How AI Training Models Compromise Privacy

HOW AI TRAINING MODELS COULD COMPROMISE PRIVACY

Why in the News?

• The growing use of Artificial Intelligence (AI) has raised concerns over the adequacy of India’s existing legal framework for protecting the Right to Privacy, recognised as a Fundamental Right by the Supreme Court in the Puttaswamy judgment (2017).
• Existing data protection laws have certain gaps in regulating the use of personal data for AI model training.
• There is a need to strengthen privacy safeguards through targeted legal and policy reforms without undermining AI innovation, productivity and economic growth.

Privacy Coverage: Gaps in India’s Data Protection Framework

Legal framework prior to the DPDP Act, 2023

• Before the DPDP Act, 2023, the regulation of personal data was primarily governed by the Information Technology Act, 2000 and IT Rules, 2011.
• These provisions primarily protected specified categories of Sensitive Personal Data, including:
  • Financial information.
  • Medical and health records.
  • Sexual orientation.
  • Biometric information.
• The transfer of sensitive personal data was subject to certain safeguards, including consent or contractual necessity and an assurance of adequate data protection by the recipient.
• Unauthorised disclosure attracted penalties of up to 3 years’ imprisonment or a fine of ₹5 lakh.
• However, the framework did not adequately cover other personal information, such as names, addresses, phone numbers, caste, religion, political affiliations, browsing history, social media activity and behavioural patterns.

Regulatory gap concerning publicly available data

• The earlier framework did not adequately regulate publicly available data scraped from social media platforms.
• This is significant because India represents a major user base for platforms such as Facebook and Instagram.
• Technology companies such as Meta, Amazon and Google use user-generated data to train their AI models, raising concerns regarding privacy and consent.

Limitations of the DPDP Act, 2023

– The DPDP Act, 2023, provides a general legal framework for regulating the processing of citizens’ personal data. However, certain gaps remain:

• Exclusion of publicly available data: Publicly accessible personal data is excluded from the Act’s protection.
• Lack of regulation of inferred data: Information inferred through social media activity, images, videos, location and other behavioural parameters may not receive adequate protection.
• No separate category for sensitive personal data: Unlike the earlier IT Rules, the DPDP Act does not create a distinct category of sensitive personal data.
• Uniform treatment of personal data: Highly sensitive information, such as passwords, biometric details and health conditions, is treated within the same broad category as names, marital status, city and shopping preferences.

Need for enhanced protection of sensitive data

• The absence of a separate sensitive data category may weaken protection for information that can cause serious harm if misused.
• Even seemingly non-sensitive personal information can reveal sensitive attributes when analysed in combination with other datasets.
• However, certain information, such as biometric details and passwords, poses an inherent risk of misuse even without further inference.
• Therefore, such categories of personal data require stricter safeguards and higher levels of legal protection.

Training AI Models on Personal Data: Privacy Concerns

Risk of personal data exposure through AI training

• The DPDP Act does not specifically address the use of personal data for training AI models.
• AI models may regurgitate or reproduce personal information memorised during their training.
• Studies have demonstrated that:
  • Personal details can be extracted from AI models by querying them with seemingly anonymised data.
  • AI techniques can combine publicly available information to identify individuals and construct their personal profiles.
• Consequently, even legally anonymised datasets may expose individuals’ identities and sensitive information when processed through AI models, without their consent.

Absence of consent and opt-out mechanisms

• Indian users currently lack an effective legal mechanism to prevent their personal data from being used for AI training.
• Since the DPDP Act’s substantive consent provisions are scheduled to take effect in May 2027, organisations are not yet fully subject to these requirements.
• Even after the Act becomes fully operational, users will have no specific right to object to the use of their publicly available personal data for AI training.
• In contrast, Meta provides users in the European Union with an opt-out mechanism to object to the use of their data for AI training.

Limitations of the Right to Erasure (Right to be Forgotten)

• Under the DPDP Act, the right to erasure applies to personal data processed with the user’s consent, subject to applicable legal requirements.
• It remains unclear whether this right extends to information inferred from the analysis of personal data.
• The Act also provides exemptions for processing personal data for research, archiving and statistical purposes.
• The conditions governing these exemptions are to be prescribed by the Union Government.
• A key concern is whether AI model training could be classified as research or service-improvement activity, potentially allowing organisations to circumvent certain data protection obligations.

What Can Be Done? (Way Forward)

Strengthening the DPDP Act, 2023

• Address existing regulatory gaps through targeted amendments, notification of Rules and judicial interpretation.
• Extend the Act’s provisions to personal data used for training AI models.
• Introduce a right to object (opt-out mechanism), allowing individuals to prevent the use of their personal data for AI training.
• Extend data protection safeguards to:
  • Publicly scraped personal data.
  • Information inferred through AI-based analysis.
• Strengthen the Right to Erasure by requiring companies to inform individuals about all entities with which their personal data has been shared.

Clarifying legal provisions and enhancing safeguards

• Clarify whether AI model training falls within the research exemption under the DPDP Act.
• Introduce a separate category of Sensitive Personal Data with stricter conditions for its collection, use and processing.
• Balance privacy protection with technological innovation and economic development through a calibrated regulatory approach.

Balancing AI innovation with data protection

• India’s objective should not be to completely prohibit the export or use of personal data for AI training, as this could adversely affect:
  • Inclusion and accessibility of AI.
  • Technological efficiency and innovation.
  • Representation of India’s diverse population in AI models.
  • Risks of discrimination and algorithmic bias.
• Instead, the focus should be on ensuring responsible data usage through adequate privacy safeguards.

Need for urgent regulatory intervention

• These concerns are supported by growing research, industry practices and government-level discussions on India’s role as an exporter of data for AI training.
• With AI increasingly being deployed by corporate giants, MSMEs and non-profit organisations, India is emerging as a major global use-case hub for AI.
• Delaying privacy reforms could weaken the effective enforcement of the Right to Privacy as a Fundamental Right under Article 21 of the Constitution.

Source: https://www.thehindu.com/opinion/lead/how-ai-training-models-could-compromise-privacy/article71658179.ece

Mains question

“Examine the challenges posed by AI-driven data processing to the Right to Privacy in India. Discuss the limitations of the DPDP Act, 2023, and suggest measures to balance data protection with technological innovation.”