MeitY Advisory on Artificial Intelligence Models (1 March 2024, replaced 15 March 2024)

Before 2024 the Ministry had used the due-diligence clause of the IT Rules, 2021 against AI-generated content in the form of advisories. In November 2023 it advised significant social media intermediaries to identify misinformation and deepfakes and to remove reported deepfakes within 36 hours, warning that failure would invite Rule 7 and loss of Section 79(1) protection; in December it advised all intermediaries to communicate the prohibited categories of Rule 3(1)(b) to users at registration, at login and at each upload.⁴⁵ Both are described in a separate entry.

The first brush with generative AI chatbots came in November 2023, when Minister of State for Electronics and IT Rajeev Chandrasekhar said that Google's Bard, which had labelled certain news and social accounts as biased, could not claim safe harbour because bias would breach the safety and trust obligations in Rule 3(1)(b); MediaNama noted that Rule 3(1)(b), which addresses content users share, might not apply to content an AI itself generates.⁶ In late February 2024 Chandrasekhar went further and alleged that Google's Gemini had violated the IT Rules by saying that some experts regarded Prime Minister Narendra Modi's policies as 'fascist'.³¹² The advisory of 1 March 2024 followed within days.¹³

According to MediaNama, which reviewed the document, the advisory was sent by MeitY to AI companies and intermediaries on 1 March 2024. Its main directions were:¹

• Government permission. Any under-tested or unreliable AI model could be made available to Indian users only after receiving explicit permission from the government.
• Labelling and consent. Such models had to be 'appropriately labeled' to show that their output may be unreliable, and had to display consent pop-ups telling users that the output is unreliable.
• Deepfakes. If a model facilitates deepfakes, the content had to be labelled or embedded with permanent unique metadata or an identifier that would show it was generated by an AI model and identify the user, the model and the creator or first originator of the deepfake.
• Rule 3(1)(b). Platforms had to ensure AI use on their service did not let users host, display, upload, modify, publish, transmit, store, update or share content prohibited by Rule 3(1)(b).
• Bias and elections. Platforms had to ensure AI use did not permit bias or discrimination or 'threaten the integrity of the electoral process'.
• User information. Terms of service and user agreements had to inform users of the consequences of dealing with unlawful information, including removal of content and suspension or termination of accounts, and of possible punishment under law.
• Report. Intermediaries had to submit a status report on action taken within 15 days.

The advisory contained no distinction by company size, no definition of 'under-tested' or 'unreliable', and no procedure for seeking permission.¹²

The advisory spread quickly through the technology community. On 4 March Chandrasekhar posted clarifications on X: the advisory was aimed at 'significant platforms' and untested AI platforms, permission-seeking would apply only to large platforms and not to start-ups, and platforms already faced 'legal consequences under existing laws (both criminal n tech laws)' for enabling or directly outputting unlawful content. He described labelling and explicit consent as the best protection for AI companies and called the advisory an 'insurance policy' against being sued by customers.¹ MediaNama pointed out that the advisory itself drew no such size distinction and that the clarifications had not been reflected in a revised text.¹²

Critics raised five main objections.¹²

• Legal force. MediaNama's founder Nikhil Pahwa asked how an advisory could be legally binding, 'not an ordinance', and how it was backed by the IT Act, and called for public consultation. MediaNama's analysis said that, even if lawful, the advisory would not be legally binding.
• Scope. The IT Rules do not expressly treat AI models as intermediaries or publishers and do not define a 'platform'. Whether an AI company is an intermediary within Section 2(1)(w), and so covered by Rule 3(1)(b), was unclear.
• Liability. AI outputs are probabilistic and depend on training data, model weights and prompts; commentators questioned whether a chatbot's owner could be treated as the publisher of its output, and whether the Shreya Singhal requirement of actual knowledge sits with that approach.
• Vagueness. The government gave no test for 'under-tested' models, and commentators asked how it would assess the hundreds of thousands of open-source models listed on Hugging Face.
• Competition and innovation. AI start-up founders told the Economic Times that the directions would hurt start-ups and favour large corporations with the resources to test and obtain approval. MediaNama observed that only well-funded companies could release models if beta testing needed permission.

The revised text followed two weeks later.

On 15 March 2024 MeitY sent a revised advisory, reviewed by MediaNama, that replaced the earlier text and 'reverses the Ministry's earlier stance' that required explicit permission. Under-tested or unreliable models could now be offered to Indian users if 'appropriately labeling the possible inherent fallibility or unreliability of the output generated', and models had to carry consent pop-ups or similar mechanisms.³

MediaNama's reading was that the permission requirement had been withdrawn in the face of start-up concern but that the advisory's other questions remained unanswered: what 'undertested' or 'unreliable' means, why AI companies should answer for their models' output, and the legality of the advisory itself.³

The advisory is a landmark in Indian AI governance for three reasons. First, the Union sought, however briefly, to condition the availability of AI models in India on government approval, and the speed of its withdrawal showed the limits of governing by advisory under pressure from industry. Second, it set the template for the government's later approach: reading an AI service as an intermediary subject to Rule 3(1)(b), requiring labels and persistent identifiers on synthetic content, and stating consequences in terms of liability under the IT Act and criminal law. Third, it made the legal-status problem explicit. The advisory was not delegated legislation, yet it was framed in mandatory language.¹²

The advisory's ideas were eventually moved onto firmer legal ground. The IT Rules amendment of 10 February 2026 requires intermediaries that offer synthetic-media tools to label synthetic content and embed permanent metadata with a unique identifier, and prohibits enabling the removal of either.¹¹¹⁰ Law-firm commentary notes that the labelling, identifier and automated-tool elements of the earlier advisories were codified as binding requirements from 20 February 2026, and lists the advisories of 26 December 2023 and 15 March 2024 among the instruments that preceded the amendment.⁸⁹ The codification was selective: the 2026 rules address the provenance of realistic synthetic audio-visual output, and do not carry over the 2024 advisory's permission requirement, its labelling of models as unreliable or its consent pop-ups, which concerned the testing and reliability of the models themselves.¹¹

The 2024 episode is also cited in current debates. In April 2026 an Ikigai Law lawyer speaking at a MediaNama discussion on draft Rule 3(4), which would make intermediaries comply with MeitY advisories and other instruments or risk safe harbour, pointed to the March 2024 advisory, which he said the government withdrew after industry pushback, as the reason advisories should remain non-binding.⁷

The 1 March 2024 advisory was replaced by the 15 March 2024 advisory. The sources reviewed do not report a formal withdrawal of the second advisory, but its core duties on labelling and provenance for synthetic audio-visual content have since been made binding on intermediaries by the February 2026 amendment to the IT Rules.⁸¹¹ The permission-based approach to AI models has not been revived. In October 2026 Union IT Minister Ashwini Vaishnaw said a consultation paper on AI regulation would follow within a month; the sources reviewed do not describe its contents.