Free DPDP compliance tools. Scan your website, map your data flows and run a DPIA yourself. Try the free tools →
    Compliance

    DPDPA: What AI Companies in India Need to Know (Before May 2027)

    DPDPA requires consent for AI training on personal data, with almost no exemptions. By May 2027, you need fresh consent, anonymization strategies, and user rights infrastructure. Audit your gaps now or risk ₹250 crore penalties.

    By TruConsent Team
    5 min read
    DPDPA: What AI Companies in India Need to Know (Before May 2027)

    Introduction

    India's new data protection law (DPDPA) isn't just another compliance checkbox. It's a fundamental shift in how AI companies can handle personal data. If you're training models or processing user data at scale, here's what changed.

    The Consent Problem

    The Digital Personal Data Protection Act (DPDPA) 2023, backed by rules notified in November 2025, operates on a simple principle: no consent, no processing.

    Most ML pipelines assume data availability. You scrape forums, repurpose user logs, fine-tune on customer interactions, build recommendation systems from behavioral traces. Under DPDPA Section 4, all of this requires explicit, informed, and specific consent. Not the vague “we may use your data to improve services” buried in a 2019 ToS.

    What “specific consent” actually means

    • Purpose-bound: “Training large language models” is valid. “Improving user experience” is too vague.
    • Free, specific, informed, unconditional, and unambiguous: Users must understand what data you're taking and why, with clear affirmative action. No legalese hiding the AI training bit. No pre-ticked boxes or bundled consent.
    • Withdrawable: Users can revoke consent anytime. Your pipeline must handle deletion or anonymization post-withdrawal.

    If you trained a model on user data collected before 2023, you're likely non-compliant. The law doesn't grandfather legacy datasets. You need fresh consent.

    The Children Problem

    Section 9 and Rule 10 require verifiable parental consent for anyone under 18. This isn't “click if you're 18+” theater. It's age verification plus guardian approval.

    For consumer AI apps (chatbots, ed-tech, social platforms), this means building dual consent systems or excluding minors entirely.

    What Counts as Personal Data in AI?

    The Rules define personal data broadly. In ML pipelines, this includes:

    • Obvious stuff: Names, emails, phone numbers, IP addresses
    • Behavioral data: Clickstreams, search queries, purchase history, session logs
    • Derived data: Credit scores, health risk predictions, personality profiles your model generates
    • Embeddings and pseudonymized data: If re-identification is possible, it's still personal data

    Anonymization is an exit, but the bar is high. K-anonymity or hashing alone may not suffice if the Data Protection Board determines re-identification risk exists.

    What You Need to Do

    1. Build a Consent Registry

    Track what data was collected, when, under which consent. Store consent artifacts. Log purpose declarations. Handle withdrawal requests with corresponding data deletion.

    For fine-tuning or transfer learning, ensure upstream datasets (OpenAI, Hugging Face, Kaggle) came with valid consent or are truly anonymized. You inherit liability for non-compliant data.

    This is the problem TruConsent.io solves. Purpose-specific consent flows with audit trails, withdrawal automation, lifecycle tracking. It integrates via API into your data layer so every row in your training data has a valid consent ID.

    2. Audit for Over-Collection

    Section 4 mandates collecting only what's necessary. Grabbing entire user profiles when you only need transaction timestamps is a violation.

    Watch for feature engineering that creates new personal data categories (deriving religion or health status from non-sensitive inputs). Model outputs predicting individual behavior are personal data tied to the input data's consent scope.

    TruScanner automates this. It crawls your pipelines and flags sensitive data inference, purpose drift (fraud detection data repurposed for marketing), and stale consent. Integrates with Snowflake, BigQuery, Databricks.

    3. Handle Cross-Border Transfers

    The Rules haven't imposed blanket localization yet, but Section 16 empowers the government to restrict transfers to countries without adequate protection.

    If you're using US/EU cloud infrastructure or offshore GPUs, mirror critical data in India. Ensure DPAs and SCCs are DPDPA-aligned, not just GDPR boilerplate.

    4. Automate User Rights

    Section 11, 12 & 13 give users rights to access, correction, erasure, and grievance redressal. Your systems must support data portability (machine-readable exports) and the right to erasure — including from training checkpoints.

    The hard part for AI companies is deleting data from frozen training checkpoints or RLHF datasets. You need model re-training protocols or proof the data was anonymized pre-training.

    TruConsent’s Data Principal Rights (DPR) engine handles this: a self-service portal for requests, automatic mapping to all systems holding that user’s data, export generation, deletion workflows, and audit logs. For ML systems, it includes workflows for marking training samples for exclusion or triggering differential privacy noise injection.

    The May 2027 Deadline

    The Data Protection Board is already operational. Penalties under Section 33 can happen today. They go up to ₹250 crore for breaches involving children’s data or cross-border violations. Even smaller fines (₹50–200 crore) are existential for startups.

    Start now:

    1. Audit your training data sources and consent gaps
    2. Implement consent infrastructure that's purpose-specific and logged
    3. Invest in privacy-preserving ML (federated learning, differential privacy)
    4. Appoint a Data Protection Officer

    Why This Matters

    Compliance is expensive, but it’s also a moat. Companies that build privacy-first AI infrastructure now will be harder to displace. Users and enterprises will prefer vendors with auditable consent.

    The startups that treat DPDPA as an engineering problem will dominate India’s AI landscape by 2028. The rest will be firefighting DPB notices.

    We built TruConsent and TruScanner for AI companies facing this problem. We've worked with enterprises processing billions of records and know the pain of retrofitting consent into ML pipelines.

    If you're training models on Indian user data without a consent strategy, you're less than 16 months from a compliance crisis.

    truConsent | Purpose Built for DPDPA

    Continue Reading