Article 50 and voice: what it actually says
Applicable since 2 August 2026. Read from the text of the Regulation. This page is not legal advice.
Most of what circulates about this article comes from press summaries and law firm briefings, not from the text. The summaries are not wrong, but they compress two separate obligations into one sentence and drop the exceptions, which is where the practical questions live.
Two obligations, two different parties
If you build the system. Providers of AI systems that generate synthetic audio must ensure the output is marked in a machine-readable format and detectable as artificially generated or manipulated. The Regulation asks for solutions that are effective, interoperable, robust and reliable, and it qualifies that with "as far as this is technically feasible", taking into account the type of content, the cost of implementation, and recognised industry standards.
If you use the output. Deployers of a system that generates or manipulates audio constituting a deepfake must disclose that the content has been artificially generated or manipulated. This one binds the person who publishes or circulates the audio, not the person who built the model.
The information has to be given in a clear and distinguishable manner, at the latest at the time of the first exposure, and it has to meet accessibility requirements.
The exceptions, which matter more than the rule in practice
The provider obligation does not apply to systems performing assistive editing functions, to systems that do not substantially alter the input, or where use is authorised by law to detect or investigate criminal offences.
The deployer obligation carries the same law enforcement exception, and a second one for artistic, creative, satirical or fictional work. There, disclosure is limited to noting that such content exists, in a way that does not spoil the work.
Read together, these carve outs cover a large share of real audio production. Anyone reading the headline version of this article and concluding that all synthetic speech in Europe is now labelled has read something the text does not say.
What this changes if you are holding a recording
Very little, and that is the part worth understanding.
The obligation runs on whoever generated or circulated the audio. It does not attach to the file in your hands. A recording produced by a system outside the Union, or by an operator who simply ignored the rule, carries no mark and breaks no law that helps you. And a mark that was applied at generation does not necessarily survive the journey: audio that has been re-encoded, compressed for a messaging app, or replayed down a phone line is not the audio that left the model.
So the practical position after 2 August is close to the position before it. If a recording matters to a dispute, its authenticity is still a question of fact to be examined, not a question answered by a label. The law creates a duty. It does not create evidence.
Where the duty and the evidence diverge
The gap that follows is straightforward. An organisation can be fully compliant, mark everything it generates, disclose everything it circulates, and still be unable to say anything about an inbound recording it did not produce. Compliance describes what you emit. It says nothing about what arrives.
That is the situation most disputes involve: a file of unknown origin, no mark, and a claim attached to it.
What is not settled
Two things, and anyone telling you otherwise is ahead of the text. The penalty regime and its application to deployers established outside the Union are not resolved on the face of Article 50 alone. And the standards that would make marking genuinely interoperable are still being written, which is why the Regulation qualifies the requirement with technical feasibility rather than stating a format.
If either of those is load bearing for a decision you are making, the answer belongs with counsel, on the current text, not with a summary of it.
Discuss a recording
If the question is whether a specific recording is authentic, see the engagement page. If a recording cannot support a confident conclusion, you will be told so.