Fashion AI Lab EU

Research and EU regulation for fashion AI, digested

Back to blog

Training Data Provenance for Fashion Generative AI: Legal and Technical Risks

· Last updated:
Training Data Provenance for Fashion Generative AI: Legal and Technical Risks

When a fashion brand or AI vendor trains an image-generation model on garment photographs, runway imagery, or a proprietary design archive, it does not simply acquire a capability — it inherits a stack of legal obligations that travel with every output the model ever produces. Copyright claims, sui generis database rights, and the EU AI Act's transparency regime for general-purpose AI models each impose distinct duties, and none of them disappear because the training happened quietly or in a private cloud environment.

Key takeaways

  • Copyright in individual garment photographs and design drawings vests in their authors; scraping or licensing those images for model training does not extinguish the underlying rights.
  • EU database rights can protect a curated image collection independently of whether any single image is itself copyrightable.
  • The EU AI Act imposes dataset documentation and transparency obligations on providers of general-purpose AI models, including those trained on fashion imagery.
  • Technical controls — dataset cards, lineage logs, cryptographic hashing — are enforceable and auditable; policy statements alone are not.
  • The risk profile differs substantially between a brand training on its own archive and a vendor training on third-party or web-scraped imagery.

Fashion imagery sits at the intersection of several overlapping rights regimes, and the interaction between them is not always intuitive.

A product photograph of a garment — even a catalogue shot on a white background — is protected by copyright in every EU member state, provided it reflects the author's own intellectual creation. The threshold is low but not zero: fully automated, purely technical shots may fall outside it, but most commercial fashion photography clears it comfortably. Design drawings, technical flats, and pattern illustrations attract copyright as artistic works in their own right.

Using those images as training data without a licence, or without falling within a recognised exception, is an act restricted by the rightsholder. The EU's Directive on Copyright in the Digital Single Market (DSM Directive) introduced a text and data mining (TDM) exception, but its scope for commercial AI training is contested. Article 4 of the DSM Directive allows TDM for any purpose unless rightsholders have expressly reserved their rights in a machine-readable way — a reservation mechanism that major stock libraries and many fashion brands have begun to implement.

For AI vendors training on web-scraped imagery, the practical question is whether any meaningful proportion of the training set carries such a reservation, and whether the vendor can demonstrate it checked. For brands training on third-party licensed imagery, the question is whether their licence agreement contemplated AI training as a permitted use — most agreements predating 2022 did not.

Sui generis database rights

Even where individual images are not protected by copyright — because they are too old, too mechanical, or their authors cannot be identified — the collection itself may be. EU database rights protect a database in which there has been substantial investment in obtaining, verifying, or presenting the contents. A curated archive of runway photographs assembled over decades, or a structured library of technical design drawings, is likely to qualify.

Database rights are infringed by extraction or re-utilisation of a substantial part of the database's contents. Training a model on such a collection extracts its contents systematically; whether that constitutes extraction of a 'substantial part' in the legal sense is a question that EU courts have not yet resolved in the generative AI context, but the risk is real and should not be assumed away.

Moral rights and design rights

In several EU jurisdictions, moral rights — including the right of integrity — are inalienable and cannot be waived by contract. A model trained to generate garments in the style of a named designer may produce outputs that a court treats as a distortion of that designer's work. Separately, registered and unregistered Community design rights protect the appearance of a product for defined periods; a model that reliably reproduces the silhouette or ornamentation of a protected design creates direct infringement risk in its outputs.

What does the EU AI Act require for training data provenance?

The EU AI Act introduces a distinct compliance layer that sits on top of intellectual property law. Its obligations for general-purpose AI models — the category that covers most large image-generation systems — are among the most documentation-intensive in the regulation.

General-purpose AI model obligations

Providers of general-purpose AI models must prepare and maintain technical documentation that covers, among other things, the data used for training, including a general description of the data, information about the provenance of that data, and the data governance measures applied. The regulation does not prescribe a single format, but the documentation must be sufficient to allow the AI Office and national competent authorities to assess compliance.

For fashion-specific deployments — a brand fine-tuning a foundation model on its own design archive, or an AI vendor building a garment-generation product on top of a general-purpose model — the question is which entity in the chain bears the documentation obligation. The Act's structure places primary obligations on the provider of the general-purpose model, but deployers who fine-tune or adapt that model take on additional responsibilities proportional to the degree of modification.

Dataset cards and lineage records

The documentation the Act expects in practice includes what the research community calls a 'dataset card': a structured record of what the dataset contains, where its constituent items came from, what rights clearances were obtained, what filtering or deduplication was applied, and what known biases or gaps exist. For a fashion training set, a dataset card would need to address the provenance of each image batch — whether it came from a brand's own archive, a licensed stock library, a web crawl, or a synthetic generation pipeline.

Lineage records go further: they trace each item or batch through every transformation step, from ingestion through cleaning, labelling, and final inclusion in the training set. The practical challenge is that many existing fashion AI training sets were assembled before these concepts were operationalised, and reconstructing lineage retrospectively is technically difficult and legally uncertain.

Retention obligations

The Act requires that technical documentation be retained for a defined period after the model is placed on the market or put into service. For legal, IP, and AI product teams, this means that the documentation infrastructure must be built before training begins, not assembled after the fact. A dataset card created after a model ships is not evidence of provenance; it is a reconstruction, and regulators are unlikely to treat it as equivalent.

Which controls are technically enforceable and which are merely stated?

This distinction matters enormously in practice. A policy document asserting that training data was lawfully obtained is a statement; a cryptographic hash of each training image, logged at ingestion time and stored in an append-only audit trail, is evidence. The two are not interchangeable under regulatory scrutiny.

Controls that are technically enforceable

Cryptographic hashing and deduplication logs. Hashing each image at ingestion creates a tamper-evident record of what entered the training pipeline. If a rightsholder later identifies an image as theirs, the hash log either confirms or excludes its presence. This is a standard practice in responsible ML operations and is directly relevant to demonstrating compliance.

Licence metadata preservation. Where images are sourced from a licensed library, the licence terms and the associated image identifiers should be stored alongside the image in the training pipeline. Stripping metadata — a common step in image preprocessing — destroys the provenance chain and is difficult to reconstruct.

Opt-out and removal mechanisms. Several jurisdictions and platforms have introduced mechanisms by which rightsholders can signal that their content should not be used for AI training. A technically enforceable control checks incoming data against opt-out registries at ingestion and excludes flagged content. A policy that promises to honour opt-outs without a technical implementation of that check is not enforceable in any meaningful sense.

Differential privacy and membership inference defences. These techniques reduce the risk that a trained model memorises and can reproduce specific training images — a risk that is particularly acute for distinctive garment designs. They are not a substitute for lawful data acquisition, but they reduce the magnitude of harm if a rights question arises later.

Controls that are merely stated

Contrast the above with controls that exist only as assertions: a terms-of-service clause stating that training data is 'properly licensed', a privacy notice claiming that personal data in images has been anonymised without specifying the technique, or a vendor's representation that its dataset is 'ethically sourced' without an auditable definition of what that means. These statements may be accurate, but they cannot be verified by a regulator, a court, or a brand's own legal team without underlying technical evidence.

The gap between stated and enforceable controls is where most current fashion AI compliance programmes are weakest. Brands and vendors that have invested in image-generation capabilities often have sophisticated ML infrastructure but underdeveloped data governance tooling.

How does the risk profile differ by use case?

Not all training scenarios carry the same exposure, and understanding the gradient is useful for prioritising remediation.

Brand training on its own archive. A fashion house training a model exclusively on photographs it commissioned and owns, design drawings created by its own employees within the scope of their employment, and technical assets it has generated internally faces the lowest third-party IP risk. The primary obligations are internal governance — ensuring the archive is accurately documented, that any third-party contributions are identified and excluded or separately cleared, and that the resulting model's outputs do not infringe registered designs held by others.

Brand fine-tuning a foundation model. When a brand adapts a general-purpose model — such as OpenAI's DALL-E or Midjourney's generation platform — on its own design data, it inherits whatever provenance questions attach to the foundation model's pre-training, while adding its own fine-tuning data to the picture. The EU AI Act's chain-of-responsibility provisions mean the brand, as a deployer making substantial modifications, takes on documentation obligations in addition to those of the foundation model provider.

Vendor training on third-party or web-scraped imagery. This is the highest-risk scenario. A vendor assembling a training set from web crawls, stock libraries, or aggregated brand imagery faces the full range of copyright, database rights, and TDM exception questions, compounded by the practical difficulty of demonstrating at scale that rights were cleared or that applicable exceptions apply. Adobe Firefly has publicly positioned its training data strategy around licensed and rights-cleared content as a differentiating factor — an approach that reflects the legal pressure on vendors in this category, regardless of whether it fully resolves every rights question.

The following steps are ordered by the degree to which they are technically enforceable rather than merely administrative.

  1. Audit your existing training sets before they are used again. Map every image batch to its source, the licence or exception under which it was acquired, and whether that licence contemplated AI training. Flag gaps for legal review before the next training run.
  2. Implement hash-based ingestion logging from the next training run forward. This is a one-time infrastructure investment that creates the audit trail regulators will expect.
  3. Preserve licence metadata through the preprocessing pipeline. Review your data pipeline for steps that strip EXIF data, watermarks, or embedded licence identifiers, and add a parallel metadata store that survives those steps.
  4. Draft dataset cards for each training set, including fine-tuning sets. Use a structured format that covers provenance, rights clearances, filtering steps, known gaps, and retention schedule. Treat these as living documents updated at each training run.
  5. Check incoming data against available opt-out registries at ingestion. Automate this check so that it runs without manual intervention.
  6. Review your foundation model provider's documentation. If you are fine-tuning a third-party model, request the provider's dataset documentation. Absence of documentation is itself a compliance signal.
  7. Align retention schedules with the EU AI Act's requirements. Documentation must outlast the model's active deployment; build retention into your data governance policy rather than relying on default storage lifecycles.

What remains unresolved

Several questions are genuinely open at the time of writing. EU courts have not yet ruled on whether training a model on a copyrighted image constitutes a reproduction in the legal sense, or whether the TDM exception's opt-out mechanism is technically adequate when implemented via robots.txt or similar signals. The AI Act's implementing acts for general-purpose AI model documentation are still being developed, and the precise format of required dataset documentation has not been finalised. The interaction between the Act's transparency obligations and the GDPR — particularly where training images contain identifiable individuals, as fashion photography routinely does — remains a live compliance question that neither regulation fully resolves on its own.

For teams building or deploying fashion generative AI today, the practical implication is that the compliance infrastructure you build now needs to be flexible enough to accommodate requirements that will become more specific over the next two to three years. Investing in technically enforceable controls — rather than policy statements — gives you that flexibility, because evidence of what you actually did is more durable than a record of what you said you would do.

FAQ

Does the EU AI Act apply to a fashion brand that fine-tunes a third-party image model on its own design archive? Yes, to a degree proportional to the modification. A brand making substantial modifications to a general-purpose model takes on documentation and transparency obligations under the Act in addition to those of the foundation model provider. Light fine-tuning with no change to the model's general-purpose character sits in a greyer zone that implementing guidance is expected to clarify.

Can a fashion brand rely on the DSM Directive's text and data mining exception for commercial AI training? Only partially. Article 4 of the DSM Directive permits TDM for any purpose, but rightsholders may opt out via machine-readable reservation. Many stock libraries and brands have begun implementing such reservations, so a web-scraped training set assembled today is likely to contain a non-trivial proportion of opted-out content. The exception does not provide a blanket clearance for commercial training.

What is a dataset card and is it legally required? A dataset card is a structured document recording a training set's contents, provenance, rights clearances, filtering steps, and known limitations. The EU AI Act requires general-purpose AI model providers to maintain technical documentation covering training data provenance; a dataset card is the practical implementation of that requirement. Its precise format is not yet mandated, but the underlying obligation is.

Does training on garment photographs create GDPR obligations? If the photographs contain identifiable individuals — models, designers, people photographed in public — then yes. Processing those images for AI training is processing of personal data under the GDPR, and a lawful basis is required. Anonymisation techniques applied during preprocessing may reduce ongoing obligations, but the initial collection and use for training must itself be lawful.

How does a brand demonstrate that its training data controls are enforceable rather than merely stated? Through technical artefacts: ingestion hash logs, preserved licence metadata, automated opt-out checks, and version-controlled dataset cards stored in an append-only system. A regulator or auditor can verify these; a policy document asserting compliance cannot be verified in the same way. Building the technical infrastructure before training begins is substantially easier than reconstructing it after the fact.

Further reading

Share this article:

Training Data Provenance: Fashion Generative AI Legal Risks