Can brands safely use generative AI on their own design archives?

Brands can use generative AI on their own design archives with manageable risk, but only if they treat two questions separately: whether the base model they build on was trained lawfully, and whether they hold the rights in the archive material they add. In 2026 the law is clearer on provider obligations than on the legality of training itself, so contracts and internal rights audits carry most of the protection.

Why does training data matter to a fashion brand?

Most brand uses of generative AI start from a foundation model trained by someone else on very large image collections, then add brand material either by prompting with reference images or by fine-tuning. If the foundation model reproduces elements of protected third-party work, the brand that publishes or manufactures the output may carry the commercial and legal consequences, whatever the provider's position. That is why the provenance of the base model is a procurement question, not only a legal one.

What does EU law require from AI model providers?

The EU AI Act places specific obligations on providers of general-purpose AI models, which apply from 2 August 2025 according to the Commission's implementation timeline, with enforcement powers for these obligations starting on 2 August 2026. Under Article 53(1)(c), a provider must put in place a policy to comply with EU copyright law, including identifying and respecting rights reservations (opt-outs) made under the text and data mining exception in Article 4(3) of Directive (EU) 2019/790. Under Article 53(1)(d), it must publish a sufficiently detailed summary of the content used for training, following a template from the AI Office. The open-source exemption does not remove these two duties.

For brands, the practical consequence is that a provider of a general-purpose model should be able to point to its copyright policy and its public training summary when asked during procurement. A provider that cannot do so is a signal to look more closely, particularly if the brand plans to publish outputs in the EU.

Read also

Who owns an AI-generated fashion design? Copyright and IP explained

What did Getty Images v Stability AI decide?

The UK High Court handed down judgment on 4 November 2025 in the first major UK case on generative AI and copyright. Getty abandoned its claims about training and about outputs before the end of trial; the training claim faced the argument that Stable Diffusion was trained outside the UK, so UK copyright was not engaged. The court rejected Getty's secondary infringement claim, holding that the model weights were not an infringing copy because they did not store copies of Getty's works. Trade mark infringement was found only to a very limited extent for early model versions.

Commentary from the University of London's law blog notes that the decision is being appealed and that, because copyright is territorial, the case leaves open whether unlicensed training is infringement. For brands, the lesson is that UK case law does not yet provide a safe harbour for either side.

What is the position in the United States?

The US Copyright Office released a pre-publication version of Part 3 of its report on generative AI training in May 2025. It states that the steps involved in building a training dataset clearly implicate the reproduction right, and it analyses whether such use can be fair use rather than giving a blanket answer. Courts decide individual cases, so the legal status of training on unlicensed material remains case-specific.

For a brand operating internationally, the practical effect of these differences is that the safest assumption is the strictest one. If a model provider cannot explain how it handled rights reservations and what kinds of content it trained on, a brand cannot easily judge the risk that outputs reproduce protected work, wherever the brand sells.

Does a brand own all the rights in its own archive?

Often not entirely. An archive accumulated over decades can include artwork by freelance designers, prints licensed from studios, campaign photography owned by photographers and images of identifiable models. Each may carry rights or contractual limits that predate AI. Before archive material is used for fine-tuning or as reference input, a rights audit should check:

• Who created each print, sketch or graphic, and whether rights were assigned to the brand or only licensed.
• Whether licences for third-party prints permit derivative works or use in software training.
• Whether photography contracts allow reuse beyond the original campaign.
• Whether images show identifiable people whose consent covers new uses.
• Whether collaborations with other brands or artists restrict reuse of jointly created designs.

Read also

AI dupes and design copying: how brands protect prints and silhouettes in 2026

What should brands put into AI provider contracts?

Provider terms decide much of the practical risk. Several providers now market enterprise options in which a model is customised on a brand's own assets and promoted as safe for commercial use. Such marketing claims should always be read alongside the actual contract. Key points to settle include whether uploaded material may be used to train models available to others, who owns outputs and any fine-tuned model, where data is stored, what happens on termination, and whether the provider offers intellectual property indemnities and on what conditions.

None of this makes archive-based AI unusable. It means the archive's value as training material depends on the same rights discipline that brands already apply to licensing and collaborations.