AI-generated article. This article was researched and drafted using AI tools and published automatically, and its featured image was generated by AI. Facts are drawn from the sources cited in the text.
Generate a logo, a header image and forty product descriptions in an afternoon, publish all of it, and a reasonable assumption follows: the business owns what it just made. That assumption is mostly wrong, and the gap between what people believe about AI generated content copyright and what the law currently says is wide enough to matter commercially.
The usual caveat, and it carries real weight on this topic: what follows summarizes published guidance and reported cases. It is not legal advice, and copyright is jurisdictional. Anyone with money riding on the answer should ask a lawyer in their own country.
Purely machine-made work is not owned by anyone
The United States Copyright Office has held consistently that copyright requires human authorship. Material generated purely by a model, with no meaningful human creative control over its expressive elements, does not qualify. That position hardened this year. On 2 March 2026 the Supreme Court declined to hear Stephen Thaler’s appeal over an AI-generated artwork, leaving the lower court rulings standing.
The practical consequence surprises people. A purely AI-generated image on a website is not protected, which means a competitor can lift it and use it, and there is no infringement claim to bring. The risk is not being sued. It is having nothing to enforce.
Where AI generated content copyright does apply
The Office has been clear that assistance is different from substitution. Human contributions to a work involving AI can be sufficient for authorship, assessed case by case, where a person exercised meaningful creative control over the expressive choices. Selecting, arranging, editing and substantially reworking output can qualify. The line it has drawn firmly is that prompting alone does not, however elaborate the prompt.
That is worth sitting with for anyone who has spent months refining prompts and reasonably regards that work as craft. Effort and copyright are not the same test. The question is not how hard the prompt was to write. It is how much of the expression in the finished piece traces back to human choices. It is also why a prompt library is better understood as an operational asset than as protected intellectual property.
Registration adds a duty most people miss. When registering a work that includes AI-generated material, applicants are expected to identify and disclaim the AI-generated portions. Claiming the whole thing without disclosure creates a problem in the registration itself.
The lawsuits are about the models, not their users
The largest development came in the litigation over training data. Anthropic settled a class action brought by authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson for $1.5 billion, over claims it trained on pirated books obtained from shadow libraries. It was the first case of that scale to close with real money.
For a business using these tools, the useful detail is where liability landed. It sat with the model developer, not with customers who used the product. That pattern holds across the other active cases, and there are many: the New York Times against OpenAI and Microsoft, authors against Meta, record labels negotiating with Suno and Udio, Disney and Universal against Midjourney. In the United Kingdom, the High Court in Getty Images v Stability AI rejected the theory that model weights are themselves infringing copies, while leaving questions about where training occurred unresolved.
None of that is settled law yet. What it has changed is procurement behavior. Buyers are now asking vendors for indemnification covering training-data claims, and the settlement gave them a number to point at when they ask.
What a small publisher should actually do
Two different exposures sit here and they pull in opposite directions. Being sued over AI output is currently a low risk for an ordinary business user. Owning nothing is a near certainty for anything published untouched.
For material that matters commercially, a logo, a signature illustration, anything that carries brand recognition, generated output is a weak foundation. Either commission it, or treat the generation as a first draft that a person then meaningfully reworks, and keep a record of what that person actually changed. For everyday content where nobody was ever going to copy it, the exposure is minor and the time saving is genuine.
Disclosure is the other half. The EU transparency rules that became enforceable in August 2026 ask a related but separate question. Copyright asks who owns it. The AI Act asks whether the reader was told. A business can comfortably satisfy one and fail the other.
The interesting shift is that “we made this with AI” has become two claims rather than one. It says something about how the work was produced, and something quite different about whether anything was created that the business can hold onto.
