Examine whether the Digital Personal Data Protection Act, 2023 and its Rules adequately protect citizens against privacy risks arising from AI training. Suggest targeted reforms.

AI models can "memorise" personal data in their weights during training and later regurgitate it [2]. India's DPDP Act, 2023 and DPDP Rules, 2025 were notified in November 2025 with an 18-month phased rollout [1]. They were designed for databases, not trained models. As a result, they give citizens a sound base but only partial protection.

Where the framework protects

• Principles: The Act is built on consent, purpose limitation and data minimisation, which gives effect to the Puttaswamy (2017) right to privacy [1].
• Time-bound rights: Requests for access, correction and erasure must be answered within 90 days [1].
• Accountability: Significant Data Fiduciaries must carry out independent audits and impact assessments. Security failures can attract penalties of up to ₹250 crore [1].

Where it falls short

• Public-data loophole: Personal data that a person has made public is excluded from the Act [3]. This is exactly the data that is web-scraped for AI training [2].
• Erasure stops at the file: Deleting the training data does not remove what the model has memorised. Full removal may need retraining, or may not be possible without destroying the model [2].
• Open-source spread: A deletion at the source does not reach copies that have been fine-tuned and redistributed [2].
• State exemptions: The government can exempt its own agencies by notification, with no safeguards specified [4]. This leaves AI built on public datasets unchecked.
• Transition gap: Duties are being phased in, so protection is incomplete in the meantime [1].

Targeted reforms

• Narrow the public-data exclusion where data is used for AI training. This fits the India AI Governance Guidelines' preference for targeted amendments over a new AI law [5].
• Data Protection Board guidance on "model erasure": Allow output filtering until the model is retrained, and back research on machine unlearning [2].
• Leak testing: Make regurgitation red-teaming part of SDF impact assessments [1][2].
• Privacy-enhancing technologies: Use differential privacy and federated learning at the training stage [2].
• Proportionate exemptions: State exemptions should be tested against Puttaswamy's necessity and proportionality standards.
• Regulator cooperation: Models are trained across borders, so the Board should work with foreign regulators [2].

The DPDP framework gets the principles right but applies them to files rather than to model weights. Targeted amendments can close this gap by covering scraping, defining what erasure means for a trained model and requiring leak testing. That would protect the Article 21 right to privacy while keeping India's AI growth on a trusted footing, so that innovation and individual dignity advance together.