India's synthetic media rules: label it, and take it down on the right clock

01The clock

India did not pass an AI act. It amended the intermediary rules that already governed platforms, which is why the duties here attach to hosting and distribution rather than to model development — and why the sanction is the loss of an immunity rather than a fine.

1. 22 October 2025 The draft is published

   It proposed a label covering at least ten per cent of the display area, and for audio the initial ten per cent of its duration. Neither figure survived.

2. 15 November 2025 A separate amendment takes effect

   G.S.R. 775(E) rewrote rule 3(1)(d) with a 36-hour window. This is a different instrument, and it is frequently conflated with the synthetic-media rules that followed.

3. 10 February 2026 G.S.R. 120(E) is notified

   The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, published in the Gazette of India, Extraordinary. Some alerts date this to 16 February; the Gazette says the tenth.

4. 20 February 2026 · in force The rules commence

   Labelling, embedded provenance, user declarations, verification for significant social media intermediaries, and the shortened takedown windows all begin.

02What counts as synthetically generated information

The test is cumulative. Content must both appear real and be indistinguishable — or likely perceived as indistinguishable — from a real person or event. Obviously stylised output that nobody would take for reality does not meet the second limb.

Text is not covered. The definition reaches audio, visual and audio-visual information only, and the ministry's own published questions and answers state that pure text and written outputs are not, by themselves, synthetically generated information. A number of law-firm alerts and news pieces say otherwise. They are wrong, and the point is not marginal — it removes every large language model's text output from the regime.

A proviso further excludes routine, good-faith work: editing, formatting, enhancement, technical correction, colour adjustment, noise reduction, transcription or compression that does not materially alter substance, context or meaning; the ordinary creation of documents, presentations, educational materials and research outputs, provided no false record results; and use of a computer resource solely for accessibility, clarity, quality, translation, description or searchability without manipulating any material part.

03The labelling duties

04Declarations and verification

The rules push part of the classification problem onto users, and then require the larger platforms to check the answer.

An intermediary must require a user uploading content to declare whether it is synthetically generated, and significant social media intermediaries must additionally deploy reasonable and appropriate technical measures — including automated tools — to verify that declaration. Where the check indicates the content is synthetic, it must be displayed with the required label.

Two consequences follow. First, a declaration is an event with a timestamp and an author, and the platform's own verification is a second event with a method and a result — two facts that together decide whether the platform behaved reasonably, and that exist nowhere unless recorded. Second, the rules protect platforms that act: removal or disabling done in compliance with the rules, including through automated tools, does not itself break the conditions of the safe harbour. Acting on a positive detection is safe; the risk sits in not acting, and in not being able to show what the check returned.

05Takedown runs on more than one clock

This is the part that most coverage reports backwards. There is no single takedown window. There are four, they run from different starting events, and the spread between the fastest and the slowest is eighteen-fold.

Three corrections worth making, because each one is widely published the other way round:

36 hours is not the court-order window. It used to be. This amendment cut the court-and-government route to three hours, and the 36-hour figure now belongs to a grievance window that was itself cut from 72 hours. Anyone still treating 36 hours as their response time for a court order is operating on a clock that expired twelve times over.

The 36-hour window excludes intellectual property. The carve-out for sub-clauses (i), (iv) and (xi) takes third-party rights, patent and trademark and copyright infringement, and general unlawfulness out of the fast window — so copyright takedowns fall back to seven days.

The two-hour duty is softer in kind than the three-hour one. Rule 3(2)(b) asks for all reasonable and practicable measures to remove or disable access. Rule 3(1)(d) is an absolute removal obligation. A shorter clock is not automatically a stricter duty.

Across an eighteen-fold spread, a guessed window is a manufactured finding. The classification of the notice is the whole question — which is why the record has to capture the class that was assigned, when the notice arrived, and when removal completed, rather than computing a verdict against a window nobody agreed on.

06What failure actually costs

Do not budget for fines. These Rules impose no penalty and create no offence. This is the single most misreported point about the regime.

Mechanically: section 79 conditions an intermediary's immunity on observing due diligence and such other guidelines as the Central Government prescribes. These Rules are those guidelines. So "due diligence" for safe-harbour purposes simply is compliance with them, and breach means the immunity ceases to be available — leaving the intermediary exposed to primary liability for content posted by other people.

That is a different risk shape from every other regime on this site, and it changes what evidence is for. There is no regulator order, no hearing, and no appeal against "losing safe harbour." It is a defence that falls away, and it is adjudicated after the fact, in whatever proceeding somebody else brings, possibly years later. You will not be told you have lost it. You will discover it when you try to rely on it.

A defence you cannot evidence at the moment you need it is not a defence. That is the whole argument for contemporaneous records here — not a fine to avoid, but an immunity to be able to demonstrate you were entitled to.

One amplifier for larger platforms: a significant social media intermediary is deemed to have failed due diligence where it knowingly permitted or promoted non-compliant synthetic content, or failed to act on it.

07Duty → receipt

This is where Planisphere enters, and not before. One event per act — and, uniquely among the regimes here, a pair of events that are bound together by a statutory clock.

Note prominence_basis. Because the notified rules dropped the percentage, prominence is a judgement rather than a measurement — so the field records the basis on which it was claimed, at the time it was claimed.

The takedown pair is the part built specifically for this regime. The notice record carries a notice_class, and the completion record points back to it with starts_from_action_key. The elapsed interval is then computed from two sealed records and measured against the window that class carries — two hours, three hours or thirty-six — rather than asserted afterwards by whoever is writing the response.

And where the class is not one the pack recognises, it records the interval and returns no verdict. That restraint is deliberate: across an eighteen-fold spread, guessing which window applied would manufacture a finding rather than report one.

08The integration

Two calls, bound together. First the notice, which starts the statutory clock:

Then the completion, pointing back at it:

The full API reference covers authentication, idempotency, batching and rate limits; pricing is per record with a monthly free allowance.

10Questions

Is the ten per cent label rule real?

Not any more. It was in the October 2025 draft and did not make it into the notified rules, which contain no percentage at all. Explainers still asserting it are quoting a draft. The notified standard is prominence — easily noticeable and adequately perceivable — and for audio, a prominently prefixed disclosure with no minimum duration.

Are we covered if we only generate text?

No. The definition reaches audio, visual and audio-visual information, and the ministry's published questions and answers state that pure text outputs are not synthetically generated information by themselves. Note this is a narrower position than the EU takes, where the marking duty covers text on the face of the provision.

We are not based in India. Does this reach us?

The rules bind intermediaries, and the intermediary regime has long applied to services offered to users in India regardless of where the company sits. The practical question is usually whether you are an intermediary at all, and whether you are a significant social media intermediary, which carries the additional verification duty.

Is there any enforcement to look at?

None that is public. We found no rule 7 proceeding, no safe-harbour revocation and no reported challenge to the amendment. Treat that as an absence of public record rather than proof of inactivity — enforcement here is characteristically not announced, and the sanction surfaces inside private litigation rather than in a regulator's press release.

Is anything else coming?

A second amendment was published in draft in March 2026 and revised in April, with consultation closed in May. As of this writing it has not been notified. We will not describe a draft as law; check the Gazette before relying on it.

11Sources

Every rule number, window and quotation on this page was read from the signed Gazette notification rather than a summary of it.

• Primary G.S.R. 120(E) — the Amendment Rules, 2026, as published in the Gazette Notified 10 February 2026, in force 20 February 2026. The definition at rule 2(1)(wa), the labelling and provenance duties at rule 3(3), and the absence of any percentage figure.
• Primary MeitY — the consolidated IT Rules and the ministry's questions and answers The confirmation that pure text outputs are not synthetically generated information. Note the consolidated text still prints words in rule 3(1)(b)(v) that were struck down by the Bombay High Court in 2024, without annotation.
• Information Technology Act, 2000 — section 79 The safe harbour that rule 7 withdraws, and the due-diligence condition that these Rules supply.