How AI Training Models could compromise Privacy?

Why in News?

The increasing use of personal and publicly available data by technology companies to train Artificial Intelligence (AI) models has raised concerns over the protection of individual privacy.

India, being a major market for platforms such as Meta’s Facebook and Instagram, faces particular challenges regarding the collection, scraping and commercial use of users’ data. This has brought into focus the adequacy of the Digital Personal Data Protection (DPDP) Act, 2023 in addressing emerging privacy risks associated with AI.

Evolution of India’s Data Protection Framework

• Constitutional Basis: The Supreme Court recognised the Right to Privacy as a Fundamental Right under Article 21 in K.S. Puttaswamy v. Union of India (2017).
• Earlier Framework: The IT Act, 2000, and its 2011 Rules primarily regulated sensitive personal data, including financial, medical and biometric information. Other personal information, including names, addresses and online activities, received relatively limited protection.
• Current Framework: The DPDP Act, 2023, and DPDP Rules, 2025, provide the principal framework for digital personal data protection. However, implementation is phased, with most substantive obligations becoming effective in May 2027.

How AI Training Models Threaten Privacy?

AI models are trained on massive datasets collected from websites, social media platforms, photographs, videos and other digital sources. This creates the following risks:

• Data Memorisation: AI models can memorise personal information in training datasets and reproduce it through carefully designed prompts, potentially exposing confidential information.
• Re-identification of Individuals: AI techniques can combine anonymised datasets with publicly available information to identify individuals and reconstruct their personal profiles.
• Inference of Sensitive Attributes: Ordinary digital footprints, such as location tags, photographs and social media interactions, can reveal sensitive attributes, including health conditions, emotional states and political preferences.
• Unauthorised Profiling: Personal information collected for one purpose may be reused for AI training and behavioural profiling without meaningful user awareness or consent.

Constitutional and Governance Implications:

Key Gaps in India’s Data Protection Framework:

• Exclusion of Publicly Available Data: Section 3(c) of the DPDP Act excludes personal data made publicly available by individuals or by persons legally required to make it public. This creates a regulatory gap concerning publicly accessible social media content, photographs and other digital footprints used for AI training.
• Absence of a Separate Sensitive Data Category: Unlike the IT Rules, 2011, the DPDP Act does not create a separate category of sensitive personal data requiring heightened protection. Consequently, inherently sensitive information, such as biometrics, passwords and health records, is not accorded a distinct statutory protection tier.
• Lack of Specific Provisions for AI Training: The Act does not explicitly address the use of personal data for AI model training, fine-tuning or development. It also does not expressly provide individuals with a separate right to object to the use of their publicly available data for AI training.
• Limitations of the Right to Erasure: Section 12 provides individuals with the right to seek erasure of personal data, subject to applicable legal requirements. However, its application to information already memorised by trained AI models and sensitive attributes inferred from personal data remains unclear.
• Ambiguity in Research Exemptions: Section 17(2)(b) exempts processing necessary for research, archiving or statistical purposes, provided it is not used to make decisions specific to individuals and complies with prescribed standards under the DPDP Rules, 2025. However, uncertainty over how these conditions apply to commercial AI training could create scope for regulatory avoidance. 
• Absence of a general Right to Object: Unlike the opt-out mechanism offered by Meta to certain users in the European Union for the use of their information in AI development, Indian users lack an equivalent general statutory right to object to such processing.

Way Forward:

• Extend DPDP Safeguards: Explicitly clarify the applicability of data protection obligations to personal data used in AI training, fine-tuning and model development.
• Introduce a Right to Object: Provide individuals with meaningful mechanisms to object to the use of their personal data, including publicly available information, for commercial AI training.
• Strengthen Sensitive Data Protection: Introduce differentiated safeguards for high-risk information such as biometric identifiers, health records and authentication credentials.
• Clarify Research Exemptions: Define the scope of research exemptions to prevent commercial AI development from automatically escaping data protection obligations.
• Strengthen Erasure Mechanisms: Clarify how erasure rights apply to AI-generated inferences and develop technical solutions such as machine unlearning to remove the influence of personal data from trained models, where feasible.
• Promote Privacy-by-Design: Encourage privacy-enhancing technologies such as differential privacy, federated learning and encryption, alongside independent audits and greater transparency in AI training datasets.

India must strike a balance between AI-led economic growth and the constitutional protection of individual privacy. Rather than restricting the use of personal data altogether, the focus should be on ensuring its lawful, transparent and accountable use.

Mains Practice Question:

Q. The growing use of personal data in training Artificial Intelligence models has exposed gaps in India’s data protection framework. Comment.

PDF