AI & personal dataUpdated 22 November 20269 min read

Training LLMs on Personal Data Under the DPDP Act

Training or fine-tuning a model on personal data is processing under the DPDP Act. Here is how lawful basis, public data, research, erasure and memorisation work.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Training is processing?
Yes — automated use of digital personal data
Usual basis
Specific consent (s.6)
Public data carve-out
Section 3(c)(ii), narrow
Research exemption
Section 17(2)(b) + Rule 16
Security baseline
Rule 6
Obligations from
13 May 2027

Quick answer

Training or fine-tuning an LLM on personal data is processing under the DPDP Act and needs a lawful basis: usually specific consent, since model improvement is a separate purpose from serving the user. Data the Data Principal herself made public is outside the Act (s.3(c)(ii)), but scraped data published by others is not. From 13 May 2027, erasure, security and one-year log duties apply to training pipelines.

  • Training is a distinct purpose; data collected to provide a service does not automatically cover model training.
  • The public-data exclusion in section 3(c)(ii) covers data the person herself made public, not everything online.
  • The research exemption needs Second Schedule standards and no decisions specific to individuals.
  • Scrub, pseudonymise and deduplicate personal data before training to reduce memorisation and erasure risk.
  • Keep dataset lineage so you can show which records trained which model version.
  • Withdrawal of consent stops future training use; plan retraining cycles to honour it.

Who this is for: ML and data science teams, AI startups building or fine-tuning models, Enterprises fine-tuning on customer data, Data Protection Officers, Model governance leads

Is training an LLM on personal data allowed under DPDP?

Yes, if you have a lawful basis and meet the fiduciary obligations. Training, fine-tuning, building evaluation sets and running reinforcement learning from human feedback on conversation logs are all processing — automated operations such as collection, storage, adaptation, combination and use of digital personal data. Whoever decides to train the model on that data, and how, is the Data Fiduciary for it.

The lawful basis is where most teams struggle:

Data sourceLikely basisNotes
Customers' chats or tickets, to improve your product modelSpecific consentA separate purpose from answering the query
Employees' work data for an internal modelSection 7(i), if genuinely for employment purposesProportionality matters
Data the Data Principal herself made publicOutside the Act (s.3(c)(ii))Narrow; see below
Licensed datasets from third partiesThe supplier's basis must cover your useDiligence the supplier's consent records
Research with no individual decisionsPossibly s.17(2)(b) with Rule 16 standardsNot a commercial shortcut

Consent must be free, specific, informed, unconditional and unambiguous, and limited to data necessary for the purpose. Bundling training consent into the terms of service for a product that does not need it for delivery is unlikely to meet that test.

Scraped and publicly available data

Section 3(c)(ii) says the Act does not apply to personal data made publicly available by the Data Principal to whom it relates, or by any other person under a legal obligation to make it public. The Act's own illustration involves an individual who shares her views on social media, publicly.

This is narrower than "anything on the open web":

  • Covered: a person's own public profile, posts she chose to publish publicly, data a company is legally required to publish (some regulatory disclosures).
  • Not covered: someone else's post about her, leaked databases, data published by aggregators without a legal obligation, data behind logins, and her data that was public once but is now private.
  • Unclear: data made public by her but later scraped and combined with other sources into a profile; the combination may be new processing. Get legal advice for web-scale crawls.

Engineering implications for crawled corpora:

  1. Record source URL, crawl date and the basis you relied on for each document.
  2. Filter out pages likely to contain third-party personal data (people-search sites, leaked dumps).
  3. Run PII detection and remove direct identifiers where they are not needed for the model's purpose.
  4. Honour removal requests even when you believe the exemption applies — the cost is low and the risk is high.

See publicly available data.

Does the research exemption help?

Section 17(2)(b) disapplies the Act to processing necessary for research, archiving or statistical purposes, if the personal data is not used to take any decision specific to a Data Principal and the processing follows the prescribed standards. Rule 16 points to the Second Schedule, which requires:

  • lawful processing;
  • processing only for the stated research purposes;
  • data limited to what is necessary;
  • reasonable efforts at completeness, accuracy and consistency;
  • retention only as long as needed;
  • reasonable security safeguards, including for processors;
  • accountability of whoever determines purpose and means.

A university lab training a language model to study linguistic patterns may fit. A company training a model that will score, recommend or respond to individual customers fails the "no decision specific to a Data Principal" condition once the model is deployed for that purpose. Treat 17(2)(b) as a narrow route for genuine research, and document how you meet each Second Schedule item.

See the research and statistical exemption.

Controls for a DPDP-ready training pipeline

1. Gate the input

  • Build training exports from a query that filters on training_consent = true (or a documented alternative basis) at row level.
  • Exclude children's data unless you have verifiable parental consent and a clear case under section 9(2) that training is not detrimental.
  • Exclude records under active erasure or withdrawal.

2. Minimise

  • Run PII detection over text, including free-text fields, file uploads and transcripts.
  • Replace direct identifiers with type tokens (<NAME>, <PHONE>) unless the model genuinely needs them.
  • Deduplicate aggressively; repeated sequences are more likely to be memorised.

3. Record lineage

  • Version every dataset (dataset_id, hash, source query, date) and every model (model_id, datasets used).
  • Store the mapping from data_principal_id to datasets, so erasure can find affected versions.

4. Secure the stores

  • Encrypt training buckets, restrict access to named roles, and log reads — Rule 6(1)(a)–(c).
  • Keep processing logs for at least one year (Rule 8(3)).

5. Test the model

  • Run memorisation probes and canary checks for known personal strings.
  • Add output filters that block emission of identifiers such as phone numbers and account numbers.

Erasure, withdrawal and model weights

Section 8(7) requires erasure when consent is withdrawn or the purpose is no longer served, and section 12 gives a right to erasure. For training data, handle three layers:

  1. Datasets. Remove the person's records from source tables and from all stored training and evaluation datasets. Mark old dataset versions as superseded.
  2. Future training. Withdrawal of consent must stop further processing within a reasonable time, and processors must stop too. Excluded records must not reappear in the next training run.
  3. Trained weights. Whether a trained model "contains" personal data in a legal sense is not settled under the DPDP Act. Practical approaches, from cheapest to strongest:
    • output filtering and refusal rules for the person's identifiers;
    • fine-tuning adapters trained only on consented data, which can be dropped and retrained;
    • scheduled retraining from cleaned datasets;
    • machine-unlearning techniques, still maturing.

Document which approach you take and why. Rule 8(3) also requires personal data and processing logs to be kept for at least one year from processing; keep any such copy in a restricted archive, not in active training stores. See erasure requests and LLMs.

Training with vendors and across borders

Fine-tuning through a model provider's API, or training on rented GPUs abroad, adds a processor and a transfer:

  • Processor contract (section 8(2), Rule 6(1)(f)): confirm the vendor uses your training data only for your model, keeps it isolated, deletes it on instruction, and will not use your fine-tuned model or data to improve its own models.
  • Cross-border transfer: the DPDP Act permits transfers except to countries restricted by Government notification (s.16), subject to any Rule 15 requirements. Sector rules — RBI payment data storage, for instance — may still require data to stay in India.
  • SDFs: once notified, an SDF must ensure Government-specified data and its traffic data are not transferred outside India (Rule 13(4)), and must verify algorithmic software is not likely to risk Data Principals' rights (Rule 13(3)).
  • EU obligations: if you place a general-purpose AI model on the EU market, EU AI Act GPAI duties apply, including a public summary of training content and a copyright policy. See GPAI obligations.

Synthetic data can reduce reliance on real personal data, but generating it from personal data is itself processing. See synthetic data.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Separate training from service purposes

    Create distinct purpose IDs for model training and evaluation, and a specific consent where no legitimate use applies.

    Section 6

  2. Filter training exports by basis

    Build dataset extracts only from rows with a valid training basis, excluding withdrawn and erased records.

    Sections 6, 8(7)

  3. Exclude or specially handle children's data

    Do not train on children's data without verifiable parental consent and a section 9(2) detriment review.

    Section 9; Rule 10

  4. Document public-data reliance

    Record source, crawl date and the section 3(c)(ii) reasoning for scraped content; filter third-party personal data.

    Section 3(c)(ii)

  5. Scrub and deduplicate before training

    Detect and tokenise direct identifiers and deduplicate repeated sequences to reduce memorisation.

    Rule 6(1)(a)

  6. Maintain dataset and model lineage

    Version datasets and models and map Data Principal IDs to dataset versions for erasure.

    Section 8(7)

  7. Run memorisation and leakage tests

    Probe models for verbatim personal strings and add output filters for identifiers.

    Rule 6(1)(g)

  8. Put training vendors under contract

    Restrict vendor use of training data and fine-tuned models, and confirm deletion and region.

    Section 8(2); Rule 6(1)(f)

Key dates

  1. DPDP Rules notified, including Rule 16 research standards

  2. Consent, security, erasure and Rule 16 apply to training pipelines

Answers

Frequently asked questions

Can I train an AI model on customer data in India?

Yes, with a lawful basis — usually specific consent, because training is a separate purpose from delivering the service — and with DPDP safeguards in place. Filter training data by consent, minimise identifiers, secure the datasets and build a path to honour withdrawal and erasure. Obligations apply from 13 May 2027.

Is scraping public data for AI training legal under DPDP?

Partly. Section 3(c)(ii) excludes personal data the Data Principal herself made publicly available, or that someone published under a legal obligation. Data published about her by others, leaked data or private data is not excluded. Document sources and filter third-party personal data from crawls.

Does the research exemption cover AI model training?

Only in narrow cases. Section 17(2)(b) applies to processing necessary for research, archiving or statistics, where data is not used for decisions specific to an individual and the Second Schedule standards in Rule 16 are met. A commercial model that serves individual customers usually fails that test.

What happens if a user withdraws consent after training?

You must stop further processing within a reasonable time and exclude her data from future training runs and datasets. Whether existing weights must be retrained is unsettled; common approaches are output filters, removable adapters and scheduled retraining from cleaned data. Document your approach.

Do fine-tuned models contain personal data?

They can memorise and reproduce personal strings, especially repeated ones. Whether weights legally contain personal data under the DPDP Act is unsettled. Reduce the risk by removing identifiers and duplicates before training, testing for memorisation and filtering outputs.

Can I fine-tune with a foreign model provider?

Yes. The DPDP Act allows cross-border transfers except to countries restricted by notification, subject to Rule 15. Treat the provider as your Data Processor with a contract limiting use of your data, and check sector rules such as RBI payment data storage that may require data to stay in India.

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Sections 3(c)(ii), 6, 8(7) (official text)
  • Digital Personal Data Protection Act, 2023 — Section 17(2)(b) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 16 and Second Schedule (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 6 (official text)

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.