to earn its own lesson; this lesson is that lesson.
• How models are pretrained, from Clawdemy’s AI Foundations track. How a model learns from an enormous pile of human writing and images. This lesson needs only that it happened; that lesson covers how.

Going deeper

• Official HKS course listing for DPI-681M, Harvard Kennedy School. The in-person course behind the open online materials this track adapts.

Adjacent topics

• The next lesson turns from who owns the work to who does the work. If a model can write, draw, summarize, and analyze, what happens to the people paid to do those things? For many readers that is the most personal question in this whole track, and the next lesson meets it with the same care this one brought to copyright.