India's rules want AI audio to announce itself out loud. For songs, a label and a tag should be enough

On 14 October YouTube Music started labelling songs by how much generative AI went into them. A day later MediaNama asked the question an Indian product lawyer would ask first: does a badge on a screen meet rules that ask for synthetic audio to announce itself out loud? We think the honest answer is "probably not, as written", and that the rule, not the badge, is what should change for music.

The news hook

YouTube's labels sit on the player page and expand when tapped. Partners declare each track "Fully Gen AI", "Partly Gen AI" or "No Gen AI", and YouTube adds its own detection; pitch correction such as Auto-Tune counts as "No Gen AI", and a label does not affect whether a track can earn money (TechCrunch).

India's Information Technology Amendment Rules, 2026, in force since 20 February, take a different shape. As MediaNama summarises them, synthetic audio must carry "a prominently prefixed audio disclosure", large platforms must ask users whether content is synthetic and verify the answer, and platforms must embed permanent metadata or identifiers "where feasible". MediaNama also notes the rules cover content that appears "indistinguishable from a real person or real-world event", and that it is unclear whether a fully AI-made song with synthetic vocals passes that test. MeitY's 8 October advisory again asked platforms to label manipulated and synthetic content (same source).

The best case for the spoken warning

Put it as a regulator would. A large share of Indian listening happens on phones in pockets, on Bluetooth speakers, through autoplay and in shared rooms. Nobody in that room sees a player page. A visual badge protects only the person holding the screen, and only if they tap it. A spoken prefix reaches everyone who hears the audio, whatever app or device it plays on, and it travels with the file when the song is ripped and reposted to WhatsApp, where no platform label follows.

The regulator also has reason to distrust self-declaration. MediaNama's own June test, cited in its October piece, found YouTube labelled AI videos mainly when creators disclosed AI use or files carried AI metadata, and missed one downloaded from a third-party channel (MediaNama). And India's music trade has not volunteered much: no Indian distributor, label or trade body was among the 27 companies and trade bodies backing IFPI-led commitments to vet AI tracks, MediaNama reported in September (MediaNama). If the industry will not police itself, a hard, audible rule is the one that cannot be quietly skipped.

That is a strong argument, and for cloned speech, fake political audio and voice-note scams we think it wins.

Where it breaks for music

It is the easiest disclosure to remove. A spoken prefix is the first few seconds of a file. Anyone reposting a track can trim it in a free editor in seconds. Metadata can be stripped too, but a platform label backed by detection is reapplied every time the track is uploaded again; an audio prefix, once cut, is gone for good. The WhatsApp argument cuts both ways.

It punishes the honest. A musician who declares "Partly Gen AI" because they used an AI drum stem would, read literally, have to put a voice in front of their song. One who lies about it would not. A rule that makes truthful declaration costly will get fewer truthful declarations, and the self-declaration the regulator already distrusts gets worse.

It tells the listener less. "This audio is synthetic" says nothing useful about a song where a human wrote, sang and mixed everything and an AI tool generated the string pad. YouTube's three tiers, and credit-style disclosure of which part was AI (our earlier argument for disclosure as a credit), carry more information than a spoken line ever could.

Music is not the harm the rule was written for. The "indistinguishable from a real person or real-world event" test points at deepfakes. An AI song in a made-up voice deceives nobody about a real event. A song cloning a real singer's voice is a different case, and there we would support the audible warning.

What we would ask MeitY to write

A music route in its guidance, not a carve-out:

1. For tracks that clone or imitate an identifiable real voice, keep the prefixed audio disclosure.
2. For other AI-assisted music, accept a visible on-platform label plus embedded, standard metadata as compliance, and require platforms to carry that metadata through downloads and shares.
3. Ask platforms to publish how often detection overrode a partner's declaration, so the self-declaration problem is measured rather than assumed.

For Indian musicians now, the practical step is the same either way: declare AI use honestly at your distributor, keep your project files, and expect a platform to ask. The rules are a draft of a norm that is still forming, and it is better to be on record as the honest declarer when it settles.