The EU AI Act's transparency requirements are not confined to a single provision. Vendors deploying AI in fashion — virtual try-on, synthetic imagery, trend-signal platforms, personalisation engines — face obligations drawn from Article 50, Article 53 and Article 13, each with a different scope, a different addressee and a different documentary form. Mapping each duty to its actual article is the first step toward a compliance file that will survive regulatory scrutiny.
Key takeaways
- Article 50 contains five distinct paragraphs of transparency duties; conflating them produces gaps in your compliance file.
- The obligation to mark synthetic content in a machine-readable format falls on the provider, not only the deployer.
- General-purpose AI model providers face a separate documentation regime under Article 53, including a public training-content summary.
- Article 13 governs instructions for use and applies to high-risk AI systems, not to all AI systems.
- The Code of Practice on Transparency of AI-generated Content offers a structured route to demonstrating Article 50 compliance.
Why does article-level precision matter for your compliance file?
Regulators and auditors reviewing a compliance file will check that each documented measure cites the provision that mandates it. A document labelled 'transparency policy' that does not distinguish between, say, the duty to inform a natural person they are interacting with an AI system (Article 50(1)) and the duty to mark synthetic output in a machine-readable format (Article 50(2)) gives an assessor no basis for concluding that either obligation has been met. Fashion AI vendors — whose products routinely generate synthetic imagery, power interactive styling assistants and process biometric-adjacent data — are exposed across several of these provisions simultaneously.
What follows is a numbered reference list. Each item names the obligation, the article and paragraph it derives from, and the documentary form it requires.
The seven obligations
1. Inform users they are interacting with an AI system (Article 50(1))
Providers of AI systems intended to interact directly with natural persons must ensure those persons are informed they are dealing with an AI system, unless this is obvious from context. For fashion applications — AI styling assistants, chatbots handling returns, virtual personal shoppers — this duty applies at the point of interaction. The documentary requirement is a disclosure mechanism: a persistent label, an opening statement, or an equivalent technical measure, with evidence that it is presented before or at the start of each interaction.
What to document: the disclosure text or UI element; the trigger logic that ensures it appears; testing records confirming it functions across deployment surfaces.
Scope note: the obligation sits with the provider of the AI system. Where a fashion retailer deploys a third-party conversational AI, the retailer is the deployer and carries a parallel duty under Article 50(1) to ensure the disclosure is in place — but the provider remains responsible for making the disclosure technically possible.
2. Label AI-generated synthetic content in a machine-readable format (Article 50(2))
Article 50 requires providers of AI systems that generate synthetic audio, image, video or text content to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solution must be effective, interoperable, robust and reliable, taking into account the specificities and limitations of various content types and the generally acknowledged state of the art.
For fashion vendors, this obligation covers AI-generated product imagery, synthetic model photography, AI-composed lookbook video, and generated copy. A watermark visible only to the human eye does not satisfy the machine-readable requirement.
What to document: the provenance-marking or watermarking standard adopted (e.g., C2PA, SynthID or equivalent); technical specifications of the implementation; interoperability evidence; and a record of how the solution was assessed against the state of the art at the time of deployment.
Scope note: this is a provider-level obligation. It applies regardless of whether the AI system is classified as high-risk.
3. Label deepfakes and AI-generated publications (Article 50(3) and 50(4))
Deployers who use AI to generate or manipulate images, audio or video that constitute a deepfake must disclose that the content has been artificially created or manipulated. A parallel obligation in Article 50(4) covers AI-generated text published in the public interest — news, analysis, public communication — requiring disclosure that the text was AI-generated, unless a human has substantially reviewed and edited it.
Fashion brands that produce AI-generated campaign imagery featuring realistic human likenesses, or that publish AI-drafted editorial content, must address both paragraphs. The disclosure must be 'clearly visible' for deepfakes; for text, it must be 'clearly and prominently' marked.
What to document: a classification procedure that identifies which outputs qualify as deepfakes; the disclosure label or notice used; the editorial review process (if claiming the Article 50(4) exception for substantially human-edited text); and records of how the procedure is applied in practice.
4. Disclose emotion recognition and biometric categorisation (Article 50(5))
Deployers of AI systems that perform emotion recognition or biometric categorisation on natural persons must inform those persons of the system's operation. This paragraph is directly relevant to fashion AI that analyses facial expressions to infer customer sentiment, or that categorises shoppers by body type, age bracket or other biometric-adjacent attributes to personalise recommendations.
What to document: a register of AI systems in deployment that perform emotion recognition or biometric categorisation; the disclosure mechanism used for each; and the legal basis under GDPR for processing any personal data involved (the AI Act obligation and the GDPR obligation are cumulative, not alternative).
Scope note: this obligation falls on the deployer, not the provider.
5. Provide a GPAI Model Documentation Form (Article 53(1)(a))
Providers of general-purpose AI (GPAI) models must draw up and keep up to date technical documentation including the information set out in Annex XI of the AI Act. The General-Purpose AI Code of Practice, published in July 2025, is a voluntary instrument designed to help providers demonstrate compliance with this and related GPAI obligations. Signing up to the Code creates a rebuttable presumption of conformity.
Fashion AI vendors who develop or fine-tune foundation models — rather than deploying third-party models via API — are likely to qualify as GPAI model providers and must produce this documentation.
What to document: the Annex XI technical documentation form, covering model architecture, training methodology, evaluation results, known limitations and intended use cases; version-control records showing the document is kept current.
6. Publish a summary of training content (Article 53(1)(d))
GPAI model providers must make publicly available a sufficiently detailed summary of the content used to train the model, according to a template provided by the AI Office. This obligation is distinct from the technical documentation in Annex XI: it is a public-facing disclosure, not a confidential compliance document.
For fashion AI vendors whose models are trained on proprietary garment imagery, pattern archives or consumer behaviour data, the summary must describe the categories and sources of training data without necessarily disclosing commercially sensitive specifics. The AI Office template governs the required level of detail.
What to document: the published summary itself, with a stable URL; the date of publication; a record of each update and the trigger for it; and the AI Office template version used.
7. Provide instructions for use (Article 13)
Providers of high-risk AI systems must provide deployers with instructions for use that are clear, complete and in a language and format that deployers can reasonably be expected to understand. Article 13 specifies the minimum content: the identity and contact details of the provider; the capabilities and limitations of the system; performance metrics; human oversight measures; and any known risks.
Fashion AI systems classified as high-risk — for example, systems used in employment-related decisions such as candidate screening for design or buying roles — must be accompanied by compliant instructions. Virtual try-on and trend-forecasting tools are generally not classified as high-risk under the current Annex III, but vendors should confirm their classification before assuming Article 13 does not apply.
What to document: the instructions for use document, version-controlled and dated; evidence that it was provided to each deployer before or at the time of deployment; and a record of updates triggered by material changes to the system.
How the Code of Practice on Transparency of AI-generated Content fits in
The Code of Practice on Transparency of AI-generated Content is a structured route to demonstrating compliance with the Article 50 obligations. Adherence to the Code creates a rebuttable presumption of conformity with those obligations, which is a material advantage when regulators or auditors review your compliance file. The Code addresses marking and detection of AI-generated content and labelling of deepfakes — obligations 2 and 3 in this list — and is designed to complement, not replace, the technical documentation requirements under Article 53.
For fashion AI vendors, the practical value of the Code is that it provides a concrete, externally validated set of measures rather than requiring each vendor to independently interpret 'effective, interoperable, robust and reliable' technical solutions. The European Commission guidelines on transparency obligations, published in August 2026, offer further interpretive guidance on each Article 50 paragraph.
Application dates and the grace period for previously released systems
Article 50 applies from 2 August 2026, as confirmed in the Commission guidelines. Systems that were already released before that date benefit from a grace period running to 2 December 2026 for the purpose of marking previously released outputs — vendors should not treat this as a general compliance extension, but as a limited window for applying machine-readable marks to content generated before the application date.
The Morgan Lewis analysis of changes to EU AI Act deadlines notes that additional preparation time granted by deadline adjustments is not a reprieve: compliance infrastructure should be operational at the application date, not at the end of any grace period.
Practical implications for fashion AI vendors
Vendors such as Aiuta — whose virtual try-on platform processes uploaded images of real users alongside AI-generated virtual models — face obligations under Articles 50(1), 50(2) and potentially 50(5), depending on whether their systems perform any biometric categorisation in the course of fitting. Trend-signal platforms, including the social-image forecasting capabilities now integrated into Luxurynsight following its acquisition of Heuritech, generate synthetic or AI-processed content that may trigger Article 50(2) marking requirements. Enterprise retail AI orchestration platforms such as Vue.ai — which cover product tagging, automated on-model imagery and personalised commerce journeys — should assess each module separately, since different components may engage different paragraphs of Article 50.
The Taylor Wessing sector analysis of fashion and the AI Act notes that most fashion AI applications fall outside the high-risk categories in Annex III, which means Article 13 instructions-for-use obligations will not apply to the majority of fashion AI products — but the Article 50 obligations apply regardless of risk classification.
Assembling your compliance file: a checklist
- Article 50(1): disclosure mechanism documented and tested across all interactive AI deployments.
- Article 50(2): machine-readable provenance marking implemented; technical standard documented; interoperability evidence on file.
- Article 50(3)/(4): deepfake disclosure procedure and editorial review records in place.
- Article 50(5): register of emotion recognition and biometric categorisation systems; disclosure mechanism for each; GDPR legal basis documented.
- Article 53(1)(a): Annex XI technical documentation completed and version-controlled (GPAI providers only).
- Article 53(1)(d): public training-content summary published at a stable URL, using the AI Office template (GPAI providers only).
- Article 13: instructions for use provided to each deployer before deployment (high-risk systems only).
FAQ
Does Article 50 apply to all AI systems or only high-risk ones? Article 50 applies to AI systems that interact with natural persons, generate synthetic content, perform emotion recognition, or categorise by biometric attributes — regardless of risk classification. It is not limited to high-risk systems.
Who is responsible for machine-readable marking of synthetic imagery — the AI provider or the fashion brand deploying it? The obligation under Article 50(2) sits with the provider of the AI system that generates the synthetic content. The deploying brand should verify that the provider has implemented compliant marking before integrating the system.
Does the Code of Practice on Transparency of AI-generated Content create a legal obligation? No. It is voluntary. However, adherence creates a rebuttable presumption of conformity with the Article 50 obligations it covers, which is a significant evidentiary advantage in any regulatory review.
What is the grace period for previously released AI-generated content? Vendors have until 2 December 2026 to apply machine-readable marks to content generated by systems released before 2 August 2026. This is a narrow, content-specific window, not a general compliance extension.
Does Article 13 apply to virtual try-on or trend-forecasting tools? Generally no, because these tools are not listed in Annex III as high-risk AI systems. Vendors should confirm their classification independently, as the Annex III list can be updated by delegated act.
Further reading
- Article 50: Transparency obligations for providers and deployers of certain AI systems — AI Act Service Desk, European Commission
- Guidelines on transparency obligations for providers and deployers of certain AI systems — European Commission
- Code of Practice on Transparency of AI-generated Content — European Commission
- The General-Purpose AI Code of Practice — European Commission
- Fashion meets the AI Act — Taylor Wessing