India’s DPDP Rules and your LLM feature: five questions to ask
This is a plain-language orientation, not legal advice. If your product handles personal data in India, talk to a lawyer who knows the Act.
India notified the Digital Personal Data Protection (DPDP) Rules, 2025 in November 2025. They come with an 18-month phased timeline. The final phase, covering notice and consent standards, the obligations of data fiduciaries and the processing of children’s data, takes effect on 13 May 2027.
Figure 1 · The phased timeline
- November 2025DPDP Rules, 2025 notified. The first set of provisions, including those setting up the Data Protection Board, comes into force.
- Phased periodFurther provisions arrive in stages while organisations prepare.
- 13 May 2027Final phase: notice and consent standards, data fiduciary obligations, children’s data.
Five questions for an LLM feature
Figure 2 · Questions to ask
| Question | Why it matters |
|---|---|
| What personal data reaches the prompt? | Names, phone numbers, and account details often slip in with the user’s text. |
| Where is it sent and stored? | Provider logs, vector indexes and caches are all copies. Cross-border transfers have their own rules. |
| What notice and consent cover this use? | A privacy notice written before you added an LLM may not describe it. |
| Are children involved? | The Rules require verifiable consent from a parent or guardian before processing a child’s data, with limited exemptions. |
| Can you delete it on request? | Erasure has to reach the chat history, the logs and the embeddings, not only the main database. |
Practical habits
- Strip or mask personal data before it reaches a model when the task doesn’t need it.
- Keep a simple map of where prompts, outputs and embeddings are stored.
- Set retention periods for logs, and stick to them.
- Treat children’s data as a separate case from the start, not as an edge case.
Security and privacy overlap. The risks we described in the prompt injection note include leaking data the model can see.