The rapid proliferation of synthetically generated information has forced governments and regulators across the world to confront and assess governance and accountability frameworks for AI systems and outputs. As the boundaries between real and synthetic media get more blurred, policy measures must address competing demands of accountability, innovation, privacy and public trust.
India’s recent amendments to the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 was a highly proactive regulatory move. By mandating the labelling of synthetically generated information, the regulatory framework seeks to create transparency and reduce risks posed by synthetic media. At first glance, the logic appears straightforward: If users are informed about the nature of the content they encounter, they can make more informed decisions. Yet, beneath this straightforward regulatory objective lies a complex web of ambiguities and practical constraints that complicate implementation.
One of the central difficulties arises from the scope of synthetic content. While the regulatory framework attempts to remain technology-neutral, it struggles to clearly delineate what constitutes routine or good-faith editing. The definition of ‘synthetically generated information’ sets out a materiality threshold but does not clarify when content will become significantly modified to trigger SGI classification. No standards are prescribed for permanence, and no factors have been set out for determining ‘technical feasibility’.
This burden is particularly evident in the obligations imposed on significant social media intermediaries (SSMIs), which are expected to detect, label and regulate synthetic content at scale. The expectation that SSMIs can reliably and accurately identify such content assumes a level of technological capability that does not yet exist. Detection systems, while improving, remain probabilistic, offering likelihood rather than certainty. In an environment where the cost of non-compliance may include the loss of legal protections, SSMIs are incentivised to err on the side of caution. The result will be a tendency towards over-labelling, where even marginally altered content may be marked as synthetic, diluting the significance of the label itself.
{{/usCountry}}This burden is particularly evident in the obligations imposed on significant social media intermediaries (SSMIs), which are expected to detect, label and regulate synthetic content at scale. The expectation that SSMIs can reliably and accurately identify such content assumes a level of technological capability that does not yet exist. Detection systems, while improving, remain probabilistic, offering likelihood rather than certainty. In an environment where the cost of non-compliance may include the loss of legal protections, SSMIs are incentivised to err on the side of caution. The result will be a tendency towards over-labelling, where even marginally altered content may be marked as synthetic, diluting the significance of the label itself.
{{/usCountry}}The amendments raise questions of interoperability and tampering. Existing tools may not be able to verify SGI created using competitors’ tools. Where SGI is used as an input in an AI system, the output may not preserve the provenance chain. ‘Overt’ watermarks can be tampered with easily, and C2PA metadata can be removed by taking screenshots and stripping EXIF data.
There are also concerns about effectiveness of SGI labelling and whether it will meet the regulatory objectives that it is purported to achieve. In a digital ecosystem already saturated with information, label fatigue is likely and such disclosures risk becoming background noise. Users, accustomed to navigating constant streams of content, may simply learn to ignore such labels, rendering them functionally redundant. Worse still, the absence of a label may be misinterpreted as a guarantee of authenticity, even when content has escaped detection. In this way, a mechanism designed to enhance trust may inadvertently contribute to its erosion.
Beyond questions of efficacy, there are also significant tensions between transparency and privacy. The requirement to embed metadata or other provenance markers introduces the possibility of increased data collection, which sits uneasily alongside the principles of data minimisation enshrined in the Digital Personal Data Protection Act, 2023. This creates a regulatory paradox where efforts to make content more traceable may compromise user privacy.
The challenges extend further when considered in a global context. Synthetic content does not recognise national boundaries, and its circulation across jurisdictions exposes the limitations of fragmented regulatory approaches. Different countries have adopted or are looking to adopt divergent strategies, from China’s layered labelling systems to harm-based models in parts of the US. In the absence of common regulatory standards, content may lose, duplicate or distort its labelling as it moves across platforms and regions, undermining the very objective of consistency and clarity.
Technological systems are evolving at a rapid pace and attempts to impose rigid frameworks may be ill-suited for fluid and rapidly changing technologies. Not to diminish the case for regulatory oversight, but regulatory and policy measures must be flexible, adaptable and subject to collaborative engagement between regulators and regulated entities. Governance must acknowledge the limits of regulatory and legal control, as no system of rules can fully eliminate uncertainty. Developing voluntary industry standards and requiring explicit user-level duties are perhaps options that are better suited as novel, initial measures than prescriptive mandates.
(The views expressed are personal)
This article is authored by Lagna Panda, partner, AP & Partners.