Can You Still Believe What You See? Rethinking Deepfake Regulation After Article 50 of the EU AI Act
[Khushi and Lakshay are students at Christ (Deemed to be University), Delhi NCR.]
A convincing fake is no longer hard to make. A voice note, a video call, a "leaked" clip of a politician or a relative in distress: each can now be generated in minutes, and each carries the same evidentiary weight, at first glance, as something real. The issue from a legal perspective is not merely the possibility of deception. The problem is rather that the legal distinctions made by the courts and by regulatory agencies between, for example, an authentic recording and a forgery, an authentic utterance and a fabrication, no longer hold. Deepfake implications include fraud in finance, sexual non-consent imagery, misinformation during elections, and a novel category of "liar's dividend" defences, whereby actual evidence can be discredited as a forgery based on the fact that doubt is cheap.
Two questions need to be kept separate. One is whether synthetic content can be identified as synthetic. The other is whether identifying it actually prevents the harm it causes. Most of the regulatory agencies in Brussels and Delhi have gone into the first question. This piece argues that labelling alone cannot solve the deepfake problem, because synthetic content can be stripped of its provenance well before it reaches the public that needs protecting.
India's Current Legal Position
Until recently, India lacked a specific law to address deepfakes, but the legislative scheme is still spread out among multiple pieces of legislation rather than consolidated within one piece of legislation. The provisions are present within the Information Technology Act 2000 (IT Act), which contains the relevant criminal and civil provisions. Section 66C criminalizes the act of identity theft, Section 66D criminalizes cheating by personation through the use of computer resources, while Section 66E deals with the infringement on privacy in terms of capturing and publishing images. None of the sections were intended for generative AI purposes.
The real regulatory response has come from delegation of powers. In February 2026, the Ministry of Electronics and Information Technology has notified the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026, which add a new statutory class called "synthetically generated information" (SGI) to the IT rules of 2021. It mandates that any intermediary providing tools for generation of content should give prominent labeling of SGI and, wherever possible, include permanent metadata or provenance in such a way that cannot be deleted by the user. The duties of prominent social media intermediaries are further heightened to get the user declaration about whether the uploaded content is fake, validate such declarations via reasonable technical means, and take action regarding flagged deepfakes in a compressed time frame, which ensures safe harbour status under Section 79 of the IT Act. The Digital Personal Data Protection Act 2023 is only marginally relevant at this point since it relates to the processing of personal data rather than the veracity of the content, except when a deepfake uses someone's biometric or voice samples without permission.
Meanwhile, Indian courts have developed their own corpus of jurisprudence on the right to personality without waiting much for the legislation to catch up. In Anil Kapoor v. Simply Life India and Others, the High Court of Delhi issued an omnibus injunction against unauthorised exploitation of the actor’s name, voice, and image through AI technology. However, this is only significant in its reactive nature, being litigative and applicable only to people who can afford a John Doe order.
Article 50 of the EU AI Act
A more codified approach to transparency obligations has been pursued in the EU. Transparency obligations under Article 50 of Regulation (EU) 2024/1689, also known as the AI Act, are applicable irrespective of whether a certain system is determined to be "high risk." The application of such obligations came into force on 2 August 2026.
Article 50 allocates duties to providers and deployers. Providers of technologies which produce artificial audio, images, videos, and text-based content are responsible for ensuring that these outputs have markers which allow for their detection as artificial manipulation or generation, which is the responsibility of the developer of the technology and not its users. On the other hand, deployers are required to undertake a disclosure duty, whereby anyone using the AI system to generate or manipulate image, audio or video content which qualifies as a deepfake is required to disclose the artificial manipulation of the content. One such responsibility, more particular, is the obligation of reporting artificial or altered texts that are placed in the public domain for informing about issues of public interest. The guidelines of the Commission regarding Article 50 are formulated in July 2026. What Article 50 does not regulate is the underlying act of producing deepfakes or liability based on harm done. This is a regime of transparency, not of harm prevention.
Why Labels May Not Be Enough
Labelling, watermarking, metadata, and provenance systems are often discussed as though they were interchangeable. They are not, and the differences matter legally.
Visible claims are claims that are made at the time of publishing, they may either be true or false, and there is no mechanism in place to ensure their continued presence once the file has moved off that platform. On the other hand, metadata attached to a file, as envisaged by the 2026 amendment in India, stays along with the file as long as it has not been removed, and removal is not a difficult task; taking a screenshot, re-encoding the file or simply removing the metadata will achieve the same in a matter of seconds. But C2PA is an industry voluntary standard implemented by a consortium of leading technology and media firms, and it does not have any legal authority of its own, nor do any pieces of content created using non-compliant technologies carry any kind of credential. Forensic detection technologies that seek to detect AI-created content without depending on the cooperative marking process at all suffer from similar problems of false negatives and false positives.
This legal problem has a systemic effect and is not accidental. Both Article 50 and India's SGI law assume that where the generation takes place is where a sustainable signal is attached and persists through any distribution. Malign entities are, by definition, least likely to meet a marking requirement at the stage of generation, while also being highly likely to remove any signal that exists prior to any further distribution. A law that only applies to entities that adhere to the marking law targets the wrong demographic.
Who Should Be Responsible?
Responsibility should not collapse into a single point in the chain. Model providers are best placed to build detectability into a system at the architectural level, machine-readable marking, robust to common transformations, is a design choice available to them and not really to anyone downstream. Those individuals who create misleading information purposefully are guilty of the act of deception because of which Article 50’s requirement of disclosure may be justified despite its inability to stop the malicious party from not disclosing anything. Platforms face a tougher balancing challenge, as intermediary liability principles, in both the EU and India, have generally not wanted platforms to be guarantors of user-provided content, and there are very real reasons not to abandon this practice entirely, the most significant being that of proportionality and the possibility of over-removal. The extremely short periods within which platforms are required to remove allegedly illegal content in India under the 2026 amendment move toward making platforms absolutely responsible for SGI, and this needs to be examined in much greater depth, considering the lack of any counter-notice procedure.
An effective assignment will link responsibility with control and knowledge and not with merely being close to the content. Providers have the duty of ensuring that safeguards are in place at the design level. Deployers have the duty of creating and knowingly distributing deceptive synthetic content. Platforms have the responsibility of doing something about knowledge of any existing deceptive synthetic content or implementing appropriate systems to detect the content.
India's Possible Way Forward
India must not adopt Article 50 lock, stock, and barrel. A purely transparency-and-labeling scheme, no matter how thoughtfully constructed, suffers from the same flaw of removability that has been highlighted in this piece: it targets the cooperative use case, but misses the dangerous one. Where India’s 2026 amendment is correct is in placing the burden of provenance at the time of creation, and not merely through platform labeling later. Where the amendment might fall short is if the only such remedy is taken up, with a narrow window for takedown.
The better alternative would have three components: first, disclosure and marking requirements very much like those included in the 2026 amendment; second, an indigenous yet interoperable provenance protocol which will either build upon or integrate C2PA, rather than rely upon a uniquely Indian model where verification is solely based on trusting a label; and third, platform obligations that adjust based on knowledge and realistic capability, rather than strict liability, along with a meaningful counter-notification mechanism for misclassified content. While this would not entirely resolve the stripping issue, it would at least acknowledge that provenance is only one piece of the puzzle.