On 30 November 2023, as voters queued in the southern state of Telangana to choose a new government, a seven-second clip began travelling through WhatsApp groups across the state. It showed K.T. Rama Rao, a leader of the then-ruling Bharat Rashtra Samithi, calling on people to vote for the opposition Congress. The Congress party’s official X handle posted it. The clip was viewed more than 5,00,000 times before fact-checkers caught up to it. The voice was a generated approximation. The face was real, but the words were not his. A senior Congress leader, speaking anonymously to Al Jazeera weeks later, conceded the obvious: it was AI-generated, and a normal voter would not have been able to tell.
Three years later, on 10 February 2026, the Ministry of Electronics and Information Technology notified an amended set of Information Technology Rules that, for the first time, codified into Indian law a definition of “synthetically generated information” — what most readers would call deepfakes — and assigned legal duties to the platforms that carry them. The rules came into force on 20 February. They are India’s first formal regulatory response to a problem that has been visible to anyone with a smartphone since the 2023 state elections, and they have arrived after a Lok Sabha campaign in 2024 that was widely described, including by Prime Minister Narendra Modi himself, as the country’s first AI-fuelled election.
This is the story of how the gap between technology and trust opened in India, and of the cautious, contested attempt to close it.
What is actually in the rules
The MeitY notification of 10 February 2026 inserts a new definition into the IT Rules: synthetically generated information (SGI) covers any image, audio, video or text that has been “artificially or algorithmically created, generated, modified or altered”. Once that definition exists, the rest of the framework falls into place. Intermediaries — the broad legal category that covers social media platforms, messaging apps and online content providers — now have to label SGI clearly and prominently, embed permanent metadata or unique identifiers where feasible to trace the computing resource that generated the content, take down deceptive or unlawful SGI within stipulated timelines, and provide grievance redressal to users who flag such content.
According to a legal analysis by the international law firm Freshfields, an earlier October 2025 draft of the rules had proposed mandatory watermarking covering at least ten per cent of the surface area of a visual or the duration of an audio. That fixed requirement was dropped from the final text after consultations. The rules now require “clear and prominent” labelling without a fixed-size mandate, while strengthening the back-end metadata and detection obligations on platforms. The Mondaq legal commentary and the PTC News explainer published in February confirm the same shift.
The DPDP Act, India’s data protection law passed in 2023, runs in parallel but addresses a different problem — the processing of personal data. The 2026 IT Rules amendments cover the synthetic content layer specifically. Together, these two frameworks form the architecture that the Indian state has built, late but not absent, to address the realities of generative AI.
The incidents that forced the response
Regulation in India usually moves after the harm has been done. Deepfakes are no exception.
The Telangana 2023 example was the first widely cited political deepfake in an Indian election context, but it was not the most consequential. The 2024 Lok Sabha campaign produced a cluster of incidents that, taken together, made the case for a legal response unavoidable. Manipulated videos involving Union Home Minister Amit Shah and Uttar Pradesh Chief Minister Yogi Adityanath circulated on social media and triggered police investigations across multiple states, as Reuters reported in May 2024. Fake clips of Bollywood actors Aamir Khan and Ranveer Singh appearing to criticise the Prime Minister and call on viewers to vote for the Congress crossed 5,00,000 views before takedowns began, according to Reuters reporting cited by ACM TechNews on 22 April 2024. The actors themselves filed police complaints. In Uttar Pradesh, the state police arrested a man named Shyam Gupta for sharing a fake video of Yogi Adityanath on X — using internet protocol address tracing, according to Reuters — and charged him under provisions that carry up to a seven-year prison term.
The Prime Minister addressed the issue directly. At an election rally in Satara on 29 April 2024, he warned voters that opponents who could not engage politically were “spreading fake videos on social media” by “misusing AI to mimic my voice or that of my colleagues”, urging people to report such content to the police. A few weeks later, on 7 May 2024, the Election Commission of India issued its first formal guidelines on the responsible use of social media during elections, including a directive that political parties take down deepfake audios and videos within three hours of notification, identify the responsible person within their party, and warn them. The Business and Human Rights Centre archive carries the text of those guidelines.
The Commission returned to the subject on 16 January 2025, with an advisory titled Labelling Synthetic or AI-Generated Content Used by Political Parties for Election Campaigning. As the Deccan Herald reported, the directive followed the registration of an FIR by Delhi Police against the Aam Aadmi Party for posting AI-generated videos of Prime Minister Modi and Home Minister Shah on the party’s X handle. The 16 January advisory required all parties to label any image, video or audio “generated or significantly altered by AI technologies” with notations such as “AI-Generated”, “Digitally Enhanced” or “Synthetic Content”.
By the Bihar Assembly elections preparation in October 2025, MediaNama was reporting that the ECI had effectively banned unlabelled deepfakes in political campaigns. The Commission warned of strict monitoring and penal action for Model Code of Conduct violations. The DMK had used AI in September 2024 to recreate the voice and image of party founder M. Karunanidhi at a public event. The Maharashtra Assembly campaign had produced a manipulated audio of NCP (SP) MP Supriya Sule. The list, by then, was long enough that a serious framework had become a political necessity rather than a technocratic ambition.
Why India is unusually exposed
India is not the only democracy facing this problem. It is, however, structurally more vulnerable than most. The country has the world’s largest WhatsApp user base, with figures regularly quoted in the range of five hundred million plus. Smartphone penetration has spread well beyond English-speaking urban segments into regional language internet users who consume video and audio rather than text. Fact-checking infrastructure, while improving — through organisations such as Alt News, BOOM Live, Reuters Fact Check and the official PIB Fact Check — covers only a fraction of the volume of content circulating in Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada and the dozen other languages in which Indian elections are actually fought.
Polarisation amplifies the problem. Voters in WhatsApp groups organised around political, regional or religious identity are predisposed to believe content that confirms their priors and dismiss content that contradicts them. Researchers call this the liars’ dividend — even authentic footage can be denied as a deepfake, because the public no longer has a stable way of knowing what is real. Kritika Goel, the head of editorial operations at Logically Facts India, told a MediaNama event that the most dangerous effect of deepfakes in the 2024 election was less the direct deception and more the erosion of trust in any video at all. That is the deeper harm. A democracy can survive disagreement. It struggles when its citizens lose the ability to share, even minimally, a picture of what just happened.
What the platforms have been doing
Platforms have responded with a mix of self-regulation, compliance with Indian directives, and quiet pushback. Meta’s own transparency disclosures, cited in multiple Indian media reports, indicate that the company restricted more than 28,000 pieces of content in India during the first six months of 2025 in response to government requests under existing IT Rules. YouTube has rolled out AI content disclosure features for creators. X has updated its synthetic and manipulated media policy. WhatsApp has expanded its “forwarded many times” labels and limited mass forwarding.
These steps are real but uneven. Detection technology lags generation technology by structural design — by the time a detector is trained on one model’s output, three new models have been released. Watermarking standards are inconsistent across platforms and across geographies. Open-source generative models can be modified to strip identifying metadata. The asymmetry between cost of creation and cost of detection is widening, not narrowing.
The new IT Rules try to address this by placing the legal burden on platforms — the entities with the resources and the infrastructure — rather than expecting individual users to detect deepfakes on their own. Whether that burden translates into actual behaviour change in the short timeframes the rules prescribe will depend on enforcement, on judicial interpretation of “reasonable” technical measures, and on whether platforms cooperate or contest.
The civil liberties question
A legitimate concern, raised by digital rights groups including the Internet Freedom Foundation, is that the same regulatory architecture designed to protect voters from deceptive synthetic content can also be used to chill legitimate satire, political commentary, parody and journalism. The line between a deepfake intended to deceive and a clearly labelled parody intended to comment is not always crisp. If platforms err on the side of takedown to avoid liability — which the financial logic of compliance suggests they will — the result could be over-removal of legitimate political speech, particularly speech that is critical of those in power.
The 2026 Rules attempt to thread this needle by focusing on deceptive intent and unlabelled SGI rather than all synthetic content. In practice, the line will be drawn by individual takedown decisions, grievance officers, government takedown orders under Section 69A of the IT Act, and eventually by courts. The Supreme Court’s 2017 Justice K.S. Puttaswamy judgment on privacy as a fundamental right and the 2015 Shreya Singhal judgment striking down Section 66A of the IT Act both indicate that Indian constitutional law treats restrictions on online speech with significant care. The political will to use the new framework with restraint, regardless of which party is in office at a given moment, will determine whether the rules end up protecting trust or eroding it from a different direction.
What this means for ordinary Indian readers
For most readers, the relevant question is not the technical detail of the rules but the practical one: how do I know what I am looking at?
The basic discipline is unchanged from the broader fight against misinformation. Pause before forwarding. Notice whether content has an “AI-Generated” or “Digitally Enhanced” label as the new rules require. Reverse-search striking images and screenshots. Cross-check viral political claims against verifiable sources — the official Election Commission website, PIB Fact Check, established fact-checking organisations, and reputable news outlets that name reporters and editors. Be especially careful with content that arrives emotionally pre-loaded, particularly in the final days before an election. The most consequential deepfakes are designed to land in the narrow window between damage and correction.
The deeper civic discipline is harder. It involves accepting that not every plausible-looking video is true, even when it confirms what one already believes. It involves treating fact-checks from credible sources as worth reading even when they contradict one’s political position. It involves an active resistance to the liars’ dividend — refusing to dismiss inconvenient authentic footage as “probably fake” simply because that is now a culturally available move.
What most reports are missing
Most coverage of India’s deepfake rules has focused on what the rules say. The under-reported angle is what they cannot do.
First, the largest social platform in India is not regulated as a publisher of content. WhatsApp is end-to-end encrypted and treated as a messaging service. Most political misinformation in India travels through WhatsApp groups before it ever surfaces on X, Instagram or YouTube. The new IT Rules cover SGI on intermediary platforms; they do not — and constitutionally cannot, without breaking encryption — reach into closed groups where the actual viral spread happens. This is the structural limit of the framework.
Second, enforcement against creators remains slow and uneven. The 2024 case of Shyam Gupta in Uttar Pradesh was an exception more than a rule. Most political deepfakes circulate without anyone being identified, let alone prosecuted, particularly when the source is anonymous or based outside Indian jurisdiction.
Third, the rules address what is on the screen. They do not address the political incentive structure that makes deepfakes profitable to create in the first place. As long as a fake video that lands in the final 48 hours of an election can shift a few percentage points of vote share, the temptation to commission one will exceed whatever penalty an after-the-fact prosecution carries.
What happens next
The first real test of the 2026 IT Rules will come in the next major electoral cycle — the state assembly elections scheduled through 2026 and 2027. Compliance reports from platforms, takedown statistics from MeitY, and the proportion of viral political content that arrives with proper labels will tell us more than any government press release. The Bihar elections, scheduled to follow the existing ECI labelling advisory of October 2025, will be the early indicator.
In parallel, the global regulatory environment is moving. China’s mandatory synthetic content labelling rules came into force on 1 September 2025. The European Union’s AI Act categorises real-time biometric identification and certain synthetic media uses as high-risk. The United States is moving more slowly at the federal level but faster at the state level. India’s rules will be benchmarked, in court and in commentary, against these alternatives.
The deepest test, however, will not be regulatory. It will be cultural. The biggest danger of deepfakes is not that people will believe every fake video. It is that people will stop believing any video, including the authentic ones. Once shared visual evidence loses its hold on a democratic public, every accountability mechanism that depends on evidence — journalism, courts, the Election Commission, the police — weakens. Restoring even a partial shared reality is the work of the next decade. The IT Rules of 2026 are the legal scaffolding. The political and civic work has to be done by everyone else.
Conclusion
India has done something this year that it did not do in 2014, 2019 or 2024. It has written a law that names the problem. The law is imperfect. It will be tested in court. It will be challenged for over-reach by some and for under-reach by others. It will fail to catch some of the worst content and will probably catch some content that did not deserve to be caught.
That is the nature of regulation arriving after a technology has already reshaped the terrain. The honest question is not whether the IT Amendment Rules of 2026 are perfect — they are not — but whether they are better than nothing. On the evidence of the past three years, of the Telangana clip and the Aamir Khan video and the Supriya Sule audio, the answer is yes. The next question, harder and more important, is whether the Indian state, the platforms and the political parties can build a culture of restraint around the new framework rather than treating it as a tool to weaponise. The country’s record on that question is, at best, mixed. The record on the question of what citizens themselves do with a video that feels too good to be true is the one we have most control over, and the one we still talk about least.