Abstract
The proposed amendments to the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, aim to regulate synthetically generated information by requiring the labelling of AI-generated content and extending intermediary due diligence obligations. Although the amendments try to improve transparency and provide definitional clarity, they put an excessive and technologically unrealistic burden on intermediaries. The framework runs the risk of over-compliance and ineffective enforcement due to the simplicity of watermark removal and the unreliability of AI detection tools. This article makes the case for uniform detection standards, a risk-based labelling approach, and a reallocation of responsibility towards AI tool developers through comparison with the US and EU models.
Keywords: SGI, Labelling, EU AI Act, US AI Labelling Act
Introduction
In a move to regulate the synthetically generated information (“SGI”), the Ministry of Electronics and Information Technology (“MeiTy”) proposed draft amendments to the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 (“IT Rules”). This step marks a positive development with the amendment aimed at strengthening the due diligence requirement under Rules 3 and 4. The amendment adds a new definition clause, rule 2(wa), which defines SGI as any information generated, modified, or altered with the assistance of a computer source in an artificial or algorithmic manner. The major changes have been made by the inclusion of multiple subclauses in Rule 3 and Rule 4, which expand the due diligence requirement for a social media intermediary (“SMI”).
Firstly, under rule 3(b), a new provision has been added which clarifies that – if any intermediary disables access to any information, including a synthetically generated one, it would not make that platform lose its safe harbour protection under section 79. Secondly, a new sub – rule 3 is added after rule 3(2). It mandates platforms that provide tools for automated content generation to label any content created with these tools with a permanent unique identifier. Additionally, under rule 4, supplementary due diligence requirements have been added for significant social media intermediaries (“SSMIs”). These requirements mandate them to ask for a user declaration to identify any content that might be generated artificially. Moreover, SSMIs must use reasonable technical tools to verify this user declaration. Lastly, they must clearly label the synthetic content to indicate that the content is generated by an automated tool.
While these new additions mark a positive step in filling the lacuna in the regulation of artificially generated content, the amendment remains incomplete with many gaps. Due to the availability of multiple tools for watermark removal, applying Rule 3(3) becomes a challenging task. Moreover, the implementation of additional duties on SSMIs becomes significantly dependent on technological tools, which will be explained in further parts. This article critiques the incomplete amendments proposed under Rules 3 and 4, which shift the burden entirely to the intermediary. Lastly, it concludes with solutions to bridge the gap highlighted in the proposed amendment.
Flawed Labelling Mandate
The IT Rules aim to amend the existing provisions to bring SGI within the regulatory framework. MeiTy has added rule 3(2) which obliges an intermediary to label or embed SGI with a permanent unique metadata or identifier. The government has shifted its baton to the intermediary, and now it is the sole obligation of the intermediary to identify, label, and even embed SGI with metadata.
The government has not prescribed any procedure for this and has left it to the intermediary’s discretion. However, the important facet is not what the procedure is, but how it is to be done. It will be very cumbersome to securely embed each and every SGI with metadata on a platform like X, which has 22 million active users in India. In a collaborative study [CE1] between the University of California, Santa Barbara and Mellon, researchers discerned that watermarks were easily removable through various simulated attacks. The modus operandi lies in treating watermarks like an image; this can be done through tweaking things like the brightness, contrast or using JPEG compression, or even simply rotating an image. The other way is using a good old Gaussian blur.
The Artificial Intelligence (“AI”) detectors are unreliable as they function on the detection of patterns used by AI. It produces a high number of false positives and false negatives. Moreover, it is very easy to bypass these detectors through various mechanisms, such as the humanising tool. This creates a situation where harmful AI-generated content like deepfakes might slip through the scrutiny of the detectors and spread on the platform. This has added an extra burden that is difficult to follow in the current technological landscape. Moreover, due to the flaws in the AI detection system, genuine content may be flagged as unreliable, demonstrating the tainted truth effect that highlights that inaccurate warnings about misinformation reduce belief in accurate information. The Forbes analysis of the Biden AI companies shows that the payload of watermarks constitutes a very small percentage of the content it is associated with, usually a few kilobytes. Consequently, simple actions like cropping, resizing and re-encoding can easily erase or damage these identifiers. Therefore, Rule 3 rests on a fragile base that requires permanent water marking of SGIs.
Moreover, rule 4(1A) prescribes a procedure for SSMIs to label SGI, which starts from users declaring that what they are posting is SGI. Then, it is on the SSMI to “deploy reasonable and appropriate technical measures, including automated tools or other suitable mechanisms, to verify the accuracy of such declaration, having regard to the nature, format, and source of such information” and then label it accordingly. The biggest flaw is relying on the judgement of the user and then verifying the accuracy of the claim based on automated tools that are not foolproof. This places a significant burden on intermediaries, with nobody to share or dilute that responsibility. By requiring multiple layers of verification, the amendment creates an unclear and excessive onus on intermediaries only for hosting content. This seems unreasonable. The Indian government should make content hosting easier by distributing this responsibility in a fair manner.
A Comparative Lens: EU and US Regulatory Models
The gap in the current amendment, due to the labelling and transparency obligations, necessitates a comparative analysis of global jurisdictions, such as the EU and the US. Since these jurisdictions have a consolidated framework on AI regulation, it becomes necessary to analyse the proposed amendment in their relation.
The EU AI Act takes a contrasting approach in this aspect; the transparency obligations regarding the identification of AI-generated content are primarily placed on automated tool providers – those who develop, train, and deploy AI systems. Article 50 stipulates that AI system providers must clearly inform their users that their interaction is taking place with an AI system. Additionally, it mandates that these platforms ensure that artificially generated content is in a machine-readable format to facilitate identification. Unlike the Indian framework, it also creates an exception to this obligation, wherein tools for assistive editing functions that do not significantly alter the input data and systems authorised by law for detective and preventive measures are exempt from this obligation’s purview. Moreover, this framework adds an important element that the IT rules lack. It states that if any AI-generated text undergoes human review, then a natural or legal person must assume its editorial responsibility. Constructively, Article 50(7) explicitly empowers the AI office to standard guidelines on the detection and labelling of AI content.
Contrastingly, the draft US AI Labelling Act proposes a disclosure-based regulatory approach. Unlike the Indian and EU systems, the US framework shifts the onus to the developer of the automated tool itself. This means that AI tool developers must create a system that attaches disclosure to the automated content itself. The Act also applies to anyone who uses or licenses these systems in the future. The US framework employs a clear system where the identification of automated content is established at the creation stage itself. By mandating a unique identifier through the system that generates the AI content, the US framework marks a practical approach in maintaining transparency.
The burden of identification on the automated tool provider or developer is a crucial step in ensuring that no harm is done through the SGIs; however, intermediaries should also share this responsibility. Even though the IT amendment shares this responsibility with SSMIs, the new requirement, as explained above, overburdens the intermediary. A balance of liability is required to correct the compliance measure under the due diligence rules.
Conclusion And The Way Forward
The issue arises from the requirement to label SGI by an intermediary. This is fraught with challenges and a lack of technological advancement. It opens the floodgates for mistakes due to human and technological error. India can draw inspiration from the EU and US frameworks. India should establish a body such as the EU AI office, which creates uniform guidelines for identifying SGIs. These guidelines will help industries to follow a uniform approach, allowing platforms to agree on a standard method. Since all platforms will rely on a uniform detection mechanism, a change made on one platform will benefit the others. Intermediaries can adjust their systems accordingly to improve their own compliance. Taking this a step further, the government can mandate intermediaries and AI tool developers to flag all the problems in this AI identification mechanism publicly. With the help of this, platforms can easily access information about possible problems that may arise in their framework. This helps them study how other platforms have prevented similar issues. It also gives them enough time to take preventive measures to avoid future problems.
Additionally, sharing the liability with the AI tool developer is another positive step that takes inspiration from the US framework. By mandating the AI software to display a watermark on any AI-generated content, we can reduce the burden of verification for the intermediary. Moreover, the Indian framework should include obligatory exceptions for minutely altered content and automated tools used for a lawful purpose. This means that any content that is minutely altered through an input in an automated tool would not be required to follow the disclosure requirement. Similarly, any tool being used for a legal purpose, such as investigation and crime prevention, will also be freed from this obligation.
Although the identification of SGI is a positive step, concerning current technological development, the complete verification of automated content is far from reality. However, placing the onus entirely on the intermediary is merely shifting responsibilities and not solving the issue. Placing the onus on users in case of SSMI is a welcome step; however, labelling on the basis of risk identification can reduce the burden for intermediaries.
A three-tiered risk framework can be established for this purpose, i.e., low-risk content, medium-risk risk, and high-risk content. Creative work, such as satire and memes, will fall into the low-risk category, followed by low-impact SGI that do not deal with sensitive topics, which will then fall into the medium-risk category. SGIs that are very sensitive and impact public opinion will fall under the high-risk category, and these contents need to be labelled as AI-generated. For this purpose, a stakeholder body with members from civil society, the press and fact-checking organisations can be created. This will ensure neutrality, transparency and public trust in the categorisation of sensitive content. Since the content classification would not be in the hands of the government, genuine dissent would not be suppressed either.
Apart from this, users on numerous instances are able to identify AI-generated content on the platforms; however, there is nothing to report it to the intermediaries. Therefore, adding an SGI reporting feature that allows users to report AI-generated content can facilitate AI identification. This will enable outside experts, such as fact-checkers, civil society organisations, audit firms, or certified trusted flaggers, to examine and mark artificial content. When a trusted flagger alerts a platform about problematic content under such a system, the platform must respond promptly and give the notice top priority.
Yash Agarwal Ashar Nezami, both are Second year student at Dr. Ram Manohar Lohiya National Law University, Lucknow
