Deepfakes and Synthetic Content: The 2026 IT Rules for Business

What changed

On 10 February 2026 MeitY notified G.S.R. 120(E), amending the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 (the IT Rules), in force from 20 February 2026 according to MeitY's FAQ. An intermediary receives, stores or transmits electronic records for others, or provides a service in respect of them (IT Act s.2(1)(w)): platforms, hosts and many SaaS tools.

A definition. Rule 2(1)(wa) defines synthetically generated information (SGI) as audio, visual or audio-visual information "artificially or algorithmically created, generated, modified or altered using a computer resource" so that it appears real and is likely to be perceived as indistinguishable from a real person or event. Excluded: routine or good-faith editing that does not materially alter meaning; routine documents, presentations and training material that create no false record; and tools used solely for accessibility, translation or searchability. MeitY's FAQ adds that pure text is not SGI (Q8), so text chatbots are outside the definition.

Prevent, label, preserve (rule 3(3)). An intermediary whose resource enables SGI must use reasonable technical measures to block unlawful SGI, including non-consensual intimate imagery, false documents and content that "falsely depicts or portrays a natural person or real-world event" in a manner likely to deceive. Other SGI needs a prominent label, or a prefixed audio disclosure, and, to the extent technically feasible, permanent metadata with a unique identifier of the resource that produced it. The intermediary must not enable removal of either (rule 3(3)(b)). Rules 3(1)(ca) and (cb) add user warnings, covering removal, suspension and, in accordance with law, disclosure of the violator's identity to a victim, and expeditious action on violations.

Declare and verify (rule 4(1A)). Significant social media intermediaries (SSMIs: 50 lakh or more registered users in India, S.O. 942(E)) must, before publication, ask users whether content is SGI, verify the declaration technically and label confirmed SGI.

Faster clocks. Removal on a court order or authorised officer's reasoned intimation: three hours, down from 36 (rule 3(1)(d)). An individual's complaint of nudity, sexual content or impersonation, "including artificially morphed images": two hours, down from 24 (rule 3(2)(b)). Other grievances: seven days, down from 15.

Who is affected

The duties bind intermediaries. In our view four groups are affected, the last indirectly.

1. 1.Vendors of AI image, video or voice tools, which MeitY's FAQ names (Q20). In our view a B2B vendor letting clients generate avatars or cloned voices is an intermediary under rule 3(3) even while it is a data processor for those clients under the DPDP Act.
2. 2.Platforms with an AI feature, such as a photo enhancer. As Trayambic's matrimonial and dating study notes, routine enhancement is excluded; material alteration is not.
3. 3.SSMIs, for declarations and verification.
4. 4.Businesses using AI in marketing, customer service or KYC. A brand making an advert with a third-party tool, or a bank with a synthetic helpline voice, is, in our view, not an intermediary for its own content, so rule 3(3) does not bind it directly. But its tool must label, and SSMIs will demand a truthful declaration. Our view: adopt the labelling standard as policy rather than design around it.

KYC teams face the reverse risk. RBI's KYC Directions for commercial banks already require liveness or spoof detection in video-based identification and checks that it is real-time, not pre-recorded (para 27). Cheating by personation using a computer resource is an offence under IT Act s.66D, which MeitY's AI Governance Guidelines cite for deepfakes.

The DPDP angle

From 13 May 2027 the Digital Personal Data Protection Act, 2023 (the DPDP Act) adds a second layer, whoever generates the content.

• A face or a voice is personal data. Personal data is "any data about an individual who is identifiable" (s.2(t)), and processing includes "adaptation" and "combination" (s.2(x)). In our view a synthetic clip of an identifiable person is personal data about her, though fabricated, and making it is processing.
• Likeness needs consent. A model release or employment contract is not DPDP consent, which must be specific, informed and withdrawable (s.6). In our view the employment ground in s.7(i) does not stretch to cloning an employee for advertising. After withdrawal, use must stop within a reasonable time (s.6(6)).
• Public photos are no licence. Section 3(c)(ii) excludes data a person has "made or caused to be made publicly available". In our view it does not reach a deepfake: the output is new data she never published. Trayambic's article on training AI on personal data discusses the exclusion's limits.
• Disclosing an uploader's identity to a victim under rule 3(1)(ca) needs a DPDP basis. In our view s.17(1)(a), enforcing a legal claim, or s.17(1)(c), offences, fits; no s.7 legitimate use does.
• Grievance routes stack. A person depicted can complain to the platform's grievance officer, appeal to the Grievance Appellate Committee within 30 days (rule 3A), report on the National Cyber Crime Reporting Portal or the 1930 helpline and, from May 2027, use the business's s.13 grievance process before approaching the Data Protection Board.

What to do now

Worked example (figures illustrative). An ed-tech company uses an AI tool to lip-sync 40 teachers' lectures into eight Indian languages. Voice and mouth movements change, so in our view the videos go beyond the translation exclusion and are SGI. The vendor labels them; the company keeps the labels, declares the clips on SSMIs, and takes a separate consent from each teacher for cloning, limited to course content. When a teacher withdraws, her dubbed lectures leave the catalogue within a set period.

1. 1.Inventory synthetic assets: avatars, cloned voices, AI product shots, dubbed videos, voice bots. Record which are SGI and which are excluded, and why.
2. 2.Decide your role per tool. Where users generate audio or video on your platform, build rule 3(3) filters, labels, metadata and warnings.
3. 3.Keep labels and metadata intact through editing, and declare SGI truthfully to SSMIs.
4. 4.Take likeness consent from each person you clone, naming the uses, with a withdrawal route and a plan to retire the asset.
5. 5.Staff the clocks if you host user content: two hours for intimate and impersonation complaints, three for orders.
6. 6.Harden KYC against injected video and cloned voices.
7. 7.Write a short synthetic-media policy covering creation, approval, labelling, consent and takedown.

Key takeaways

• G.S.R. 120(E), in force 20 February 2026, defines synthetic audio, visual and audio-visual content and excludes routine editing, document creation and accessibility or translation tools; pure text is outside.
• Intermediaries offering AI generation tools must block unlawful synthetic content, label the rest with permanent metadata where feasible, and not enable label removal; SSMIs must collect and verify user declarations.
• Takedown clocks are now three hours for court orders and official intimations and two hours for intimate or impersonation complaints, including morphed images.
• In our view, brands, banks and other businesses using AI tools for their own content are not directly bound by rule 3(3), but SSMI declarations and the DPDP Act still reach them.
• In our view a deepfake of an identifiable person is personal data, needs that person's consent, and is not covered by the public-data exclusion in s.3(c)(ii).