AI for UX · Step 3 · 6 min read
Copyright and AI-generated assets
What designers need to know in late 2026 about owning and using AI-generated work: when output can be protected, what recent court cases decided, what the EU requires of model makers, which vendors offer legal protection, and how to keep a safe record of your own creative contribution.
Two questions sit behind every AI-generated asset. Can anyone own it, including you or your client? And could using it infringe someone else's rights? The answers depend on how much of the work is yours, which tool you used, and where you are. This note gives you the map.
What you'll be able to do after reading:
- Explain when AI-assisted work can be protected by copyright, and when it can't
- Summarise the court cases and rules designers hear about most
- Check whether a tool's terms, and any legal protection it offers, cover your use
- Keep a simple record of your human contribution for work that matters
1. Can AI output be owned?
In the US, the Copyright Office's January 2025 report is the clearest guidance. AI output is protected only "where a human author has determined sufficient expressive elements"; prompts alone are not enough. What can be protected: your own expressive input that stays visible in the result, and your creative selection, arrangement or modification of AI material (US Copyright Office).
The courts agree. In Thaler v. Perlmutter, the US appeals court ruled that copyright requires a human author, and in March 2026 the Supreme Court declined to hear the case, so that ruling stands (Supreme Court docket 25-449).
- 1Prompt only
You typed a request; the machine made every expressive choice.
Not protected (US) - 2Your input visible
Your own sketch, photo or text stays recognisable in the result.
Your part can be protected - 3Selected and arranged
You chose, combined and ordered AI pieces creatively.
The arrangement can be protected - 4Substantially modified
You reworked the AI output with your own expressive changes.
Your changes can be protected
Other countries take their own positions; if ownership matters (a logo, a brand mascot, a book cover), ask a lawyer where you'll use it.
2. Could using it infringe someone else's rights?
This is about how models were trained and what they output.
| Case or rule | What it decided | What it means for you |
|---|---|---|
| Bartz v. Anthropic (US) | A US$1.5 billion settlement, about US$3,000 per book, for books obtained from pirate libraries; final approval July 2026 (settlement site) | How training data was obtained matters, and it can be expensive |
| Getty Images v. Stability AI (UK, Nov 2025) | The main copyright claim failed: a model that doesn't store or reproduce the works isn't an "infringing copy". Getty won only a narrow trade-mark point about generated watermarks (judgment) | Outputs that reproduce marks or recognisable works are where the risk is |
| US Copyright Office, training report (May 2025) | Analysis of when training may or may not be fair use; still a pre-publication version (USCO AI) | The question is not settled in the US |
| EU: AI Act and the Code of Practice | Since 2 Aug 2025, makers of general-purpose AI models must have a copyright policy, respect machine-readable opt-outs and publish a summary of their training content (Commission Q&A; Code of Practice) | Obligations fall on model makers, not on you as a user |
| EU text-and-data-mining opt-out | Commercial mining is allowed unless the rightholder reserved it, for example in machine-readable form (DSM Directive, Art. 4) | Publishers can opt their content out of AI training |
The practical risk for a designer is rarely the training debate; it's output that looks too much like something specific: a recognisable character, a living artist's signature style, a brand's logo, a real person's face.
3. Legal protection from vendors (indemnities)
Several vendors promise to defend business customers if their AI output is accused of infringing copyright. All have conditions, and none covers free or personal plans:
Commercial Copilot and Azure OpenAI, if you use the built-in guardrails.
Since Sep 2023Listed Google Cloud and Workspace services: training data and output.
Since Oct 2023API and business plans; not if you knew it infringed or disabled safety features.
Beta features excludedListed features, for customers whose agreement includes it.
Mostly enterpriseRead the conditions: most require using the tool's built-in safety filters, and none protects you if you knew, or should have known, that the output infringed. Partner models inside a tool may not be covered (see AI for visual design: images and video).
4. Faces, voices and likeness
- In the EU, realistic AI images, audio or video of real people, places or events (deepfakes) must be labelled as AI-generated since 2 August 2026 (AI Act, Article 50), and AI that generates non-consensual intimate imagery is banned from 2 December 2026.
- In the US, there is no federal law on digital replicas yet; the Copyright Office recommended one in 2024 (USCO AI). Many states have their own rules.
- Everywhere: don't generate real, identifiable people (customers, colleagues, celebrities) without their consent, and don't present generated people as real customers or testimonials.
5. A record of your contribution
For assets whose ownership matters, keep a short record. It costs minutes and can matter later:
- The tool, model and plan you used, and the date
- Your prompts and reference images, and where the references came from
- What you selected, changed and combined: sketches, edits, layers, final choices
- The licence terms of the tool at that date (save a copy)
6. Vocabulary
| Term | Meaning |
|---|---|
| Copyright | The right to control copies of an original creative work |
| Fair use / exceptions | Legal room to use protected works without permission in some cases |
| Indemnity | A vendor's promise to defend you, and pay, if a covered claim arises |
| Text and data mining | Automated analysis of large amounts of content, including for AI training |
| Likeness | A person's recognisable face, voice or identity |
| Trade mark | A protected sign (a name, logo) that identifies a brand |
7. Try it (no code needed)
- Pick one AI-generated asset you've used (an illustration, an icon set, a hero image).
- Find the tool's current terms: who owns the output? May you use it commercially? Is there an indemnity on your plan?
- Write a four-line record for it (section 5).
- Decide: is this an asset where ownership matters? If yes, what would you change about how you make the next one?
Reflection questions: Which of your deliverables would you never generate with AI, and why? Who in your organisation should see the indemnity conditions?
8. If you lead a team
- Keep the approved tools list with each tool's ownership terms and indemnity status.
- Agree which assets must not be AI-generated (logos, brand mascots, anything with real people) and which need a record.
- Ask legal to review the indemnity conditions before relying on them.
Checklist
Learn more
- US: Copyright Office, Copyright and AI · Part 2: copyrightability
- UK: Getty Images v Stability AI judgment
- EU: General-purpose AI Q&A · GPAI Code of Practice · AI Act, Article 50
- Vendors: Microsoft Customer Copyright Commitment · Google Cloud indemnification · OpenAI service terms · Adobe Firefly legal terms
Books, courses and videos to go deeper
Related
AI for visual design: images and video · AI governance and data protection · Ethics, bias and inclusion in AI design · AI tools landscape for designers