Lookalike and trademark checks for AI-generated designs: a checklist
A pre-release checklist for prints, graphics and campaign images made with generative tools, covering similarity, trade marks, watermarks, tool terms and records.
KEY TAKEAWAYS Summary by the editors
- In Getty Images v Stability AI (High Court of England and Wales, 4 November 2025), the court accepted that earlier Stable Diffusion versions could produce images carrying Getty or iStock watermarks under realistic prompting, and found limited, historic trade mark infringement.
- A European Parliament study from July 2025 notes that where outputs substantially resemble training data, questions of copyright infringement may arise, and that there is little guidance on how to assess this.
- The study cites EUIPO work showing image and language models can memorise and regurgitate protected content, particularly where material is original or unique.
- The US Copyright Office says AI output is protectable only for human-authored contributions, so an unprotectable generated design can still infringe someone else's rights.
- The Getty judgment is likely to be appealed, so UK guidance on trade marks and generative AI may change.
Before release, every AI-generated design should pass five checks: does it contain any logo, watermark or brand element; does it resemble an existing protected design or artwork; is the intended use clear of third-party trade marks; do the tool terms allow commercial use; and is there a record of who reviewed it. Generative tools can reproduce things they were trained on, and the legal risk sits with the company that sells the product, not only with the tool vendor.
This is general information, not legal advice. Trade mark and design law differs by country, and anyone launching a significant line should involve a specialist.
Why can an AI-generated design infringe someone else's rights?
A European Parliament study published in July 2025 explains that where outputs are substantially based on or resemble training data, questions of copyright infringement may arise, and that current legislation and case law give little guidance on how to assess this. The study cites work by the EU Intellectual Property Office showing that language and image models can memorise and regurgitate long sequences or images from training data, including protected content, particularly where the material is original or unique.
Separately, trade marks protect signs, not creative works. A generated graphic that includes a recognisable logo, a distinctive pattern associated with a house, or a watermark can raise trade mark issues whatever the copyright position.
What did the Getty Images v Stability AI judgment say about trade marks?
On 4 November 2025 the High Court of England and Wales gave judgment in Getty Images v Stability AI. According to a law firm summary, the court accepted that some earlier versions of Stable Diffusion could, under realistic, non-contrived prompting, generate images containing the Getty or iStock watermark, and that such outputs may infringe if produced in the course of trade and capable of causing confusion. It found only limited and historic infringement from earlier versions, and no detriment to distinctive character or reputation and no unfair advantage. Filtering and hosted environments reduced the risk.
The same summary reports that Getty abandoned its primary training claim because it could not show that reproduction took place in the UK, that the court held the model weights are not an infringing copy, and that the decision is likely to be appealed. For fashion teams, the lesson is narrow but practical: outputs can carry brand marks, and checking for them is sensible.
What belongs on the checklist?
- Scan for visible marks. Look for logos, watermarks, signatures, text fragments and brand-like monograms, including partial or distorted ones.
- Run reverse image searches. Search the whole design and its main motifs to find close matches in existing images and products.
- Review against known references. Compare with the brief's reference set, competitors' bestsellers and iconic designs in the category.
- Check trade mark registers. Search any words, names, slogans, shapes and patterns used, in the classes and countries where you will sell.
- Check prompts for risk. Confirm prompts did not name living artists, brands or characters, and keep the prompt log.
- Review tool terms. Confirm commercial use is allowed, who bears third-party claims, and any restrictions on outputs.
- Assess distinctiveness. If the design is close to something existing, change it by hand or discard it.
- Record the review. Name the reviewer, the date, the searches run and the decision.
How should the risk level change the depth of review?
| Use | Exposure | Suggested review |
|---|---|---|
| Internal moodboard | Low | Basic visual check; keep logs |
| Print or graphic on a product | High | Full checklist, reverse search and register checks |
| Hero campaign image | High | Full checklist plus disclosure and likeness review |
| Social content | Medium | Visible marks, similarity and labelling check |
| Private-label or licensed line | High | Full checklist plus licensor approval |
Who is responsible when something slips through?
Contracts matter. Check whether the tool provider offers any indemnity for outputs and under what conditions, and do not assume it does. In supplier and licensing agreements, require suppliers to say which tools they use, and agree who bears the cost of a third-party claim. The US Copyright Office's January 2025 report adds a related point: AI output is protectable only for the human-authored contribution, so a generated design may be both hard to protect and capable of infringing, which is the worst combination for a commercial product.
What records should you keep?
- The prompt, tool, version and settings for each released design.
- Screenshots or exports of searches carried out and their dates.
- The reviewer's decision and any changes made as a result.
- Tool terms of use as they stood at the time of generation.
- Supplier declarations about the use of AI tools.
How does this differ for text and names?
Generative tools also produce words: product names, slogans, print text and campaign copy. Trade marks protect names and slogans, so a generated collection name should be searched in the relevant registers before use, in the countries and product classes where it will be sold. Generated text on a print can also contain garbled or partial brand names. Check it by eye and by search, and avoid prompts that ask a tool to write in the style of a named brand.
Characters and licensed properties deserve particular care. A prompt that asks for a cartoon character or a famous figure may produce an output that is close enough to infringe, and the fact that it was machine-made does not change that. Block these prompts in team guidelines, and treat any output that resembles a known character as unusable without a licence.
Finally, build the checklist into the workflow rather than adding it at the end. A review gate in the product development process, with a named owner and a record, is cheaper than a recall or a dispute after launch.
Frequently asked questions
Can an AI-generated design infringe copyright or trade marks?
Yes. A European Parliament study from 2025 notes that outputs resembling training data may raise infringement questions, and the 2025 Getty Images v Stability AI judgment accepted that earlier Stable Diffusion versions could generate watermarked images that may infringe trade marks.
How do I check an AI-generated print for lookalikes?
Look for visible marks, run reverse image searches on the whole design and main motifs, compare against competitors and iconic designs, and search trade mark registers for words and shapes. Record each check and the decision.
Who is liable if my AI-generated design copies something?
Usually the company that sells the product bears the first exposure. Tool terms may or may not offer any protection, so check them and agree responsibilities with suppliers and licensors in writing.
Did Getty win its case against Stability AI?
Only in a limited way. According to a law firm summary, the court found limited, historic trade mark infringement relating to watermarks in earlier versions, rejected the secondary copyright claim about model weights, and the decision is likely to be appealed.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
