Since 2 August 2026, Article 50 of the EU AI Act has applied. For businesses using AI-assisted music, synthetic voices, audio logos or pre-recorded announcements, the practical issue is not whether every piece of audio needs a public label. It is whether the organisation understands its role in the supply chain and can distinguish technical marking from a disclosure aimed at listeners.
This matters to retail groups, hospitality operators, fitness chains and restaurant businesses using audio across multiple locations. A supplier may provide a generated track, a synthetic voice-over may be commissioned for a campaign, or a local team may use an AI tool to adapt messages. Those scenarios may involve different responsibilities.
The legal starting point: Article 50 of the EU AI Act
Article 50 creates transparency obligations for certain providers and deployers of AI systems. It is a binding provision of Regulation (EU) 2024/1689. The Commission’s guidance, published on 20 July 2026, helps interpret the scope and practical operation of those obligations, but it is not a separate law and does not create a blanket new duty to label all AI-made media.
For providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text, Article 50 requires outputs to be marked in a machine-readable format and detectable as artificially generated or manipulated. The relevant technical solution must be effective, interoperable, robust and reliable as far as technically feasible, taking account of the content type, implementation costs, limitations and the generally acknowledged state of the art.
That obligation is directed at the provider of the AI system. A hotel group or retailer using an audio asset is not automatically the provider merely because it plays the asset in its premises. However, procurement and legal teams should establish who provides the relevant system and what evidence supports the technical handling of generated outputs.
Machine-readable marking is not the same as a customer notice
The marking obligation concerns a technical signal that enables generated or manipulated content to be detected. It is not automatically an audible statement in a shop, a sign at reception or a label on every in-store message.
Public-facing disclosure is relevant to deployers in more narrowly defined cases. Most importantly for audio, a deployer using an AI system to generate or manipulate content that constitutes a deepfake must disclose that the content has been artificially generated or manipulated. The information must be clear and distinguishable by the time of first exposure, and meet applicable accessibility requirements.
Not every synthetic voice, music bed or jingle is a deepfake. The AI Act definition concerns image, audio or video that resembles existing persons, objects, places, entities or events and falsely appears authentic or truthful. A generic synthetic voice used for a store opening-hours message therefore should not be treated automatically as a deepfake. A highly realistic imitation presented as the voice of an identifiable real person calls for a much closer assessment.
Examples for in-store and customer-facing audio
Background music and AI-assisted production
A business playing AI-assisted background music should not assume that Article 50 has settled its music licensing position. The AI Act’s transparency rules are separate from copyright, contractual rights, public-performance permissions and the terms governing an audio catalogue.
The sensible operational approach is to document both tracks: rights and usage permissions on one side; AI-system provenance, provider information and any relevant technical documentation on the other. This complements the broader work involved in choosing licensed music for shops, rather than replacing it.
Synthetic voice-over for promotions or service messages
Consider a restaurant chain that runs a synthetic voice message announcing table service, loyalty benefits or seasonal offers. The first questions should be factual: Was a real person’s voice cloned or imitated? Does the output resemble an existing individual? Could a reasonable listener be led to regard it as authentic or truthful? Where the answer may be yes, the deepfake rules need a case-specific legal review.
Where a voice is created as an unmistakably synthetic character and does not purport to be a real person, the deepfake analysis may be different. Context remains central: audience expectations, the manner of presentation and the content’s substantive message all matter.
Audio identity and creative work
An audio logo or fictional campaign character may form part of a wider creative programme. Article 50 provides that where deepfake content forms part of an evidently artistic, creative, satirical, fictional or analogous work, disclosure is limited to an appropriate manner that does not hamper the display or enjoyment of the work. This is not a general exemption for branded content; it only becomes relevant if the content first meets the deepfake test.
For operators building a coherent experience across sites, a sound design approach to in-store radio can make governance easier: content has an owner, an approval route and a defined purpose instead of being uploaded informally by different teams.
What the Commission guidelines clarify
The Commission’s guidelines explain the distinction between providers and deployers, set out practical examples and address exclusions. They state that the provider marking obligation does not apply where an AI system performs an assistive function for standard editing or does not substantially alter the input or its semantics. The guidance also explains that provider-side machine-readable marking alone cannot fulfil a deployer’s duty to disclose a deepfake to people exposed to it.
That distinction is particularly useful for multi-site businesses. A technical marker embedded by a platform may be relevant evidence, but it does not remove the need to assess whether a particular campaign audio requires listener-facing disclosure.
A practical governance checklist
- Create an inventory of AI-generated or AI-manipulated music, spoken announcements, digital-signage audio and campaign assets.
- Identify the organisation’s actual role in each use case: deployer, provider, customer of a provider or commissioner of a system developed under its own name.
- Ask suppliers for clear information on generated outputs, technical marking, source files, contractual rights and permitted uses.
- Assess realistic voice imitation separately from ordinary synthetic audio. Do not assume that all generated sound is a deepfake.
- Set an approval and retention process for voice-over, jingle and promotional content, especially where a real person could be represented.
- Keep AI Act assessment separate from music licensing and other rights checks.
EU scope and the UK
This article concerns the EU AI Act. It should not be presented as a UK labelling rule simply because a business is established in the United Kingdom. At the same time, the AI Act can apply to providers outside the EU where the output of their AI system is used in the Union. UK-based groups operating EU venues or supplying audio into EU markets should therefore assess the relevant facts with specialist advice.
Responsible governance is also consistent with the need for human oversight, traceability and output review discussed in our analysis of AI-assisted music governance. Those measures are useful controls, but they are not a substitute for a legal classification under Article 50.
Conclusion
The most useful response to Article 50 is disciplined classification, not indiscriminate labelling. Businesses should know which party provides the AI system, retain evidence about supplied audio and give particular attention to realistic imitations that might qualify as deepfakes. Before rolling out a campaign or changing supplier terms, obtain specialist legal review for the actual deployment context.
This article is general information, not legal advice.
FAQ
Does the EU AI Act require every AI-generated song to be labelled publicly?
No. Article 50 requires providers of relevant generative AI systems to mark synthetic outputs in a machine-readable format. Public disclosure by deployers is required in specific situations, including deepfakes.
Is a synthetic voice announcement always a deepfake?
No. A deepfake must resemble an existing or plausibly existing person, object, place, entity or event and falsely appear authentic or truthful. The intended audience and deployment context are relevant.
Who is responsible for machine-readable marking?
The obligation in Article 50(2) is aimed at providers of AI systems generating synthetic audio, image, video or text. Customers using generated assets should nevertheless seek appropriate supplier documentation.
Does Article 50 deal with music licensing?
No. AI Act transparency obligations do not determine copyright ownership, music licensing, public-performance permissions or contractual rights to use an audio asset.
