IRIS8 Talk to us
Regulation

DPDP Act and LLMs: what the law asks of enterprise AI

What India's data protection law means for training and running a language model on personal data, and where a model in your own environment helps. This page is general information, not legal advice.

Your environmentYour cloud account (VPC), or on-prem as a scoped data-centre install in a managed engagement
You own the modelWeights, prompts and logs stay with you
8-week pilotOne workflow, a pass mark agreed in week 1
JurisdictionIndia

Last reviewed

01

The timeline

  1. 13 Nov 2025

    Rules 1, 2, 17–21 (Data Protection Board set-up and procedure)

    DPDP Rules, Rule 1(2)
  2. 13 Nov 2026

    Rule 4 (Consent Managers)

    Rule 1(3)
  3. 13 May 2027

    Rules 3, 5–16, 22, 23 (notice, safeguards, breach intimation, retention, children's data, Significant Data Fiduciaries, transfers)

    Rule 1(4)
02

What the Act asks, and what it means for an LLM

Duty What it requires (paraphrased) LLM implication
Lawful grounds (Act ss. 4–7) Process personal data with consent or for certain legitimate uses Check that your notice and consent cover the AI use, or that a legitimate use applies
Notice (Rule 3) Itemised description of the data and the specific purpose "Improving our services" may not describe model training clearly enough — get legal advice
Security safeguards (Act, s. 8(5); Rule 6) Reasonable safeguards including encryption, access control, logs and monitoring Prompts, outputs, training sets and logs all count
Breach intimation (Act, s. 8(6); Rule 7) Tell the Board and each affected person A leaked training set or log is a breach
Erasure (Act, s. 8(7); Rule 8) Erase data once the purpose is served or consent is withdrawn Track which records trained which model version
Significant Data Fiduciaries (Act, s. 10; Rule 13) Periodic DPIA and audit; Rule 13 also asks for due diligence that algorithmic software doesn't risk Data Principals' rights Large banks, insurers and platforms may be designated — plan for model documentation
Cross-border transfer (Act, s. 16) Transfers allowed except to countries the government restricts; sector rules that are stricter still apply Sector rules such as RBI's payment-data localisation can be stricter than the Act

Act text: Digital Personal Data Protection Act, 2023 (Act No. 22 of 2023), Gazette of India, 11 Aug 2023. Rules: MeitY.

See how this would work on your data.

Book a call
03

Why where the model runs matters in India

Section 16 lets the government restrict transfers to named countries, and it preserves sector laws that set higher protection. For banks, RBI already requires payment-system data to be stored only in India (RBI circular RBI/2017-18/153). A model trained and served in an Indian cloud region or your own data center avoids the question of where prompts, logs and support access go.

04

What we'd do first

  1. 1
    Map the personal data a target workflow uses, check that your notice covers AI use, and set up lineage so each model version records which data trained it — so erasure and audit requests have an answer by May 2027.
  2. 2
    Then a narrow task, trained in-country, in an 8-week pilot.
  3. 3
FAQ

Frequently asked questions

When does the DPDP Act apply to our AI?

Most Data Fiduciary duties apply from 13 May 2027 (Rule 1(4)).

Does DPDP ban sending data abroad?

No. It allows transfers except to countries the government restricts, and stricter sector rules still apply (Act, s. 16).

Is publicly available data exempt?

The Act excludes personal data made publicly available by the person or under a legal duty (Act, s. 3(c)(ii)). Scraped data won't always meet that test; get advice.

What are the penalties?

The Schedule sets penalties up to ₹250 crore for failing to take reasonable security safeguards (Act, Schedule, item 1).

Do you deploy in India?

Yes, in your Indian cloud region, or on-prem as a scoped data-centre install in a managed engagement.

Talk to us

A 30-minute call with the IRIS8 team about your data, your environment and a first pilot.

Talk to us