Training AI Models With Personal Data: The true test of Puttaswamy

Nayona Roy | Partner

When does association with an event become evidence of criminal participation?

The recent 80’s trend on ChatGPT took social media by storm when millions of people tasked the model with reimagining their images in the fashion and style of the 80s. The consequence of the AI model having been permitted and instructed to read and process the personal attributes of each such individual was hardly made as 'viral. ' In 2025, the MIT Technology Review reported that a training data pool for one of the largest open-source image generation tools, DataComp CommonPool scraped from the web, contained thousands of images of personally identifiable information of millions of people. What is even more concerning is the fact that none of this data was actually uploaded with the intent of it ever being used to train a black-boxed technology tool on the physical attributes of real people.

Well before AI entered the living room conversations around the world, in 2017 the Supreme Court of India delivered a landmark judgement that changed the trajectory of privacy jurisprudence in the country. Justice K.S. Puttaswamy v. Union of India read the right to privacy into the fundamental right to life and personal liberty under Article 21 of the Indian Constitution. More pertinently, it sought to balance the commercial need for technological development and the right of privacy against the parameters of legality, necessity, and proportionality as primary tests for any intrusion into informational autonomy of an individual. In 2023 the Digital Personal Data Protection Act (“DPDPA”) gave teeth to this ruling, at least in theory, and the question remains if using personal data in AI model training sets can first be detected and second, be averted using the provisions of statute and the intention of the judgement.

Today every business is undoubtedly collecting huge volumes of personal data — a customer provides personal information to obtain a service; an employee shares information during the induction into an organisation; a professional publishes a photograph or employment history on a social media or networking site; and a patient provides information to a healthcare provider. This data is inevitably clocked into a data management system hosted on the servers of the organisation and/or a third-party SaaS provider, and the primary collector is held liable in a vicarious and fiduciary capacity even on behalf of the SaaS provider. Three critical questions are raised:

1.Is the third-party software provider bound to open themselves up to a governance audit by the client, or alternatively an uncapped indemnity in the event the client faces a breach?

2.What is the probability of the personal data being used directly or indirectly to train an AI model, and have either of the entities procured traceable consents from the concerned individual (the data principal) for each touchpoint of such training?

3.Does the training of the model on such data result in the transfer of the data overseas and a cross-border processing of data?

Globally, the debate is being shaped by the Global Data Protection Rules (GDPR) which has been mirrored in the DPDPA to a large extent and in the absence of case laws under the recently enacted DPDPA, one turns to the cases under GDPR for the interpretation of the statute.

As a rule of thumb, digital personal data can only be processed under the DPDPA under free, explicit and specific consent obtained from the Data Principal. However, a limited set of narrowly defined exemptions under Section 7 of the DPDPA permits data to be conditionally processed as legitimate uses without explicit consent. Section 7(a) carves out personal data made publicly available by the Data Principal as such a legitimate use, and most organisations interpret this provision to mean that data scraped from the open internet is "publicly available" for being used in training AI models. This couldn’t be farther from the truth, and the intent of disclosure gains primary importance in distinguishing between what a person has chosen to make public and what a scraper has simply been able to reach.

Irrespective of the ground of processing, the DPDPA mandates that the Data Principal must be given an explicit and specific notice for the processing of personal data in plain language, and such processing must be limited to the purpose explicitly provided in such notice. The exemption under Section 7(a) is contingent on the processing being limited to the specified purpose for which the Data Principal has voluntarily provided her personal data to the Data Fiduciary, and in respect of which she has not indicated to the Data Fiduciary that she does not consent to the use of her personal data. Therefore, a developer training a model on a data set containing personal data without an explicit notice to each and every individual whose data is used is in breach of the DPDPA irrespective of whether the data was voluntarily uploaded by the concerned individual with a different intention.

The Puttaswamy judgement compels us to ask whether an intrusion was necessary and proportionate to a legitimate aim and not merely convenient or utilitarian. Consent is no longer a checkbox compliance for an organisation. Its jurisprudence underlines the culture of the organisation in itself instilled across the lifecycle of every operation.