India's AI Labelling Rules (2026): What Every AI Site Must Show

Last checked: 7 October 2026. This is not legal advice. We run an AI chat product in India and wrote this summary for ourselves; it is a plain-language reading of public analysis, not a substitute for a lawyer. Primary sources are linked throughout.

Key takeaways

• Notified 10 Feb 2026, in force 20 Feb 2026.
• "Synthetically generated information" is audio, visual and audio-visual — text is excluded.
• Four different clocks: 3 hours, 2 hours, 36 hours, 7 days.
• The proposed 10% watermark rule was dropped in favour of "clearly and prominently labelled".
• Safe harbour is the lever — non-compliance puts it at risk.

What counts as "synthetically generated information"

The definition is narrower than most headlines suggest. Per Hogan Lovells' summary, SGI is:

audio, visual, or audio-visual information that is artificially or algorithmically created, generated, modified, or altered using a computer resource in a manner that appears real and authentic, portraying individuals or events as indistinguishable from actual persons or real-world occurrences

Two things follow.

Text is outside it. Freshfields' analysis states plainly that text generated by AI does not fall within SGI. So a text-only chat product's labelling obligations are materially narrower than those of an image, voice or video generator.

That is not a blanket exemption. The labelling duty is one obligation among several. Takedown timelines, grievance handling and the general due-diligence obligations under the IT Rules apply regardless of whether your output is text. Reading "text is excluded" as "the rules don't apply to us" would be a mistake.

The rule is also aimed at a specific harm: synthetic media that appears real and authentic and portrays real people or events. A clearly fictional character is a different thing from a convincing fake of a real person — which is precisely why we don't build characters from living people.

The four takedown clocks

Deadline	Applies to
3 hours	Court orders and government notices to remove unlawful content (previously 36 hours)
2 hours	High-risk content: nudity or sexual acts, non-consensual intimate imagery, impersonation, artificially morphed images
36 hours	Complaints relating to content concerning a specific individual
7 days	General user grievances

The two-hour window is the one to design around. It covers impersonation and morphed imagery, which is exactly the category a character-AI platform could drift into if it allowed likenesses of real people. Two hours is not enough time to build a process after the first complaint arrives — it has to exist beforehand.

Labelling: what the final rules actually require

The draft proposed hard numbers: a watermark covering at least 10% of the surface area for visual content, or a disclaimer across the first 10% of the duration for audio. Both were dropped.

The final requirement is that SGI be "clearly and prominently labelled", and embedded with permanent metadata or provenance markers — including identifiers linking the content back to the intermediary's systems.

Qualitative standards are easier to meet in spirit and harder to measure. In practice: a label a normal person would notice without looking for it, that survives the content being downloaded and re-shared.

Who is a "significant social media intermediary"?

An SSMI is a platform with more than 50 lakh (5 million) registered users in India. Crossing that threshold brings additional obligations.

Most Indian AI products are nowhere near it — but the threshold is about registered users, not daily actives, so it arrives sooner than teams expect.

Why safe harbour is the real stake

The enforcement mechanism isn't a fine. It's safe harbour — the protection under Section 79 that stops a platform being treated as the publisher of what its users post. Lose it and every piece of user content becomes a potential direct liability.

That is why the compliance question is worth taking seriously even at small scale.

What this means for a text-only AI chat product

Speaking only about our own position, and without claiming a clean bill of health:

• We label every character as AI. It is on the character card, the chat screen and our AI disclaimer, which opens: "Every character is an AI. None of them are real people — even the famous ones. They make things up, confidently."
• We don't generate images, audio or video, so the SGI labelling duty largely doesn't reach our output today. If we ship voice, that changes immediately — original synthetic voices only, never a clone of a real person.
• We don't build characters from living people. That keeps us away from the two-hour impersonation clock by design rather than by process.
• We have a report flow on every chat, and a published content policy.

We are not claiming to be fully compliant. That is a judgement for counsel, not for a blog post, and any platform telling you otherwise on its own website is overreaching.

Primary sources

• Hogan Lovells — mandatory labelling and 3-hour takedown
• Freshfields — key changes under MeitY's 2026 amendments

For anything that matters, read MeitY's notification itself and talk to a lawyer.