1 November 2026 • 6 min read
If your product keeps chats, documents, voice notes or support tickets from Indian users, someone on your team will ask whether you can train or fine-tune a model on them. India's Digital Personal Data Protection Act, 2023 does not mention AI training. You have to work it out from four provisions: the consent rule, the public-data exclusion, the research exemption, and the Rules that fill in the research standards. This page sets those out in plain words, says where the law is unsettled, and gives a short list of what a product team can build now. It applies to any business that decides why and how personal data is processed (a "Data Fiduciary" in the Act), including Indian startups and foreign firms that offer goods or services to people in India.
The Act was enacted on 11 August 2023. The Rules were published on 13 November 2025 (the Gazette is dated 13 November; the PIB release is dated 14 November). The Government has switched on the Act in stages. A section-by-section commencement notice, as reproduced by dpdpa.com, brings sections 3 to 5, most of section 6, sections 7 to 10 and sections 11 to 17 into force 18 months after publication. In the Rules, Rule 16 (the research exemption) follows the same 18-month date.
That puts the main duties at about mid-May 2027. On 1 November 2026 they are not yet in force. Datasets you build today will still be in use after that date, so design for the rules now.
Section 6 says consent must be "free, specific, informed, unconditional and unambiguous" and limited to the personal data necessary for the specified purpose. The Act's own example is a telemedicine app that also asks for the phone contact list: consent covers only what the stated service needs.
Applied to training: a user who agreed to your terms to use a chatbot has not agreed to have their chats used to improve a model, unless your notice said so as a separate purpose. The Rules require a standalone notice that is clear and states the specific purpose, and the Act requires the notice to be offered in English or any Eighth Schedule language. A buried line in a privacy policy is the weak spot.
Consent can be withdrawn at any time, and withdrawing must be as easy as giving consent. On withdrawal the fiduciary must stop processing within a reasonable time and cause its processors to do the same. The Act also says a fiduciary must erase personal data when consent is withdrawn or the purpose is no longer served, unless a law requires it to keep the data. Whether personal data that has already shaped a model's weights counts as "held" is not answered by any source we opened. Build the practical part now: keep a list of which user IDs went into which training set, so a withdrawn user is left out of the next run.
Section 3(c)(ii) says the Act does not apply to personal data made public by the person it relates to, or by someone legally required to publish it. The Act's example is a person who posts her own views, with her own details, on social media.
Read it narrowly. On the text, it covers data the person published about themselves, and data that a law forces someone to publish. It does not clearly reach a forum post in which a third party names a person, or a database that is public only because it leaked. A law-firm note from Cyril Amarchand Mangaldas (August 2026) calls this a narrowly conditioned exemption that offers a much less clear safe harbour for AI training. Treat each web source as its own question: who published it, and were they the person it describes?
Section 17(2)(b) exempts processing "necessary for research, archiving or statistical purposes" if the data is not used to take any decision specific to a Data Principal and the processing follows prescribed standards. Rule 16 repeats this and points to the Second Schedule of the Rules. The Schedule asks for technical and organisational measures so that processing is lawful, limited to the data necessary for the purpose, reasonably accurate, kept only as long as needed, protected by reasonable security safeguards, and that someone is accountable.
Can commercial model training be "research"? The statute does not say. The Delhi High Court's interim judgment of 24 July 2026 in ANI v OpenAI held, under the Copyright Act, that training an LLM qualifies as research. The same law-firm note argues the reasoning is persuasive for the DPDP Act but is not binding on it, and says the judgment "does not finally resolve" how the Act applies to AI training. Our reading: a general-purpose model that does not score, rank, approve or price individual users is the easier case. A model fine-tuned to make decisions about each customer, such as credit or pricing, is the harder one, because of the "decision specific to a Data Principal" condition. Do not rely on this route until the Board or a court says so.
Related pages on this site: consent withdrawal workflows, fine-tuning versus prompting and labelling rules for AI-generated content.
This is a reading of the published texts, not legal advice; check the Act, the Rules and your own facts with a qualified lawyer before you train on user data.
No. It does not mention AI training. Personal data used for training must fit a lawful basis: consent for a specified purpose, the narrow public-data exclusion, or the research exemption under Rule 16.
Only if the Act does not apply to it, and that exclusion is narrow. It covers personal data the person made public themselves, or that a law requires someone to publish. Other scraped personal data stays inside the Act.
With consent for that purpose, yes. Without it, you would have to rely on the research exemption, which is untested for commercial training and fails if the model is used to take decisions about individuals.
Not yet. The Rules and the commencement notice set the main duties, including Rule 16, 18 months after the November 2025 Gazette.
Nine stories, every morning at 09:15 IST.
Get the brief free →