EU AI Act Article 50 Deepfake Labeling for Synthetic Voice in Bank Marketing

What Article 50 asks of providers and deployers

Two separate layers apply to synthetic audio, and treating them as one is the most common source of confusion.

The first layer is machine-readable marking. Article 50(2) of Regulation (EU) 2024/1689 requires providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, taking into account content-type limitations, implementation costs and the generally acknowledged state of the art. The statutory text does not name a specific marking method.

The second layer is disclosure. The Act defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Recital 134 says deployers who use an AI system to generate or manipulate audio that appreciably resembles existing persons and would falsely appear authentic should clearly and distinguishably disclose the artificial origin by labelling the output. Marking is a technical property of the output. Disclosure is something a listener can perceive.

A bank that licenses a voice platform is ordinarily a deployer, so the marking duty falls on the vendor and the disclosure duty falls on the bank. A bank that builds or substantially controls its own generation system should assess whether it also carries provider obligations.

Mapping bank marketing scenarios to the deep fake test

The deep fake test has two elements: resemblance to an existing person or thing, and a false appearance of authenticity. Recital 134 adds the word appreciably to resemblance, and recitals are interpretive aids rather than operative text. The scenario readings below are analytical, not settled legal conclusions.

• Cloned spokesperson voice: Resembles an existing person, so it is the clearest candidate for deep fake treatment. Expect marking by the provider and clear disclosure by the bank.
• Personalized outbound audio in a cloned voice: Same analysis as the cloned spokesperson. Scale does not change the test, but it raises the need to show that disclosure was applied to each item.
• Generic synthetic voice in a campaign: Marking applies. Whether it is a deep fake depends on whether it resembles an existing person and would falsely appear authentic, so case-by-case analysis is needed.
• Voice assistant output: Marking applies to generated audio. Disclosure turns on the same test, so a voice modelled on a real person is treated differently from a generic voice.

Settled readings, open questions and guidance in development

Several points can be stated with confidence. Article 50(2) marking applies to providers of systems that generate synthetic audio. The deep fake definition and the recital 134 disclosure expectation for deployers are in the text. The Commission describes Article 50 as applicable from 2 August 2026.

Other points remain open. The Commission says the AI Office will issue guidelines on Article 50 obligations for providers and deployers, and a voluntary Code of Practice on Transparency of AI-generated Content is intended to help providers and deployers comply with Article 50(2), (4) and (5). The Commission states that signatories can rely on the code's measures to demonstrate compliance after a positive assessment. Adherence is voluntary, but the Article 50 requirements are legal obligations. Marking methods for audio and the exact disclosure format are the kind of detail these instruments are expected to supply, so classifications should be revisited when they change.

The timeline is also moving. A provisional Council-Parliament agreement of 7 May 2026 on the Digital Omnibus on AI would shorten the grace period for providers to implement transparency solutions for artificially generated content from six months to three, with a new deadline of 2 December 2026. As of November 2026, treat that date as provisional until formal adoption, and do not plan around it.

Exceptions, such as those for evidently artistic or creative content or for assistive editing, should be handled as documented legal determinations rather than defaults. A marketing team's view that a campaign is creative is not a basis for skipping disclosure without legal sign-off recorded against the use case.

Enforcing labeling policy in the execution path

Reviewing synthetic audio after publication leaves a gap between a stated policy and a demonstrable outcome. Trussed is runtime control and evidence infrastructure for regulated enterprises. It enforces policy in the execution path on individual interactions and tool calls, with allow, redact and block decisions, so labeling rules are enforced at runtime, not reviewed after the fact. Each enforcement decision produces audit-ready evidence, mapped to the EU AI Act among other frameworks.

The bank still owns the Article 50 determinations: which use cases are deep fakes, which exceptions it relies on, and what disclosure looks like. A runtime layer applies those determinations consistently and records the outcome. Trussed does not claim financial-services deployments. Its production experience is deepest in insurance, another regulated financial sector, and it supports self-managed deployment in the customer cloud with configurable retention and data residency. The value for a bank is the evidence examiners ask for: what the system was allowed to do, and what it did.