India's AI Deepfake Labelling Rules: What Changed on Feb 20

India has just become one of the first countries with binding rules forcing artificial intelligence to wear a name tag. From 20 February 2026, the IT (Amendment) Rules, 2026 — notified by the Ministry of Electronics and Information Technology on 10 February — require that AI-generated photos, videos and audio circulating online carry a clear, machine- and human-readable label. If you make memes, edit reels, run a page, or simply scroll, the AI deepfake labelling rules quietly change what you see and what you are allowed to post.

This is not a vague advisory. It plugs a new category of content directly into India's intermediary law, with deadlines measured in hours and consequences that reach down to the individual user. Here is what actually changed, in plain language, and what you should do about it.

What the law now calls a deepfake

The rules introduce a precise term: synthetically generated information, or SGI. It covers audio, visual or audio-visual content that is artificially or algorithmically created, modified or altered so that it appears real, authentic or true — and is likely to be perceived as indistinguishable from a real person or event.

That definition is deliberately broad. A face swap, a cloned voice note, an AI-generated 'photo' of a politician, a fabricated bank-CEO video — all qualify. So does a genuine clip altered enough to mislead. Crucially, the test is whether an ordinary viewer could mistake it for reality, not whether the tool used was famous or obscure.

What the rules do not target is harmless, obviously creative AI: cartoons, clearly stylised art, or content nobody would confuse for a real event. The whole regime turns on that line between deception and expression.

The label rule — and the 10% twist

The headline obligation is simple: permitted AI content must be labelled. But the most-discussed part of the original draft was quietly killed.

The October 2025 draft had proposed that every label cover at least 10% of the visual frame, or the opening 10% of an audio clip's duration. Creators and platforms pushed back hard, arguing it would deface legitimate work. The final notification dropped the fixed percentage entirely.

In its place is a flexible, qualitative standard. Labels must now be:

• Prominent, easily noticeable and adequately perceivable for visual content — no exact size mandated.
• A clearly audible disclosure prefixed to the start of audio content.
• Backed, where technically feasible, by embedded provenance metadata that travels with the file and is not easily stripped out.

That last point matters most. A visible badge can be cropped; invisible provenance data baked into the file is meant to survive downloads and re-uploads, so detection tools can still flag the content later. This is the real spine of the regime.

What changes when you hit 'upload'

For ordinary users, the friction shows up at the moment of posting. Large platforms — the significant social media intermediaries (SSMIs) with millions of Indian users — now have to do three things.

1. Ask you. At upload, you will be prompted to declare whether what you are posting is synthetically generated.
2. Verify you. The platform must cross-check your answer using reasonable, automated detection tools — it cannot simply take your word.
3. Override you. If its tools conclude the content is AI-made and you said it was not, the platform must apply the label anyway.

In other words, a self-declaration that you tick is now a legal statement, not a formality. Falsely marking a deepfake as 'real' content does not get it off the hook; it just adds a layer of misrepresentation on top.

Platforms must also send users an advisory every three months, reminding them that unlawful synthetic media can trigger account suspension, disclosure of identity to authorities, and referral to law enforcement.

The three-hour clock

The rules tighten takedown timelines dramatically. Where the old framework gave intermediaries up to 36 hours to act on flagged content, the amendment introduces a fast lane for the worst material.

• Content deemed unlawful by a court order or government direction must be removed within 3 hours.
• Complaints involving nudity, sexual content, morphed imagery or impersonation must be acted on within 2 hours.

For a victim of a non-consensual deepfake — overwhelmingly women — that compression from a day and a half to a couple of hours is the single most consequential change. Speed is the whole point, because a viral morph does its damage in the first few hours, not the second day.

Platforms are also expected to deploy automated measures to stop the very creation and spread of clearly illegal SGI, including child sexual abuse material, non-consensual intimate imagery, forged government documents, and deceptive impersonation.

Who carries the burden

The rules split responsibility across the chain, and it helps to know where you sit.

• AI tool providers (the apps and models that generate content) must label what they produce and embed provenance signals at the source.
• Social platforms must collect declarations, verify them, label confirmed SGI, and run the takedown clock.
• Users and creators must declare honestly and avoid generating prohibited content.

The genuine worry, flagged by digital-rights groups, is over-removal. When platforms face a two- or three-hour deadline and legal liability, the safe move is to take content down first and ask questions later. Satire, journalism, political commentary and legitimate AI art could get swept up by nervous automated filters — a real free-speech cost the rules do not fully resolve.

What this means for you

You do not need a law degree to stay on the right side of this. A few practical habits cover most situations:

• Declare AI content honestly when a platform asks. The label is now expected, not optional.
• Keep the provenance intact — don't strip metadata or crop out labels from AI media you share.
• Treat 'AI-generated' tags as a trust signal, not a stigma; their absence on a shocking 'leaked' clip is itself a red flag.
• If you are targeted by a deepfake, file a complaint immediately and cite the 2-hour category for impersonation or intimate imagery — the faster clock is your leverage.
• Brands and creators using AI in ads or campaigns should build labelling into their workflow now, not retrofit it after a complaint.

The bigger picture

India is wagering that transparency, not prohibition, is the workable answer to synthetic media. The bet is that if every AI artefact is tagged and traceable, the public can keep trusting what it sees without the state banning the technology outright.

Whether it works depends on enforcement and on whether detection tools are good enough to avoid mislabelling real content as fake — and vice versa. Expect friction, court challenges and a few high-profile mistakes in the first months. But the direction is set: in India, AI content now has to introduce itself. For a country heading into a dense election and festival calendar, where a single convincing fake can move crowds, that label may be the most important two words on your screen.