India must make AI labels durable and verifiable

A few years ago, a San Francisco family received a call from their son claiming he had been in an accident. A lawyer urged them to wire money. They did. But there had been no accident, and it had not been their son speaking. The voice was AI-generated.

Such scams prompted California's AI Transparency Act, which took effect on August 2. India had notified similar rules 6 mths earlier. Both governments reached the same conclusion: when people can no longer distinguish real from synthetic, public trust erodes.

California's law applies to gen AI systems with over 1 mn monthly users, covering major AI labs and large image, audio and video platforms. Covered providers must do 3 things:


Embed a latent disclosure, an imperceptible mark that software can read, which carries provenance data, in generated images, audio and video. The statute requires that mark to be permanent or extraordinarily difficult to remove.

Offer users option of a visible label.

Operate a free, publicly accessible detection tool so that anyone can check a file against that provider's system at no cost, and without an account.

The penalty is $5,000 per violation, with each day of non-compliance counting as a fresh offence. Enforcement rests with the attorney general, city attorneys and county counsel, with costs and legal fees recoverable by a successful plaintiff. From January 1, 2027, obligations extend from model providers to platforms that host and distribute AI-generated content.

Two details matter:

Obligation sits on the generator, at the point of creation, where marking is cheap and evasion hardest.

California moved its start date to August 2 to land on the day Article 50 of EU AI Act began to apply. The latter requires synthetic output to be marked in a format that machines can read.

India was ahead of both. IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026, effective from February 20, gave synthetic media a statutory identity - 'synthetically generated information' - and defined it as AI-created or altered audio, visual or audiovisual content that appears real and is likely to be indistinguishable from an actual person or event. Routine editing, subtitling, translation, compression and ordinary document preparation are exempt. This is a sensible distinction.

Obligations are demanding.

Tools that generate synthetic media must deploy technical measures to prevent unlawful output, label what they permit, embed permanent metadata or provenance mechanisms, including a unique identifier, and must not offer features that strip those marks.

Large social platforms must obtain a user declaration before publication, and verify it technically, rather than take uploader's word. Content flagged by a court or an authorised officer must come down within 3 hrs. Complaints about morphed intimate imagery or impersonation must be acted on within two.

That is a serious framework. Its weakness is not ambition but durability. A visible label lasts until someone crops it. Filming a video off a screen removes it. Signed file metadata is stripped constantly, usually automatically and without any malicious intent, by social media processing pipelines, screenshots, messaging compression and conversion between formats.

Five additions are required:

Any provider offering image, audio or video generation to Indian users above a threshold scale should embed an imperceptible watermark that software can read in every output at the point of creation, in addition to any visible label.

Adopt EU's 4 technical standards alongside California's requirement that AI labels be permanent, or extremely difficult to remove. Ask India's standards and certification agencies to set measurable benchmarks and certify compliant technologies, giving intermediaries a safe harbour they can rely on.

A mark only its creator can read is transparency in name only. Courts, fact-checkers, Election Commission and ordinary citizens need a tool that costs nothing to use.

Compliance should require either a detection standard that's publicly available or licensed on fair, reasonable and non-discriminatory terms to certified Indian verifiers, or participation in an open provenance framework backed by a certified durable watermark.

Otherwise, India's ability to authenticate evidence in its courts will depend on the discretion of foreign firms and shut Indian companies out of a market the IndiaAI Mission aims to build.

Phase the duties - providers within 12 mths of notification; then, platforms required to preserve provenance through their processing pipelines and surface it to users, from a date aligned with California's January 2027 obligations.

The global AI industry is building durable labelling to comply with laws in Europe and California. India should require the same, at little extra cost. It moved early to regulate synthetic media. But that advantage will count for little if AI labels disappear before a forwarded video reaches a citizen's phone.

The writer is programme director (Eastern India), WHEELS Global Foundation