AI and copyright — who actually owns what an AI creates?

Every time you ask an AI to write something, generate an image, or create any original-seeming output, you are operating in legal territory that courts and legislators are still actively mapping. The questions are not hypothetical. They are being litigated right now, in multiple jurisdictions, with outcomes that will affect everyone who creates content professionally.

Here is the current state of the three most important unresolved questions.

Question one: did AI companies have the right to train on your work?

Every major AI model was trained on data scraped from the internet — text, code, images, audio. Most of that data was created by people who did not consent to its use as training material and were not compensated for it.

The legal question is whether this constitutes copyright infringement or falls under fair use doctrine. The answer varies by jurisdiction and is not settled. In the US, several cases are pending. The most significant ongoing cases involve major news publishers (the New York Times vs OpenAI), book authors (multiple class actions vs Anthropic and Meta), and visual artists (vs Stability AI and Midjourney).

A US ruling in early 2025 found in one case that certain training uses could qualify as fair use, while another ruling went the other way on different facts. The legal landscape is genuinely unsettled, with different courts reaching different conclusions on similar fact patterns.

Question two: who owns the output of an AI?

This is cleaner in some jurisdictions than others. The US Copyright Office’s current position is that purely AI-generated output — content created by AI without meaningful human creative input — is not eligible for copyright protection. You cannot own the copyright in something a machine generated without your authorship.

The nuance is in “meaningful human creative input.” If you write a detailed prompt, make selections among AI outputs, edit and arrange the results, and integrate them into a larger original work — a human copyright claim over the resulting work becomes more defensible. The more creative decisions a human makes in the production process, the stronger the copyright basis.

Practically: if you are producing commercial work using AI and you want to hold copyright over it, document your creative choices. The prompts, the selections you made, the editing you did, the original elements you added. The paper trail strengthens your claim if the question is ever contested.

Question three: what is the liability if AI output copies protected work?

AI models trained on large datasets sometimes reproduce fragments of training data in their outputs. This happens rarely and usually incompletely, but it has been documented and it creates potential downstream liability.

Most major AI companies have implemented indemnification programmes: if you use their commercial API or product and are sued for copyright infringement based on AI-generated output, the company will cover your legal defence under defined conditions. OpenAI, Microsoft, and Google all have versions of this programme. Anthropic has one for Claude commercial API users.

The conditions matter. These programmes generally apply when you are using the product within its terms of service, not actively trying to reproduce specific copyrighted works, and have not made public claims about the provenance of the output that create separate liability.

What creators need to know today

For people generating commercial content with AI, the practical risk level is low but not zero. The most common real-world risk is not a copyright lawsuit over AI-generated text — it is misunderstanding what rights you hold over AI-generated images when you try to license them, or what rights a client holds over work you delivered that was substantially AI-generated without disclosure.

The disclosure question is increasingly important. Some clients include AI provisions in their contracts. Some stock photography platforms reject AI-generated images. Some publishers have explicit policies. Understanding what the other party expects is a practical step that costs nothing and prevents problems.

The law is moving. The direction of movement, across most major jurisdictions, is toward greater disclosure requirements, some form of compensation mechanisms for training data, and clearer ownership rules for human-AI collaborative works. The specifics are two to five years away from resolution. Operating with awareness of the uncertainty — rather than assuming either that AI output is freely usable or that it is completely off-limits — is the grounded position.

Related reading

• The Deepfake Problem — What AI-Generated Media Is Doing to Trust
• The 7 Biggest AI Moments of 2025 That Changed Everything
• Will AI Replace Writers? An Honest Answer