India's Digital Personal Data Protection Act 2023 (DPDP) applies to every AI system that processes personal data of any individual in India — regardless of where the model is trained, where the company is incorporated, or which cloud the inference runs on. The five compliance pillars that matter for AI specifically are lawful consent, purpose limitation, data fiduciary obligations, significant data fiduciarydesignation (likely for most AI companies), and cross-border transfercontrols. The DPDP Rules notified in 2025–26 narrow several open questions, but leave model-training, automated decision-making, and synthetic-data treatment to sector-specific guidance.
- DPDP Act (Digital Personal Data Protection Act 2023)
- India's comprehensive data-protection law, enacted in August 2023, with phased enforcement through 2025–26. It governs processing of digital personal data of individuals in India by any entity (data fiduciary) — Indian or foreign — and creates the Data Protection Board of India (DPBI) as the regulator. Penalties go up to ₹250 crore per breach class.
- If your AI processes personal data of any individual in India, DPDP applies — full stop. Extra-territorial scope is real.
- Consent under DPDP is narrower than GDPR's lawful-basis menu — for most AI use cases you need explicit, specific, free, informed consent.
- Most AI companies will be designated Significant Data Fiduciaries, triggering DPIA, audit, and DPO obligations.
- The DPDP penalty ceiling (₹250 cr per breach class) is high enough that AI governance is now a board-risk item, not an engineering item.
- Cross-border transfer of personal data is permitted except where the government notifies a restricted list — design for both regimes.
Why DPDP matters more for AI than for any other software category
AI systems are unusually exposed to data-protection law for five reasons:
- Training data is personal data. If your training set includes any identifiable Indian individual's data, you are processing personal data — even if the model itself is "just numbers."
- Inference is processing. Every query about an individual that the model answers is a processing event, requiring a lawful basis.
- Outputs may be personal data. A generated image, biography or analysis of a named individual is personal data about that individual.
- Re-identification risk. Aggregated or anonymised data feeding AI is increasingly re-identifiable. DPDP treats this seriously.
- Automated decision-making. While DPDP is less explicit than GDPR on automated decisions, the principles around purpose limitation, fairness, and grievance redressal pull in the same direction.
The five DPDP pillars for AI compliance
1. Lawful consent
DPDP's lawful-processing basis is narrower than GDPR's. The default basis is consent — and DPDP consent is specifically defined as free, specific, informed, unconditional, unambiguous, signified by clear affirmative action. Bundled consent, dark-pattern consent, and pre-ticked boxes are non-compliant. For AI training data specifically, this means: pre-2023 data sets gathered without DPDP-grade consent need re-consent or removal before they can be used in models that process Indian personal data after enforcement begins.
2. Purpose limitation
Data collected for purpose A cannot be processed for purpose B without fresh notice and (typically) fresh consent. The implication for AI: collecting customer support transcripts for "to provide customer service" does not give you a basis to train a model on them. This is the single most-overlooked DPDP failure in Indian AI products.
3. Data fiduciary obligations
Every entity processing personal data is a Data Fiduciary, with obligations: notice, consent management, accuracy, retention limits, security, breach notification, grievance redressal. These are non-negotiable baseline duties that apply whether you're a 10-person AI startup or a 50,000-person bank.
4. Significant Data Fiduciary (SDF) designation
The Central Government may notify any data fiduciary as an SDF based on volume, sensitivity, risk to rights, sovereignty, public order, and electoral democracy. AI companies — particularly those processing high volumes of personal data, or deploying models that influence consumer decisions at scale — will mostly fall inside. SDF designation triggers three heavy obligations: appointment of a Data Protection Officer based in India, periodic Data Protection Impact Assessment, and periodic independent audit.
5. Cross-border transfer
DPDP permits cross-border transfer of personal data except to countries the government notifies as restricted. The list is reverse-blacklist — open by default, restricted by exception. The practical design: assume transfers are permitted, but build the architecture so you can localise specific jurisdictions when the list updates. EU/UK customer data already has parallel GDPR obligations — see GDPR in India vs DPDP for the dual-regime design pattern.
The DPDP penalty structure (and why AI boards should care)
Penalties are class-based, with a ceiling per class. The headline numbers:
- Failure to take reasonable security safeguards leading to breach: up to ₹250 crore
- Failure to notify the Board and affected Data Principals about a breach: up to ₹200 crore
- Failure to fulfil additional obligations of SDFs: up to ₹150 crore
- Non-fulfilment of obligations to children's data: up to ₹200 crore
- Breach of any other DPDP provision: up to ₹50 crore
The numbers are large enough that AI governance is no longer an engineering line item — it's a board-level risk register entry. Most Indian AI startups treat DPDP as a checklist; the right posture is to treat it as material risk.
DPDP vs other AI-relevant Indian frameworks
| Question | Primary framework | Where to read more |
|---|---|---|
| Can we process this personal data for AI? | DPDP | This article |
| What controls do we need around the AI system? | ISO 42001 / NIST AI RMF | ISO 42001 essay |
| What broader Indian laws apply to AI? | DPDP + IT Act + MeitY guidelines + sector rules | AI laws in India essay |
| What does an AI audit look like? | ISO 42001 + sector frameworks | AI audit India essay |
| How do we govern third-party AI use? | ISO 42001 A.9 + DPDP data fiduciary | AI governance framework essay |
The DPDP + AI compliance roadmap for an Indian company in 2026
- Data inventory. Map every personal-data flow into and out of every AI system — training, fine-tuning, inference, logging, evaluation.
- Lawful basis check. For each flow, document the lawful basis. For most AI use, this will be specific informed consent.
- Notice rewrite. Update privacy notices to DPDP-grade specificity, including the AI-specific processing.
- Consent re-collection. Where pre-DPDP consent does not meet the new standard, re-collect or remove the data from AI pipelines.
- DPIA programme. Run a DPIA on every AI system before launch and on a defined cadence afterward.
- DPO appointment if SDF designation is plausible.
- Audit cadence. Pair the DPIA with an AI audit against ISO 42001 or NIST AI RMF — DPDP is about personal data, the audit is about the AI system end-to-end.
- Cross-border architecture. Design so you can localise specific jurisdictions when (not if) the restricted list updates.
- Breach playbook. 72-hour-style notification readiness; the DPDP timeline is "without delay."
- Grievance redressal. A real, named, contactable mechanism — not a generic support email.
Where this fits in the wider Indian AI legal landscape
DPDP is the privacy spine. The broader legal landscape — IT Act, MeitY guidelines, sector rules from RBI/SEBI/IRDAI, the upcoming Digital India Act — is covered in our companion essay on artificial intelligence laws in India. The governance posture that lets you operationalise all of it sits in the AI governance framework. And the privacy-management certification pathway most Indian AI companies will choose to demonstrate DPDP alignment is ISO 27701 — where Dr. Sodhi is lead auditor of record.
Frequently asked
- Does DPDP apply to AI systems?
- Yes — to every AI system that processes personal data of any individual in India, regardless of where the model is trained, where the company is incorporated, or which cloud the inference runs on. Training data, inference inputs, and identifiable outputs are all in scope.
- What lawful basis can I use for AI training data under DPDP?
- The default basis is consent — and DPDP consent is specifically defined as free, specific, informed, unconditional, unambiguous, signified by clear affirmative action. Bundled consent and pre-ticked boxes are non-compliant. Pre-2023 training data often needs re-consent or removal.
- Will most AI companies be designated Significant Data Fiduciaries (SDFs)?
- Probably yes. SDF designation triggers Data Protection Officer (DPO) appointment, periodic DPIAs, and periodic independent audits. AI companies with high data volumes or models that influence consumer decisions at scale will mostly fall inside.
- What's the maximum DPDP penalty?
- Up to ₹250 crore per breach class — failure to take reasonable security safeguards leading to a personal-data breach carries the headline penalty. Other classes range from ₹50 cr to ₹200 cr. AI governance is now a board-risk item, not an engineering item.
- How does DPDP interact with cross-border data transfers for AI?
- DPDP permits cross-border transfer except where the government notifies specific countries as restricted (reverse-blacklist). Design the architecture so you can localise specific jurisdictions when the list updates. EU/UK customer data has parallel GDPR obligations on top.
Audit your AI system against DPDP this week.
Pick the DPDP frame and our auto-audit produces a ~25-page report mapping every AI data flow to a DPDP obligation, with severity ratings and a ranked remediation list. Or book a 1-hour consult to scope an ISO 27701 + DPDP engagement. ₹799 for the auto-audit, ₹2,500/hour for the consult.
Bharat NeuroTech offers self-serve AI audits across 12 global and Indian standards from ₹799, with Dr. Sodhi personally signing engagements under ISO/IEC 42001, 27001 and 27701.
— Bharat NeuroTech · /ai-audit