Consider the following scenario. A brand manager at a well-known consumer goods company. One morning, a colleague sends him a link. It’s a sixty-second video with slick production values, in which the company’s logo sits prominently on a product that the company has never manufactured, their trademarked jingle plays in the background, and a synthetically generated voice that is indistinguishable from that of the contracted celebrity ambassador recommending a rival’s product. The video has 800,000 views within 12 hours of upload.
They seek advice from the legal counsel who turns to the recently amended IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 (hereinafter ‘IT Rules’) as amended on 10 February 2026, specifically to deal with AI-generated synthetic content and the Trade Marks Act, 1999 (hereinafter ‘TMA 1999’). Upon closer examination, however, neither regime furnishes a clear or reliable remedy. The loophole at the intersection of these two laws is quietly becoming one of the most commercially significant unaddressed risks in Indian Intellectual Property law.
The central gap is that the IT Rules’ newly introduced definition of ‘synthetically generated information’ is confined, as a threshold matter, to content depicting an identifiable individual or a real event. AI-generated content that replicates registered trademarks, brand trade dress, or sound marks without depicting any recognisable natural person falls outside this regulatory framework. Simultaneously, the TMA 1999’s infringement provision under Section 29, which requires ‘use in the course of trade’ by an identifiable party, remains wholly untested against AI-generated infringement by Indian courts. The result is a regulatory lacuna of substantial and growing commercial consequence for brand owners operating in the Indian digital economy.
This piece examines the scope and structural limitations of the 2026 IT Rules amendment as applied to AI-generated brand impersonation. It further analyses the ‘use in the course of trade’ requirement under Section 29 of the TMA 1999 and the unresolved question of attribution in AI-mediated infringement. It then considers the viability of passing off and how the Section 79 intermediary safe harbour interacts with it. Finally, it maps three distinct categories of AI brand impersonation against the applicable legal framework and advances specific legislative and regulatory recommendations.
I. What the 2026 Amendment in IT Rules Actually Says
The February 2026 amendment to the IT Rules introduces a formal statutory definition of synthetically generated information (SGI) and imposes obligations on platforms that enable its creation. These obligations include mandatory content labelling, permanent metadata, embedding with unique identifiers, proactive monitoring, and a tightened takedown window reduced from 36 hours to just three hours for unlawful content.
However, the definition itself in the new Rule 2(wa) contains a critical limitation that defines SGI as content that “depicts or portrays any individual or event” in a manner indistinguishable from reality. The phrase “individual or event” is where the problem begins, reflecting the legislature’s evident concern with protecting natural persons from non-consensual deepfake content, an entirely commendable legislative objective. However, this definitional choice has an unintended consequence of considerable commercial significance: A deepfake advertisement that synthetically replicates a registered logo, packaging design or trademark sound without depicting any identifiable human being doesn’t meet the statutory threshold of the Rule 2(wa) definition of SGI. It falls outside the framework entirely before considering any of the provided exemptions. So, an AI-Generated video that faithfully replicates your brand’s visual identity but doesn’t depict any identifiable natural person is not considered within the IT rules. An AI-generated audio clip that reproduces trademarked audio with high acoustic similarity is also excluded from the SGI framework on identical grounds.
Even where the Rule 2(wa) definition does apply, for instance, because your celebrity ambassador’s face appears in the content, the proviso to that definition introduces three carve-outs that can exempt content from the SGI framework entirely. The most dangerous of these is Proviso (b).
It reads, in relevant part, as exempting content arising from “routine or good-faith creation, preparation, formatting, presentation or design of documents, presentations… educational or training materials, research outputs, including the use of illustrative, hypothetical, draft, template-based or conceptual content, where such creation or presentation does not result in the creation or generation of any false document or false electronic record.”
The operative terms Illustrative, hypothetical, Research Output, Conceptual content are not defined anywhere in the rules. They are broad, indeterminate, and susceptible to expansive self-serving interpretation by a platform motivated to avoid the obligations that SGI classification would trigger. More fundamentally, the determination of whether particular content falls within the Proviso is, under the Rules as currently drafted, a matter of platform self-assessment rather than independent regulatory or judicial determination.
“Here’s a video of your spokesperson recommending our product. We made it as a hypothetical illustrative example for our content moderation training programme.”
Now this characterisation will exempt this creation under Proviso (b). Generative AI companies routinely describe their outputs as demonstrations, research tools or educational materials. And there is no mechanism in the current Rules for a brand owner or anyone else to challenge the platform’s self-characterisation before a regulator with jurisdiction over the matter. The Grievance Appellate Committee under Rule 3A was designed to hear appeals against the Grievance Officer’s decisions. It was not designed to adjudicate definitional disputes about whether content is SGI in the first place.
The consequence is significant. Content that qualifies for the proviso doesn’t just escape penalty liability; it escapes the entire regulation: No labelling obligation arises, no metadata is required, no monitoring duty applies, no grievance timeline is triggered. The platform is, by the Rules’ own terms, under no obligation to take any step in relation to such content.
II. The Trade Marks Act, 1999: ‘Use in Course of Trade’ and Distributed Attribution in AI-Generated Infringement
It might be assumed that, notwithstanding the limitations of the IT Rules, the TMA 1999 provides adequate protection for brand owners whose registered marks have been synthetically reproduced. Though rights exist in the registration of logo, sound and ambassador’s contract, including image rights, however, the difficulty does not lie in the existence of these rights but in the absence of a tested mechanism for their enforcement against AI-mediated infringers.
Section 29 of the TMA requires that the infringing party “uses in the course of trade” a mark identical with or deceptively similar to a registered mark in relation to the same or similar goods or services. Judicial construction of this requirement has engaged with domain name appropriation, metatag embedding, and keyword advertising. It has not, however, been authoritatively applied to AI-generated brand impersonation, and the existing doctrinal framework requires careful extension to address this category of harm.
The attribution difficulty is structural. When a user prompts an AI platform to generate a synthetic advertisement incorporating your brand, three parties are involved: the user who types the prompt, the platform whose model generates the output, and the social media service where it gets posted. The question which now arises is, which of these entities is “using the mark in the course of trade”?
The user may satisfy the requirement where the synthetic content is commercially circulated for the user’s own commercial benefit. But identifying individual users is frequently uncertain and difficult and most of the users cannot satisfy any award of damages or costs after suing them. [5] [pj6] The practical utility of this head of liability is constrained by the difficulty of identifying individual users and the limited financial capacity of most such persons to satisfy any meaningful remedy. The [7] [pj8] AI platform? It presents a more analytically significant question. The platform’s generative model processes your mark as training data and produces an output, but does that constitute trademark use in a legally operative sense? In Satyam Infoway Ltd. v. Sifynet Solutions Pvt. Ltd., 2004 , the Supreme Court extended trademark principles to domain names, meaning domain name registration and use could constitute trademark use, reasoning that the internet had become a commercial marketplace in which trademark principles operated with full force.
The Delhi High Court extended similar reasoning to metatag embedding in Tata Sons Ltd. v. Manu Kosuri, 2001. In Google India Pvt. Ltd. v. Visakha Industries, 2020, the question of keyword advertising and trademark use was engaged, albeit without definitive resolution. Each of these extensions involved a human actor making a deliberate commercial decision to employ a mark for the purpose of trade. The generative AI context is structurally different in a manner that none of these authorities directly addresses.
Further, Indian Courts have employed John Doe or Ashok Kumar orders to grant interim injunctive relief against anonymous or unknown defendants in intellectual property matters, including digital infringement cases. In cases where the individual user responsible for prompting and disseminating the infringing content cannot be identified at the point of filing, such orders may offer brand owners a degree of interim relief. A John Doe order presupposes that the unknown defendant is a natural person whose identity is capable of being subsequently ascertained and against whom the injunction can be meaningfully enforced upon identification. However, in the AI brand impersonation context, this presupposition fails. The most significant actor in the infringement claim is not the anonymous user, but the AI generative platform whose identity is known, whose role in introducing the infringement output is direct, and against whom the brand owner could, in principle, proceed directly.
Such an order secures nothing more than interim relief against the party who is anonymous and does not resolve the underlying question of the AI platform’s liability. They only secure the removal of one piece of infringing content, while doing nothing to prevent the same platform from generating functionally identical content the following day. Further, since any user of a freely accessible generative AI platform can produce infringing content at negligible cost within seconds, the class of potential wrongdoers is effectively unbounded, and Courts may be reluctant to grant such orders against such an indefinite class.
When a generative AI platform produces synthetic brand-impersonating content in response to a user prompt, no simple party in the chain performs an act that corresponds cleanly to conventional trademark ‘use’. The user inputs a prompt but may have no direct commercial interest in the output and may not ‘use’ the mark in any commercially operative sense. The AI platform processes the prompt and generates the output, but the platform’s model doesn’t ‘use’ the mark in the manner of a trader who affixes it to goods or deploys it in advertising; rather, it reproduces a statistical approximation of visual or sonic patterns derived from training data, without any volitional act directed at the mark as a commercial identifier. The social media intermediary hosts the output but exercises no editorial judgement over its content. The result is that the act most casually proximate to the infringement, i.e., the generation of the synthetic content by the AI model, is performed by an entity whose relation to the concept of ‘use in the course of trade’ is, at best, analogically attenuated.
This difficulty has direct consequences for the remedies available to a brand owner. Section 29 does not impose strict liability; it requires a defendant who uses the mark. If the AI platform cannot be characterised as a user of the mark in the legally operative sense, it cannot be an infringer under Section 29, regardless of the commercial harm caused by its output. The brand owner is then left with the user, who is anonymous, impecunious, or both, and the social media intermediary, which benefits from the Section 79, which provides for safe harbour. The infringement goes without a defendant capable of satisfying a remedy in practical terms.
The comparative experience of other jurisdictions offers limited but instructive guidance. In the United States, courts applying the Lanham Act have begun to deal with analogous questions in the context of AI-generated content, and the ‘use in commerce’ requirement may need to be reconceptualised to account for automated generation. In the European Union, the trade mark ‘use’ doctrine developed under Arsenal Football Club v. Reed, [2002] and L’Oréal SA v. Bellure NV, [2009] emphasises the functional impairment of the mark’s guarantee of origin as the operative criterion, an approach that may be more readily adaptable to AI generated impersonation because it focuses on the effect of the use rather than the identity or volition of the user. Indian courts could draw on both traditions, and there is no doctrinal impediment to the adoption of a similarly effect-oriented construction of Section 29 in an appropriate case.
The AI platform does not ‘use’ the mark in the manner contemplated by a conventional Section 29 analysis; its output is the probabilistic product of a generative model rather than the result of a deliberate commercial act directed at the mark. In a conventional counterfeiting case, there is a single identifiable defendant who produced the counterfeit goods. In an AI brand impersonation case, the “misrepresentation” is a distributed act across user, model and platform. The Trade Marks Act was not written for distributed authorship.
(This post has been authored by Priyanshi Jain, 4th Year student at National Law University, Jabalpur)
CITE AS: Priyanshi Jain, ‘AI Generated Brand Impersonation and the Limits of ‘Use in Course of Trade’ under the Trade Marks Act, 1999’ (The Contemporary Law Forum, 06 July 2026) <https://tclf.in/2026/07/06/ai-generated-brand-impersonation-and-the-limits-of-use-in-course-of-trade-under-the-trade-marks-act-1999-part-i/ > date of access.