When the Supreme Court declined to hear Thaler v. Perlmutter in March, headlines declared that AI content "cannot be copyrighted." That is not quite what happened, and the actual legal picture is messier in ways that matter.
The Thaler case involved a specific and unusual claim: that an AI system called DABUS should be named as the sole author of a creative work, with no human involvement at all. The courts said no to that, and the Supreme Court's non-intervention confirms it. The Copyright Act requires a human author. That much is now settled.
But "AI cannot be the sole author" is a long way from "nothing made with AI assistance can be protected." The Copyright Office has been consistent on this distinction for two years. Works where a human makes genuinely creative choices, selecting, arranging, directing, and refining AI output, can qualify for copyright protection on the human-authored portions. The question in any given case is always how much creative expression came from the human and how much was generated automatically. There's no bright line, and the Copyright Office has been reviewing applications on a case-by-case basis.
The cases that actually aren't settled
What the courts haven't resolved, and what will take years more to work out, is the training data question. Multiple lawsuits are still active. Anthropic settled a class action over training data for $1.5 billion in August 2025 without admitting liability. A court ordered OpenAI to produce 20 million output logs in January 2026, in a case that will turn partly on whether model outputs are substantially similar to copyrighted training material. Thomson Reuters won on summary judgment against an AI legal research startup on fair use grounds, a ruling that has made the legal AI sector rethink its data sourcing practices.
The training cases involve genuinely unsettled law. Courts are applying decades-old copyright doctrine to technology the doctrine was not designed for, and they're reaching different conclusions depending on the circuit and the specific facts. The Supreme Court's non-intervention in Thaler says nothing about whether training on copyrighted data is permissible. Those are separate questions, proceeding through separate cases.
What this means practically
A few things follow from the current state of play. Pure AI output with no meaningful human creative input cannot be copyrighted in the US. AI-assisted work with genuine human creative contribution probably can be protected, though how much contribution is "enough" remains fuzzy. The legality of training large models on copyrighted data is still live litigation, and outcomes are likely to vary by circuit and use case.
The situation is also evolving internationally. The EU, Japan, and the UK have taken meaningfully different positions on both authorship and training data, which creates real complexity for companies operating across jurisdictions.
One more thing worth noting: the same courts that are ruling on AI copyright are also sanctioning lawyers who use AI tools to generate case citations without verification. The judiciary is engaging with AI on multiple fronts simultaneously, and the resulting body of doctrine will be built incrementally, through cases with specific facts, not through any single landmark ruling.
The AI copyright situation is genuinely complex, and that complexity is doing real work. Anyone telling you it's been definitively settled in either direction is probably oversimplifying.
Sources: Holland & Knight, Baker Donelson, Norton Rose Fulbright
Commentarii · 0