[Aviral Singhai is a fifth-year B.A. LL.B. (Hons.) student at the National Law Institute University, Bhopal. In this piece, the author interrogates the technological architecture behind AI-generated deepfakes and mirror websites and the challenges they pose for India’s legal framework. The piece argues that while existing laws provide some remedies against unlawful online content, they struggle to regulate AI-generated deepfakes that can be rapidly replicated and endlessly disseminated through mirror networks, exposing significant gaps in both the substantive and enforcement framework and necessitating targeted legal reforms.]
Introduction
In the last three years, Indian courts have repeatedly been asked to protect individuals from AI-driven misuse of their identities. In cases Anil Kapoor v. Simply Life India, Arijit Singh v. Codible Ventures LLP, and Sadhguru Jagadish Vasudev v. Igor Isakov, courts restrained the unauthorised use of various traits of public personalities through mirror websites. Similar concerns have recently surfaced in disputes involving Naga Chaitanya, Varun Dhawan, and Preity Zinta.
Generative AI has developed enough to recreate a person’s face, voice, expressions, and mannerisms from material already available online. The resulting content may look authentic despite having no connection with the individual concerned. Material like Deepfake videos, cloned voices, synthetic endorsements, and AI-generated chatbots can be produced within minutes and distributed across multiple platforms at almost no cost. Now while this may be addressed by the existing laws, the issue becomes difficult to address when they are replicated by way of mirror sites.
This article examines the challenges faced by laws in dealing with highly technical and complex problems caused by AI deepfakes of public personalities, which are distributed using mirror sites. It first explains how AI-generated content and mirror networks work and the technology involved in their creation. It then analyses how the existing framework, both substantive as well as enforcement, is insufficient for dealing with the dissemination of AI deepfakes using mirror sites. Finally, it considers possible reforms in the existing framework to better deal with these modern technologies.
How AI-Deepfakes their Mirror Networks work
In order to understand whether or not the existing laws are able and sufficient to deal with AI and mirror networks, it is important to understand how they work.
The creation of a deepfake begins with training a machine-learning model on large datasets containing photographs, videos, interviews, podcasts, and social media content. During training, the system converts identifiable features into numerical representations known as embeddings. The model then learns patterns that define a person’s facial features, voice, expressions, and mannerisms. Technologies such as Generative Adversarial Networks (GANs) and Diffusion Models use these patterns to generate entirely new content. They rely on techniques such as facial landmark detection and frame synthesis to create videos that appear real despite depicting events that never happened. Voice-cloning systems work in a similar manner. They analyse pitch, timbre, cadence, and pronunciation, they can generate fresh audio from a text prompt while sounding like the original speaker. Modern systems such as VALL-E can recreate a person’s voice using only a short audio sample.
The more difficult problem is that these digital files can be copied endlessly without any loss in quality. According to UN Women, a deepfake uploaded once can be downloaded, edited, and re-uploaded across multiple platforms within minutes. Removing the original upload therefore does not necessarily remove the content itself. This process of copying is usually done via a mirroring site. A mirror site is a replica of an existing website, including its databases, media files, directory structure, and search functions. Through synchronisation tools such as GNU Wget and rsync, administrators can automatically copy and update content across multiple servers. Mirroring was developed for legitimate purposes such as load balancing, backup, disaster recovery, and maintaining access when the primary server fails.
However, the same architecture can also be used for unlawful purposes. As a result, removing a single webpage or domain may have little effect because identical copies may already exist across multiple servers, hosting providers, and jurisdictions. This makes it significantly harder to eliminate once it is online.
Why Existing Laws Struggle with this
Existing laws struggle in dealing with dissemination of AI-Deepfakes through mirror sites. This can be seen by analysing the regulatory and enforcement framework.
Regulatory Framework
Under Indian law injunctive remedy against mirror sites can be issued only when the underlying content of the mirror sites is illegal. This means that unless AI-Deepfakes are classified as clearly illegal under a law, there would be no remedy against their mirrors.
This issue can be seen very clearly in the Copyright Act. Sections 13 and 14 of the Copyright Act, 1957 protect original literary, dramatic, musical and artistic works, cinematograph films, sound recordings, and performances. Where infringing copyrighted content is disseminated through mirror websites, Indian courts have addressed the problem by granting dynamic injunctions, as recognised in UTV Software Communication Ltd. v. 1337X.to. However, this remedy depends upon the existence of copyright infringement. AI-generated deepfakes often replicate a person’s voice, likeness and persona without reproducing a protected copyrighted work. Voice-cloning models can generate fresh audio, while deepfake systems can create synthetic videos that imitate an individual’s identity without copying any particular recording. While, it is true that sometimes the deepfake may fall within the purview of this act but often times this is not the case. Consequently, where AI-deepfakes falls outside the scope of copyright, the Copyright Act cannot be invoked to obtain copyright-based dynamic injunctions by relying on the UTV case.
Trademark law faces a similar limitation. While dynamic injunctions have been specified for trademarks as well as per the case of UltraTech Cement Ltd. v. www.ultratechcements.com. However, again the issue arises in specifying whether deepfakes are violative of trademarks act. Under Sections 2(1)(zb) and 29 of the Trade Marks Act, 1999, liability largely depends on commercial use and confusion regarding source, sponsorship, or association. Many AI-generated harms have nothing to do with commercial transactions. A fabricated interview, manipulated political speech, or cloned voice may cause significant reputational harm without advertising a product or misleading consumers about origin.
The Digital Personal Data Protection Act, 2023 also addresses only part of the problem. Although Section 2(t) of the Digital Personal Data Protection Act, 2023 adopts a broad definition of “personal data” as “any data about an individual who is identifiable by or in relation to such data”, enabling photographs, videos, voice recordings and AI-generated deepfakes depicting identifiable individuals to fall within its scope, the protection afforded by the Act is substantially curtailed by Section 3(c)(ii)(A). This section expressly excludes from the Act’s application personal data that has been made or caused to be made publicly available by the Data Principal themselves. This limitation assumes particular significance in the case of public figures, who routinely make their photographs, videos, interviews, speeches and social media content publicly available to engage with their audience and maintain their public presence. Such publicly available material frequently serves as the primary dataset for training AI models and generating deepfakes.
Article 21 and the Limits of Judicial Protection
Realising the limitations of these laws, the courts have tried to move beyond using a scattered framework to protect public personalities against AI-Deepfakes. They have instead started to use personality rights for the same. They have used Article 21 of the Constitution of India to protect personality rights in cases such as Anil Kapoor v. Simply Life India & Ors., Arijit Singh v. Codible Ventures LLP, and Sadhguru Jagadish Vasudev v. Igor Isakov & Ors. because they provide a much wider and blanket protection. This is done along with Section 69A of the Information Technology Act, 2000, by providing a dynamic injunction which creates an order not just for removal of the Deepfake but also its mirror sites.
However, in the absence of a dedicated statutory framework for Personality Rights, this leads to a big issue. Personality rights are not clearly defined in any statute. Indian law does not explain what forms part of a person’s “persona” or “personality”, or what kind of unauthorised use amounts to a violation of these rights. As a result, courts have to decide these questions on a case-by-case basis. This creates uncertainty because there are no fixed legal standards to determine when the use of a person’s identity in a deepfake will amount to a violation of personality rights.
Although this mechanism does provide a much wider protection than the scattered laws used before. All of these still face one common limitation i.e., enforcement.
Limited enforcement mechanism
Even if the underlying content of mirror sites is classified as illegal under any of the above laws, the enforcement mechanism for removal of mirrors still faces issues.
Currently, courts remove mirror sites by granting temporary and dynamic injunctions under Order XXXIX Rules 1 and 2read with Section 151 of the Code of Civil Procedure, 1908, restraining the dissemination of infringing content and extending relief to identified mirror, redirect and alphanumeric websites. These orders are implemented through Section 69A of the Information Technology Act, 2000, read with the Information Technology (Procedure and Safeguards for Blocking for Access of Information by Public) Rules, 2009, which empowers the Central Government to direct intermediaries and internet service providers to block access to such websites. Similar dynamic injunctions have also been recognised in copyright and trademark disputes, enabling rights holders to seek extension of blocking orders to identified mirror websites without instituting fresh proceedings each time. However, this framework has its own issues:
- The existing framework is reactive rather than preventive. It comes into operation only after a deepfake has already been created and uploaded. Digital files can be downloaded, edited and re-uploaded across multiple platforms within minutes without any loss in quality, which means that by the time an injunction is granted and blocking directions are issued, the deepfake may already have been widely circulated, causing significant reputational and commercial harm.
- Even where a dynamic injunction is granted, effective enforcement remains difficult. Since mirror websites replicate the databases, media files, directory structure and search functions of the original website using synchronisation tools such as GNU Wget and rsync, content continues to remain accessible despite blocking orders. For instance, despite several celebrities obtaining injunctions against the online circulation of AI-generated deepfakes, such content continued to remain available through mirror platforms such as DesiFakes and Celebdeepfakes. Similarly, although pornographic websites have been blocked in India, websites such as XHamster and XVideos continue to remain accessible through their mirror domains. Likewise, despite repeated blocking orders against pirated anime websites, platforms such as AniSuge, Anikoto and Zoro have repeatedly resurfaced through mirror websites.
Concluding Beyond Case-by-Case Protection
Deepfakes, voice clones, and mirror websites expose gaps that traditional legal frameworks were never designed to address. While judicial intervention has provided some protection, the same simply WILL NOT be able to keep up. This is why a more structured response is necessary. The following measures may help address some of these concerns.
A. Expand Intermediary Obligations Under the IT Rules
A practical solution can be implemented through the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 framed under Section 79 of the Information Technology Act, 2000. Platforms should be required to create digital fingerprints of court-identified deepfakes, voice clones, and AI impersonations through image, video, and audio fingerprinting systems. Future uploads can then be matched against those signatures and blocked automatically, while similarity-detection tools identify modified versions of the same content.
A framework similar to the one adopted by Google can be used. Google has adopted this approach specifically for Child Sexual Abuse Material (CSAM), where digital hashing and AI tools is used to prevent the repeated circulation of unlawful content. This technology assigns images and videos a unique digital signature, a “hash” and then compares it against a database of known signatures. If the two match, the content is considered to be the same or closely similar. If this is made a requirement in the context of deepfakes, it can considerably reduce instances of such misuse.
B. Issue DPIIT Guidelines on Personality Rights and Digital Identity
The Department for Promotion of Industry and Internal Trade (DPIIT) could issue guidelines drawing from the Guernsey Image Rights Ordinance, 2012, particularly Sections 1 and 3, which define “personnage” and “personality” broadly to include a person’s name, image, likeness, appearance, voice, signature, gestures, mannerisms, and other attributes capable of identifying an individual. The guidelines could further clarify how these concepts apply to deepfakes, voice cloning, AI-generated avatars, and synthetic endorsements. This would provide courts with a more structured framework when determining the scope of a protected persona in AI-driven identity disputes.
C. Express Recognition of Dynamic Injunctions Under the IP Laws
Indian courts have developed dynamic injunctions through decisions such as UTV Software Communication Ltd. v. 1337X.to & Ors. and Sadhguru Jagadish Vasudev v. Igor Isakov & Ors.. However, the Code of Civil Procedure, 1908 or any other statute contain no express provision recognising such relief. Since this jurisprudence is concentrated largely within the Delhi High Court, courts across India may adopt different approaches regarding the scope and availability of dynamic injunctions. In Singapore, Sections 193DDA–193DDC of the Copyright Act In Singapore, Section 193DDA of the Copyright Act provided an express statutory framework for website-blocking orders against flagrantly infringing online locations. The Singapore High Court subsequently relied on this provision in Disney Enterprises, Inc. v. M1 Ltd to develop dynamic website-blocking orders extending to new domain names providing access to the same infringing content. A similar amendment to the IP laws could authorise Indian courts to extend injunctions to mirror websites, redirect domains, reverse proxies, successor websites, and substantially identical future uploads carrying the same unlawful content.