Deepfake disclosure laws: when synthetic audio must say so
The labeling rules arriving for AI audio: the EU AI Act's transparency article, US state election-deepfake laws, platform policies, and where Cantari already stands.
Updated June 11, 2026
The idea behind all of it
Every disclosure law reduces to one sentence: people should be able to find out that synthetic media is synthetic. Regulators disagree about scope, penalties, and timing, but not about the principle, and audio that imitates a real person sits at the strict end of every version.
If you publish voice work, this is the area of law most likely to apply to you personally, because the duty to disclose usually lands on the publisher, not on the tool.
The EU AI Act: transparency as law
The EU AI Act (Regulation 2024/1689) writes the principle into Article 50: providers of systems that generate synthetic audio must ensure outputs are marked as artificially generated in a machine-readable way, and people who deploy deepfakes, content that realistically impersonates real persons, must disclose that the content was artificially created. These transparency obligations apply from August 2026.
The reach matters: it applies when the output is used in the EU, not only to European companies. If your audience includes Europe, Article 50 is part of your publishing checklist.
US states: elections first
The United States has no general federal labeling law for synthetic audio yet, so states started where the stakes felt highest. Texas was early: SB 751 (2019) made it a criminal offense to fabricate a deceptive video with intent to influence the outcome of an election. Minnesota Statutes 609.771 criminalizes knowingly disseminating a deepfake of a person without their consent, with intent to injure a candidate or influence an election, in the window around conventions and voting.
Dozens of states now have some variant, differing in time windows, whether a disclaimer cures the violation, and civil versus criminal remedies. The trend line is the only stable fact: more states, broader scope, every year. The political-voice cases that pushed this wave are covered in impersonation and public figures.
Platform rules fill the gaps
Ahead of most statutes, the large video, social, and podcast platforms added their own synthetic-media policies: disclosure toggles for realistic AI content, labels applied on the platform's side, and takedown lanes for unconsented voice imitation. Enforcement varies wildly, but the practical norm is settled, and it is the same norm the laws are converging on.
Our plain advice: when you publish audio that a listener could mistake for a real person's recorded speech, say it is synthetic where the listener can see it. A line in the description costs nothing; the trouble it prevents is not hypothetical.
What this means on Cantari
Cantari's public surfaces already work the way these rules point. Every clip in the community is framed as made with AI and carries a visible credit naming the engine and the voice that spoke it; nothing shared here presents itself as a human recording. Sharing itself is consent-gated, and the credit is not optional.
Your private exports are a different surface, and we are honest about it: you own them, they carry no watermark, and Cantari does not stamp anything into the file. That ownership is a feature, and it comes with the corresponding duty: when a law like Article 50 or a platform policy requires a label where you publish, applying it is your job, not something the file does for you.