On 30 March 2026, the Ministry of Electronics and Information Technology (MeitY) published the draft IT (Intermediary Guidelines and Digital Media Ethics Code) Second Amendment Rules, 2026. The consultation closed on 7 May 2026 — extended once already from the original 14 April deadline — and the volume of pushback explains the extension. SFLC.in has called for the draft's complete withdrawal. The Internet Freedom Foundation titled its first read "Sound the Alarm." The Editors Guild of India and DIGIPUB News India Foundation jointly demanded an unconditional withdrawal. The complaint is not that India should ignore deepfakes. It is that the draft asks platforms to do something that does not work, in a way that will end up censoring lawful speech.
The harm that motivates the draft is real
Before dismissing the rules, the case for them deserves a fair hearing. India ran a national election cycle in which AI-cloned audio of politicians circulated within hours of recording. Non-consensual intimate imagery generated from a single social-media photo now reaches victims' families before any takedown request can be filed. Voice-clone fraud has hollowed out elderly savings in cases reported across Mumbai, Bengaluru and Hyderabad. The Supreme Court's 2015 ruling in Shreya Singhal v. Union of India protected online speech from vague criminal sanctions, but it did not absolve the state of any duty to address downstream harms from a technology that did not exist when Section 79 of the IT Act was written. A regulator that did nothing about synthetic media would be failing.
The February 2026 amendment, notified on 10 February and in force from 20 February, did the foundational work: it defined "synthetically generated information" (SGI) as audio, visual or audio-visual content that is artificially or algorithmically created and appears indistinguishable from a natural person or real event. It introduced labelling, embedded provenance metadata, and a three-hour takedown window for court orders or notices from a Joint Secretary-rank official, per the Freshfields analysis of the rules. Significant social media intermediaries — those with five million or more Indian users — were assigned a heavier load: requiring user declarations, deploying automated verification, and posting prominent on-content notices.
The draft asks for a label that cannot persist
The Second Amendment Rules push past what is technically achievable. Draft Rule 3(3)(a)(ii) replaces the earlier "prominent visibility" wording with a requirement that a label be displayed "continuously and clearly visibly throughout the duration of the content" — that is, baked into every frame of every synthetic video, end-to-end. The phrasing is reported verbatim in SocialSamosa's coverage of the draft and confirmed in the IAPP's Asia-Pacific roundup.
The problem, as SFLC.in's formal comments note, is that any such overlay is defeated by a screenshot, a screen recording with a crop, or a trivial pass through a re-encoder. Provenance metadata, the more durable alternative, is itself stripped by every social-media compression pipeline currently in production. SFLC.in points to the Coalition for Content Provenance and Authenticity (C2PA) standard, which Adobe, Microsoft, Google, OpenAI and the BBC have already adopted, as a less brittle approach — but C2PA is a cryptographic credential, not a burnt-in visual label, and it relies on viewing software that respects the credential. MeitY's draft conflates the two.
For large platforms, the requirement is expensive but compliable on the visual layer. For smaller intermediaries — the regional news aggregator, the Indic-language video host, the open-source federated network with no detection budget — the rule is a wall. The LiveLaw analysis describes the operational asymmetry plainly: automated detection accuracy of "at least 50%" with quarterly audits, hard three-hour clocks, and safe-harbour loss for non-compliance. A 50% accuracy floor on a recall-precision tradeoff weighted toward removal is not a fact-finding tool. It is a deletion tool with a name.
The chilling effect is structural, not theoretical
The constitutional concern is older than the draft. The February rules already prohibit creating a "false document or false electronic record" and content "likely to deceive" — terms with no statutory definition. Satire, parody, political memes, and fictional storytelling all satisfy a literal reading. When the penalty for guessing wrong is loss of safe harbour, the rational platform behaviour is to over-remove. That is the precise pathology Shreya Singhal identified in striking down Section 66A: a vague provision "takes within its sweep protected speech" and chills expression that the Constitution protects.
The Second Amendment compounds this. It expands Part III obligations to cover user-generated "news and current affairs" content, which would draw ordinary citizens commenting on policy into the same publisher regime as registered news outlets. It also makes compliance with non-statutory MeitY advisories a condition of safe harbour — a route around Parliament that the Internet Freedom Foundation and SFLC.in both flag as censorship by proxy: an opaque committee, undefined directives, and platforms incentivised to comply early and broadly.
A proportionate path exists
India does not have to choose between deepfake fraud and over-removal. A workable Second Amendment would (1) drop the continuous visible label requirement and reference C2PA-style cryptographic provenance as the standard, harmonising with the EU AI Act's Article 50 disclosure regime; (2) scale obligations by platform size and by risk category, not impose uniform three-hour windows on every intermediary; (3) keep the takedown framework tied to a judicial or statutory authority rather than executive advisories, preserving the Shreya Singhal threshold; and (4) carve out clear exemptions for satire, parody, journalism and consented synthetic media so platforms are not forced to delete the lawful in order to catch the unlawful.
None of this requires India to soften its position on the underlying harms. It requires the rule text to keep pace with what watermarks and detectors actually do. Without that, the Second Amendment will not stop the next election-eve voice clone. It will only ensure that the platform hosting the takedown response is fined for hosting it too slowly.