India's 3-hour takedown rule: AI audio and video must now carry a label

01Three hours, not 36: four response clocks got shorter

The largest change is in Rule 3(1)(d). When an intermediary gets “actual knowledge” of unlawful content, which under the rule arises only from a court order or a reasoned written intimation from an authorised officer, it must now remove or disable access “within three hours”. The rule had said thirty-six hours.¹³ MeitY’s own example: an intimation received at 11:00 AM must be complied with by 2:00 PM.²

Three other limits move with it. Under Rule 3(2)(b), a complaint from a person about content that exposes them, shows them in nudity or a sexual act, or impersonates them “including artificially morphed images” must now be acted on in 2 hours instead of 24. Removal requests under the proviso to Rule 3(2)(a)(i) fall from 72 hours to 36, and other grievances must be resolved in 7 days instead of 15.¹³

The 3-hour clock is not limited to deepfakes; it covers any unlawful content named in an order. The safeguards on who may issue one came three months earlier: G.S.R. 775(E), in force since 15 November 2025, requires an officer not below Joint Secretary (Deputy Inspector General for the police), a stated legal basis and the exact URL, and a monthly review by a Secretary.⁷ The new rule lets a police administration authorise more than one such officer.¹ The case against is serious. The Internet Freedom Foundation, a day after notification, called the shortened deadline “an impossibly short timeline, which eliminates any meaningful human review”.¹⁰

02What must carry a label: realistic audio, images and video, not text

The new term is “synthetically generated information”, and the definition is narrower than the word “deepfake” suggests.

Three things are carved out: routine or good-faith editing that does not change the meaning; documents, presentations, training material and research outputs that do not create a false document; and tools used only for accessibility, translation, captions or search.¹ MeitY’s FAQs say plainly that “pure text / written outputs, by themselves, are not SGI”, and give examples on both sides: an AI-designed slide deck or subtitles are not covered; a realistic AI presenter, or a synthetic voice that reproduces a real person’s, is.²

Covered content then splits in two. Four kinds must be blocked outright: sexual abuse and non-consensual intimate imagery, false documents or records, material on explosives or arms, and content that falsely depicts a real person or event in a way likely to deceive. Everything else must be labelled so it is “easily noticeable”, with audio carrying “a prominently prefixed audio disclosure”, and embedded with permanent metadata, including a unique identifier, “to the extent technically feasible”. The intermediary may not let anyone remove either.¹ The FAQs spell out the last part: no “remove watermark” or “export without metadata” option.²

The draft’s 10% rule, which would have put a label over a tenth of every frame, is gone.⁴ So is any doubt about text: an AI-written blog post or chatbot reply is not synthetic information under these rules, though it can still be unlawful information under the older ones.²

03Who it binds: the tool, the platform, and your company through both

The rules regulate intermediaries, and two kinds carry the new duties. Rule 3(3) applies to any intermediary offering a computer resource that can create, alter, publish or share synthetic content; MeitY lists AI image and video generators, voice-cloning tools and platforms that help publish such content.¹² Rule 4(1A) adds duties for significant social media intermediaries, those with 50 lakh or more registered users in India: before anything is published, ask the user to declare whether it is synthetic, “deploy appropriate technical measures” to check the declaration, and label it if confirmed.¹⁶ A platform that knowingly lets unlabelled synthetic content through is deemed to have failed due diligence.

Two duties reach every intermediary, synthetic content or not. Users must be told of the platform’s rules and the penalties at least once every three months, where the FAQs say the earlier requirement was once a year.¹² And the price of non-compliance is unchanged: under Rule 7, an intermediary that fails to observe the rules loses the Section 79(1) protection from liability for its users’ content.³

04Europe binds the publisher, China labels text, India binds the platform

The European Union’s version applies later and reaches further. Article 50 of the AI Act, Regulation (EU) 2024/1689, applies from 2 August 2026. Providers of systems that generate synthetic audio, image, video or text must mark the output in a machine-readable format, detectable as artificially generated. Deployers, meaning the businesses that use such a system, must disclose a deep fake they publish; evidently creative or satirical work gets a lighter disclosure, and AI text published to inform the public on matters of public interest must be disclosed unless a person has reviewed it and holds editorial responsibility. Fines reach €15 million or 3% of worldwide annual turnover, whichever is higher.⁸

Two pieces of the EU picture are still moving. The Commission published the first draft of a code of practice on marking and labelling on 17 December 2025.⁹ Its Digital Omnibus proposal of 19 November 2025 would give generative systems already on the market before 2 August 2026 until 2 February 2027 to add marking; that is a proposal before the Parliament and Council, not law.¹¹¹²

China went first. Its Measures for Labelling AI-Generated Synthetic Content (人工智能生成合成内容标识办法), issued by the Cyberspace Administration with three other agencies, have applied since 1 September 2025: a visible label on text, audio, images, video and virtual scenes, plus a hidden label in the file’s metadata naming the provider and a content ID.¹³ The United States has taken the takedown route instead: the TAKE IT DOWN Act of 19 May 2025 gives platforms 48 hours to remove intimate images, including software-made “digital forgeries”, after a valid request, with the process due within a year of enactment.¹⁴ India’s clock for the same harm, on the complaint of the person shown, is 2 hours: a twenty-fourth of that.¹

05What a business should change now

Most of the work is an inventory and a few contract lines. The order below starts with what every company that publishes AI media can do this week and ends with the duties that apply only if you run a platform.