August 6, 2026. The Supreme Court declined to hear the one case that asked whether a machine can write. The lower courts have already answered. Every ground-level question about who owns what, in the age of the machine, now runs through the cases courts are still deciding.
On March 2, 2026, the Supreme Court denied certiorari in Thaler v. Perlmutter. That leaves standing the D.C. Circuit's ruling that the Copyright Act requires a human author. A machine alone cannot hold a copyright. The Office has said this for years. The courts have now said it at every level asked to hear the case.
Thaler was never the hard question, though. Dr. Thaler disclaimed all human creative input, and the case asked only whether a machine could be named the author. The court answered no, and doing so cost little, because nobody had to decide how much human involvement the law actually requires. That question was left to the next case.
The next case is here.
Jason Allen used Midjourney to make a prize-winning image, entered it with hundreds of iterative prompts, cleaned it up in Photoshop, and asked the Copyright Office to register it. The Office refused, saying the machine made the expressive choices. Allen sued. His case is now fully briefed for summary judgment in the District of Colorado.
The Office's argument is that prompts are unprotectable ideas. Run the same prompt twice, it says, and you get a different image. If you cannot predict or control the specific output, you are not the author of that output. Allen's answer is that specifying genre, tone, color, composition, and style through hundreds of iterations is not supplying ideas. It is making expressive choices, the same kind a photographer makes in framing a shot.
Allen's lawyers lean on the photography analogy for good reason. The Office has long held that creative selection and arrangement of a camera's output is authorship. The fight is whether steering a generative model toward a fixed outcome is meaningfully different from pointing a camera and choosing the moment.
No matter which way the court lands, the case matters. If prompts count as authorship, far more AI-assisted work becomes registrable. If they do not, creators who rely on generated material keep only the human portions they can isolate. Either way, this is the case that finally gives the line a shape.
The authorship question runs parallel to a second one: whether using copyrighted works to train a model is fair use. The early cases are converging on a single standard, and it is the market test. When a model's output competes with the licensed market for the works it was trained on, fair use gets harder to defend. When it does not, the defense survives.
The first trial court to reach the question, Thomson Reuters v. Ross Intelligence, used exactly that logic. Ross trained a legal research AI on Westlaw's headnotes, and the court rejected fair use partly because the product competed directly with Westlaw. That ruling is now on appeal to the Third Circuit, argued in June 2026, with a decision expected within months. It will be the first appellate-level ruling on AI training and fair use in the country.
Anthropic is fighting the same fight on two fronts. The music publishers moved for partial summary judgment in March 2026, asking the court to reject the fair use defense on training Claude on song lyrics. Separately, Anthropic paid $1.5 billion to settle the authors' suit over pirated books in its training set. The message for creators is direct. The more a model substitutes for the market that supports you, the stronger your claim.
On July 31, 2026, the Regional Court of Munich handed down the first significant ruling against a music generator. GEMA, the German collecting society, sued Suno over stream-ripped songs in its training data. The court found that Suno's model itself contains copies of GEMA's repertoire, which takes the case outside the European text-and-data-mining exception. It also found the model's outputs reproduce GEMA's works, and it ordered Suno to disclose its income from the infringement.
The ruling is appealable and is not binding in the United States. It matters anyway. It is the first time a court has looked inside a music model, found the works at issue embedded in the weights, and acted on it. The American cases have spent years arguing over outputs and market harm. Munich found the copies in the model itself.
Here at home, the landscape splits in two. Warner settled with Suno and Udio. UMG settled with Udio. Sony holds the line, and its fair-use cases against both companies are now scheduled for dispositive motions in April 2027. The independent artists whose songs trained these models were offered nothing. The class actions run by Hagens Berman and Delgado Entertainment Law are the only path left to them, and those will take years.
Read together, the developments point one way. Human-authored work keeps full protection, and AI assistance in the creative process does not erase the human portions. The cases now being decided will determine how much of the machine's output a human can claim, and whether the companies that trained on unlicensed work owe the people who made it.
The practical steps have not changed, but the reasons to take them have sharpened. Register your work; it is the gatekeeper for statutory damages. Document your catalog and its publication dates. Check whether your music is in the training datasets, and if you find it, keep the record. The class action windows for works published to streaming platforms since January 1, 2021 are open, and the eligibility rules are gathered on our Class Actions page.
And watch the Third Circuit. Its ruling in Thomson Reuters v. Ross Intelligence will set the fair use rule for the whole country. The authorship line is holding. The training line is next.
Beaumont & Sheridan is an information resource for individual creators. This summary is based on publicly available court filings and press reporting and does not constitute legal advice.