What DPDP means if your product uses AI or machine learning
AI products process vast amounts of personal data. Here is how the DPDP Act applies — and what your engineering and legal teams need to do before May 2027.
Does the DPDP Act apply to AI and machine learning products?
Yes. Any company that determines the purpose and means of processing personal data is a Data Fiduciary under the DPDP Act, and that includes AI product companies. If your model is trained on user data, if you collect feedback from users to improve your system, or if your API outputs personal data about identifiable individuals, you are in scope. The fact that processing happens inside a model rather than a traditional database does not change the classification.
Do you need consent to use personal data for AI model training under DPDP?
One of the most operationally significant requirements for AI products is that you need a lawful basis — most commonly consent — before using personal data to train or fine-tune a model. Scraping data from public sources without consent, or repurposing user data collected for one purpose to train a model for a different purpose, are both likely to be unlawful under the Act. Build consent into your data collection flow, not as an afterthought, and document which training datasets are covered by which consent basis.
What must a Data Processing Agreement with an AI or ML vendor include under DPDP?
If you use a third-party foundation model or ML API — OpenAI, Google Vertex, AWS Bedrock — and pass personal data to it in prompts or fine-tuning datasets, that vendor is your Data Processor and you need a Data Processing Agreement. The DPA must specify what the processor may do with the data, prohibit onward sharing without your approval, require deletion on termination, and mandate breach notification. Most major vendors now offer DPA templates, but you must review them against DPDP requirements specifically.
What rights do individuals have over automated AI decisions under DPDP?
Data Principals — the people whose data you process — have the right to access, correct, and erase their personal data, and to withdraw consent. For AI products this has a direct engineering implication: you need to be able to locate and delete an individual's data across your training sets, fine-tuning datasets, and inference logs. If your product makes automated decisions that significantly affect individuals — credit scoring, hiring, content moderation — those individuals may have the right to challenge the decision, which means you need explainability built in, not bolted on.
How do you implement privacy by design in an ML data pipeline?
The cleanest way to reduce DPDP risk in an AI product is to reduce the personal data in the pipeline. Anonymise or pseudonymise training data before use. Use synthetic data where the use case allows. Apply differential privacy techniques on sensitive datasets. Limit inference logs to what is needed for debugging and delete them on schedule. These are engineering choices, and making them early is far cheaper than retrofitting them under regulatory pressure.
What should AI product teams do to prepare for DPDP by May 2027?
Map every flow where personal data enters your AI system: collection, labelling, training, fine-tuning, inference, logging. Establish the lawful basis for each. Put DPAs in place with your model vendors. Build a data-subject rights workflow that can reach into your datasets. Review your privacy notice to accurately describe AI-specific processing. And set a May 2027 milestone to have all of this audit-ready.
Frequently asked questions
Does the DPDP Act apply if I only use anonymised data to train my model?
Genuinely anonymised data — where re-identification is not reasonably possible — falls outside the Act's definition of personal data, so you are not a Data Fiduciary for that data. The challenge is proving anonymisation is robust. Pseudonymised data, or data from which an individual could be re-identified by combining with other datasets, remains personal data and is in scope.
Do I need consent to use publicly available data to train my model?
Consent or another lawful basis is required for any processing of personal data, even if the data was publicly available. Using a LinkedIn profile or a public tweet in your training set without a lawful basis is likely unlawful under the DPDP Act. The public availability of data is not itself a lawful basis.
What happens if a user asks me to delete their data but it is baked into my model weights?
This is one of the hardest operational challenges in AI compliance. The Act requires you to erase personal data when consent is withdrawn or when it is no longer needed. Where personal data is entangled in model weights, you will need to either retrain the model without that individual's data, or demonstrate that re-identification from the weights is not technically feasible. Machine unlearning techniques are an active research area precisely because of this requirement.
Assess your AI product's DPDP readiness
Niti Bharat's DPDP Readiness Assessment covers AI and ML products specifically — consent frameworks, DPAs, rights workflows, and privacy-by-design checks across your data pipeline.
Start DPDP Readiness Assessment