Who owns what AI makes?

Copyright and AI is genuinely unsettled, and anyone who tells you it’s simple is selling something. The law varies by country and is still being decided in courts and parliaments. The useful thing isn’t a definitive answer — it’s knowing where the open questions are so you don’t get caught out.

The two live questions

• The inputs. Models were trained on large amounts of existing work. Whether that training was permitted is being fought over in multiple jurisdictions, with no settled global answer.
• The outputs. Who, if anyone, owns what a model generates? Several places hold that purely machine-generated work may not attract copyright the way human authorship does — which has real consequences if you need to protect what you create.

Where things have got to

Still unsettled, but less shapeless than a year ago. Two things have firmed up:

• Human authorship remains the requirement in US law for a work to attract copyright. That position was reaffirmed rather than disturbed in 2026. If you need to own what you produce, your own creative contribution is what secures it.
• Training is not automatically infringement, nor automatically fair use. The direction in US case law treats training a general-purpose model as substantially transformative, which helps a fair-use argument, but courts are deciding case by case on whether the output competes with the market for the original. The UK produced its first substantive judgment on AI training and copyright in late 2025.

Live litigation involving most major providers continues, and none of it is finished. Expect the question to keep moving from “was the training lawful?” toward “does this particular output substitute for the original?”

Staying on safe ground

• Check the tool’s terms for what rights you’re granted over outputs — they differ.
• Don’t assume outputs are clear of others’ rights. A model can reproduce something close to existing protected work; review before publishing commercially.
• Add real human authorship — your judgement, edits, and arrangement — both for quality and because it strengthens your position.
• Disclose where it matters — see disclosing AI use.

When the stakes are commercial or legal, treat AI output as a draft to clear, not a finished asset you automatically own. For anything significant, get proper legal advice for your country.