Guide
AI systems processing personal data (DPDP)
Training, fine-tuning and running AI on personal data is processing under DPDP. The ordinary duties apply, with an extra Rule 13 duty for Significant Data Fiduciaries.
Short answer
Training, fine-tuning or running an AI system on digital personal data is processing under s.2, so the ordinary DPDP duties apply: a ground under s.4, a notice that itemises each purpose (Rule 3), consent limited to the data necessary for the specified purpose (s.6), accuracy where outputs drive decisions (s.8(3)) and erasure when the purpose is served (s.8(7)). Treat training and inference as separate purposes, and put AI vendors under processor contracts (s.8(2)). If you are a Significant Data Fiduciary, Rule 13 also requires due diligence that algorithmic software does not risk Data Principals' rights.
Where personal data enters an AI system
| Stage | Example (sample) | DPDP points |
|---|---|---|
| Training | Fine-tuning a support model on past tickets at Example Finance | Is training a purpose in the notice, and is the data necessary for it? (Rule 3, s.6) |
| Inference | A model scoring a loan application | Accuracy where the output drives a decision (s.8(3)) |
| Prompts and logs | Staff pasting customer details into a chat assistant | Security, access control and log retention (Rule 6) |
| Outputs | A summary that names a customer | Output about an identifiable person is personal data (s.2) |
| Vendors | A hosted model API | Processor contract, location and sub-processors (s.8(2), s.16) |
Vendor AI tools that staff already use are easy to miss, so start the inventory there.
Training and inference are different purposes
Data collected to provide a service was collected for that purpose. Using it to train a model is often a further purpose, and then it needs its own ground: consent that names it, or a legitimate use under s.7 that genuinely covers it. See purpose limitation.
Section 3 excludes personal data made publicly available by the Data Principal, or under a legal obligation. That is narrower than anything that can be found online, so check with counsel before you collect public web data for training.
When consent is withdrawn or the purpose ends, s.8(7) requires erasure. Decide how you will handle data already used in training before you start, and record your approach.
Steps to govern AI use of personal data
- 01Inventory models and datasets
List every model you build or buy and every dataset that feeds it, including vendor tools that staff use.
- 02Name the purposes
Write each purpose separately: training, evaluation, inference and monitoring.
- 03Check the ground
Confirm consent or a s.7 legitimate use for each purpose, and update notices where a purpose is new (Rule 3).
- 04Minimise
Remove or mask identifiers the purpose does not need before training.
- 05Assess risk
Run an impact assessment for models that make or support decisions about people.
- 06Contract vendors
Put AI providers under processor contracts that cover their use of your data for their own training, location and deletion.
- 07Monitor and erase
Check accuracy over time (s.8(3)), and act on withdrawals and expired purposes (s.8(7)).
AI vendors as processors
- Does the vendor use your inputs or outputs to train its own models? If so, it may be acting as a fiduciary for that use.
- Can you switch off any use of your data for the vendor's model improvement, and is it off by default?
- Where are prompts, outputs and logs stored, and for how long?
- Which sub-processors host the model?
- Can the vendor delete one person's data on request?
- How soon will the vendor tell you about a breach?
The vendor due diligence guide covers the full questionnaire.
AI governance checklist
- Model and dataset inventory kept current
- Purposes named separately in notices
- Ground recorded for each purpose
- Identifiers masked or removed where not needed
- Impact assessment for decision-making models
- Vendor contracts cover training on your data
- Approach to withdrawal and erasure documented
- Rule 13 algorithmic due diligence, if you are an SDF
This guide covers DPDP. If you also serve people outside India, other laws may apply; check with counsel.
How TruePrivacy helps
AI governance keeps an inventory of the AI models and datasets you use, with assessments. Impact assessments run as a DPIA or PIA from templates or blank, and for SDFs the SDF workspace runs a yearly clock for algorithmic due diligence once you record the designation.
Data discovery and classification find personal data in training stores, with samples masked by the scanner before they leave the source. Vendors and processors assesses AI providers. Related template: the DPIA template for SDFs.
Can we train AI on customer data under DPDP?
Only for a purpose with a valid ground. If training was not part of the purpose you collected the data for, you need consent that names it or a s.7 legitimate use that covers it.
Is publicly available data free to use for AI training?
Section 3 excludes only personal data made publicly available by the Data Principal or under a legal obligation. Check with counsel before relying on it.
What does Rule 13 say about algorithms?
A Significant Data Fiduciary must carry out due diligence that algorithmic software it uses does not risk Data Principals' rights.
Is an AI vendor a Data Processor?
Where it processes personal data only on your behalf, yes. If it also uses the data for its own purposes, it may be a fiduciary for that use.
Your DPDP plan, starting today
Core obligations for data fiduciaries begin on 13 May 2027.
Core obligations begin