India IT Rules 2026 Amendment Guide: AI labelling, deepfake duties and three-hour takedowns

Since 20 February 2026, any platform that lets people in India create, host or share realistic AI-generated audio, images or video has hard legal duties to label it, trace it and remove unlawful versions within hours. The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, notified by the Ministry of Electronics and Information Technology (MeitY) on 10 February 2026, amend the IT Rules, 2021 to define "synthetically generated information" (SGI) and to cut takedown clocks from 36 hours to 3 hours. This guide explains who is in scope, what each obligation requires in engineering and operations terms, how the rules fit with provenance standards such as C2PA, and how to build a compliance programme that survives the new timelines. It is written for trust and safety leads, security and platform engineers, product counsel and GRC teams at intermediaries, AI tool providers and global platforms with Indian users.

Source note: the official MeitY and e-Gazette text could not be read when this guide was prepared, so dates, timelines and rule numbers are taken from law-firm analyses (AMS Shardul, Khaitan & Co, Lexplosion, Nishith Desai Associates) that agree on the points used here. Confirm against the official text before relying on it.

Key takeaways

• SGI is defined by realism, not by the tool. Audio, visual or audio-visual content created or altered algorithmically so that it appears real and is likely to be perceived as a real person or event is in scope; pure text, good-faith editing and accessibility uses are out.
• Lawful SGI must be labelled and traceable. Visual content needs a prominent visible label, audio needs a prefixed audio disclosure, and permanent metadata with a unique identifier must be embedded where technically feasible.
• Clocks are now measured in hours. Lawful takedown directions must be met within 3 hours (was 36), non-consensual intimate imagery complaints within 2 hours (was 24), and general grievances within 7 days (was 15).
• Significant social media intermediaries carry extra duties. Platforms above the 50 lakh (5 million) Indian user threshold must collect a user declaration, verify it with technical measures and label confirmed SGI before publication.
• Safe harbour is the stake. Failure to observe the Rules removes protection under Section 79(1) of the IT Act, 2000, exposing the intermediary to liability under the IT Act and the Bharatiya Nyaya Sanhita, 2023.

Status, scope and where the rules sit

The amendment changes the IT Rules, 2021, which are made under the Information Technology Act, 2000. It is in force: MeitY notified it on 10 February 2026 and it took effect on 20 February 2026, giving intermediaries about 10 days to implement. A draft published in October 2025 had proposed that visible labels cover at least 10% of the surface of visual content; the final text dropped the fixed size in favour of a qualitative "prominent" standard and narrowed the definition to exclude routine and good-faith uses. MeitY has also published FAQs to explain the definition. This guide was checked against law-firm analyses of the notified text in November 2026 because the MeitY and e-Gazette pages could not be read directly; confirm sub-rule numbering against the official gazette before quoting it in a contract or filing.

The rules apply to intermediaries as defined in the IT Act. In practice that reaches four groups:

• AI generation services: tools that let users create or alter realistic audio, images or video, which carry the preventive, labelling and metadata duties.
• Hosting and sharing platforms: social media, video, messaging, marketplace and community services that enable publication, transmission or sharing of SGI.
• Significant social media intermediaries (SSMIs): social media intermediaries above the 50 lakh registered user threshold, which carry the declaration and verification duties in Rule 4(1A).
• Enterprises: organisations not acting as intermediaries are not directly regulated by these rules, but they are the victims of deepfake fraud and will use the faster complaint routes. See the Deepfake and AI Impersonation Guide.

The amendment sits alongside the wider Indian framework covered in the India IT Act 2000 and SPDI Rules guide and the non-binding approach in the India AI Governance Guidelines.

What counts as synthetically generated information

The new definition (Rule 2(1)(wa) in the analyses read) covers 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, authentic or true, and depicts a person or event so that it is, or is likely to be, perceived as indistinguishable from a natural person or real-world occurrence. MeitY's FAQs describe the test as technology-neutral: what matters is realistic resemblance capable of deceiving, not the model used.

Two grey areas need written internal criteria: satire and parody, and multimodal content where a real clip is paired with a misleading synthetic caption or audio track (MeitY FAQ 21 treats multimodal deception as covered).

Obligations by role

Rule 4(4) on proactive technical measures for SSMIs moved from "shall endeavour to deploy" to an obligation to deploy appropriate measures, which commentators read as a continuing duty to keep detection current.

The new clocks

A 3-hour clock cannot be met by a queue that a lawyer reads the next morning. It needs a 24x7 intake channel for government and court notices, an authenticated sender list, pre-approved decision rules for common categories, the ability to geo-restrict or disable a specific item and all its re-uploads, and a timestamped record of receipt and action. Industry bodies have raised feasibility concerns, and commentators warn of over-removal; neither changes the current obligation.

Building labelling and provenance that holds up

The rules do not name a standard, but the requirement for a visible label, embedded metadata and a unique identifier tied to the creating resource maps closely to existing provenance technology. NIST AI 100-4 groups the options into provenance data tracking, watermarking, metadata recording and synthetic content detection, and is blunt that no single technique is sufficient: metadata is often stripped when files are shared, most watermarks can be removed, and detectors degrade on unfamiliar generators and compressed media.

Practical design: generate the label and the manifest at the point of creation, not at upload, so content leaving your service already carries both. Preserve incoming C2PA manifests and watermarks through your own transcoding pipeline, because stripping them is the behaviour the rules prohibit. Use the manifest or watermark as a positive signal in the SSMI verification step, and route undeclared, high-confidence detections to labelling rather than deletion when the content is lawful. See the AI Content Provenance and C2PA Guide and the AI Model Provenance and Watermarking Guide.

Where identity fits

Provenance is only as trustworthy as the keys that sign it. A C2PA claim signature proves which signing identity attested to the content, so the signing keys and certificates used by generation services are high-value machine credentials: an attacker who steals them can make forged media look like it came from your tool, and a leaked key forces revocation across everything you signed. Hold signing keys in an HSM or cloud KMS, scope each key to one service, rotate on a fixed schedule and log every signing operation. The same discipline applies to the API keys and service accounts that connect detection, labelling and takedown systems. See the Cryptographic Key Management Guide and the Code Signing Security Guide.

The rules also allow disclosure of a violating user's identity to a victim in line with applicable law. Treat those requests as you would law enforcement data requests: verify the requester, record the legal basis, and release the minimum data, keeping the India DPDP Act Guide obligations in view.

How to implement: a phased programme

1. Classify your role (week 1): record whether you are an intermediary, whether you enable SGI creation, and whether you are an SSMI. Document the reasoning; it decides which rows of the obligations table apply.
2. Stand up the clock (weeks 1 to 2): a 24x7 monitored intake for court orders and government intimations, a named on-call decision maker, runbooks for the 3-hour and 2-hour categories, and timestamps on receipt, decision and action. Test with tabletop drills before relying on it.
3. Update terms and notices (weeks 1 to 2): add SGI misuse warnings, the user declaration requirement and the quarterly notice schedule, in English or an Eighth Schedule language.
4. Label and sign at creation (weeks 2 to 6): add visible and audio disclosures and C2PA manifests to generation outputs. Pilot on one product surface, measure label persistence after download and re-upload, then expand.
5. Preserve provenance in the pipeline (weeks 2 to 6): audit every transcoder, thumbnailer and CDN step for metadata stripping. Fix or record technical infeasibility with evidence.
6. SSMI declaration and verification (weeks 4 to 10): add the declaration control at upload, run detection in shadow mode first to measure false positive rates, then turn on labelling. Keep a manual override with reason codes.
7. Prevention filters (ongoing): prompt and output filters for CSEAM, intimate imagery, forged documents and deceptive impersonation, with hash matching for known abuse material.
8. Rollback and review: any detector or filter change ships behind a flag so it can be reverted if false positives spike; review the whole programme quarterly against MeitY FAQs and any further amendment.

Decision guidance

• Visible label versus metadata only: the rules require both for lawful SGI. Metadata alone fails the visible label duty; a label alone fails traceability.
• Label versus remove: lawful SGI is labelled, not removed. Removal is for prohibited SGI or content subject to an order. Over-removal creates its own user grievance volume against a 7-day clock.
• Build versus buy detection: small platforms usually license detection and rely heavily on declarations and incoming provenance; large SSMIs need in-house evaluation sets, including Indian languages and regional accents, because detectors perform worse outside the data they were trained on.
• Global platforms: one provenance architecture can serve India, the EU AI Act Article 50 and the China AI labelling rules, but the label wording, takedown clocks and declaration steps differ by jurisdiction.

Detection, measurement and evidence

• Log: receipt time, sender verification, decision, action time and scope for every order and complaint; declaration value and detector score for every SSMI upload; signing events for every manifest.
• Metric 1: percentage of takedown directions actioned within 3 hours, and of intimate imagery complaints within 2 hours (target 100%).
• Metric 2: median and 95th percentile time from receipt to action.
• Metric 3: label persistence rate: share of your generated SGI still carrying its label and manifest after a download and re-upload test.
• Metric 4: declaration mismatch rate: undeclared uploads later confirmed as SGI.
• Metric 5: detector false positive rate on a labelled Indian evaluation set, reviewed monthly.
• Metric 6: grievances resolved within 7 days.

Common failure patterns

• Business-hours legal queue: orders arriving at night miss the 3-hour window. The clock needs on-call ownership, not an inbox.
• Your own pipeline strips the label: transcoding drops C2PA metadata and the platform ends up enabling the suppression the rules prohibit.
• Treating text-only exemption as a general AI exemption: a chatbot that also generates images or voice is in scope for those outputs.
• Removing whole accounts without preserving evidence: the rules require action without vitiating evidence, and police may need it.
• Unprotected signing keys: generation services sign manifests with keys in configuration files, so a leak makes provenance worthless.
• Assuming deepfake harm is reputational only: realistic impersonation drives payment fraud, as the Arup deepfake fraud showed when a synthetic video call with a fake CFO led to transfers of HK$200 million. Fast takedown and clear labels matter to the enterprises targeted.

Framework mapping

Practitioner checklist

• Record your intermediary and SSMI status with reasoning.
• Run a 24x7 intake and decision rota able to act on orders in 3 hours and intimate imagery complaints in 2 hours.
• Publish SGI misuse warnings and schedule user notices at least every three months.
• Add visible and audio disclosures plus C2PA manifests to all realistic generated media at creation.
• Test that your upload and transcoding pipeline preserves labels and metadata.
• For SSMIs, ship declaration, verification and labelling at upload, with detector shadow mode first.
• Protect manifest signing keys in an HSM or KMS and log every signing operation.
• Preserve evidence before removal and document identity disclosures to victims.
• Check the official gazette and MeitY FAQs quarterly for amendments.

Standards and references

• MeitY: Acts and Policies (IT Rules, 2021 and amendments)
• Khaitan & Co: MeitY notifies the IT Amendment Rules 2026
• Nishith Desai Associates: Deepfakes, Disclosure and Due Diligence, Unpacking India's SGI Amendments
• Shardul Amarchand Mangaldas: Legislative Update, IT Rules Amendment
• NIST: AI 100-4 Reducing Risks Posed by Synthetic Content
• C2PA: Content Credentials Technical Specification 2.2
• Lexplosion: MeitY notifies IT Amendment Rules 2026 on synthetically generated information

Related NHI Mgmt Group resources: India IT Act 2000 and SPDI Rules · India AI Governance Guidelines · AI Content Provenance and C2PA Guide · Deepfake and AI Impersonation Guide. This guide summarises the IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026 and is not legal advice.