AI & personal dataUpdated 22 November 20268 min read

Synthetic Data and Anonymisation Under the DPDP Act

When synthetic or anonymised data falls outside the DPDP Act, when it does not, and how to generate, test and govern it for AI training and evaluation.

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

Test for coverage
Whether the individual is identifiable
Anonymisation standard in DPDP
None defined
Generation from real data
Is processing; needs a lawful basis
Research exemption
Rule 16 with Second Schedule standards
Rules apply from
13 May 2027

Quick answer

The DPDP Act applies to data about an individual who is identifiable by or in relation to that data. Truly synthetic or anonymised data that cannot be linked to anyone falls outside it, but generating it from real records is itself processing that needs a lawful basis and Rule 6 safeguards. The Act has no anonymisation standard, so test re-identification risk before treating any dataset as non-personal.

  • DPDP turns on identifiability: non-identifiable synthetic data is outside the Act, re-identifiable data is not.
  • Creating synthetic data from production records is processing of personal data and needs a lawful basis.
  • The Act and Rules define no anonymisation test, so document your own method and residual-risk evidence.
  • The research and statistics exemption in Rule 16 applies only when Second Schedule standards are followed.
  • Synthetic training data is different from 'synthetically generated information' under the IT Rules amendments, which concern deepfakes and labelling.

Who this is for: ML and data science teams, Data platform engineers, QA teams using test data, Research and analytics groups, Data Protection Officers

Is synthetic data personal data under the DPDP Act?

The DPDP Act defines personal data as any data about an individual who is identifiable by or in relation to such data, and applies to digital personal data processed in India (and outside India where connected to offering goods or services to people in India).

So the answer depends on identifiability, not on the label "synthetic":

  • Fully synthetic data generated from statistical properties, where no record corresponds to or can be linked with a real person, is not personal data and sits outside the Act.
  • Partially synthetic data — real records with some fields replaced — usually remains personal data, because the untouched fields can identify people.
  • Pseudonymised data with a key or lookup you control remains personal data; it is a security safeguard under Rule 6(1)(a), not an exit.
  • Model-generated text that reproduces real individuals' details because a generator memorised them is personal data, whatever the pipeline was called.

The Act and the 2025 Rules contain no anonymisation definition or test. That means the burden is practical: you need evidence that a dataset cannot reasonably be linked back to individuals, using the information available to you and to plausible attackers.

Generating synthetic data is still processing

Even if the output is non-personal, the input is not. Training a generator — a GAN, a diffusion model, a tabular synthesiser or an LLM prompted with real records — processes personal data. That step needs:

  • A lawful basis. Consent covering the purpose, or a legitimate use under section 7. Product analytics or model development is rarely within the purpose a customer originally consented to, so check your notices before using production data.
  • Rule 6 safeguards. The generation environment holds real data: encryption, access control, logging and the one-year log retention under Rule 6(1)(e) apply.
  • Minimisation. Feed the generator only the columns it needs. Drop direct identifiers before training it.
  • Erasure handling. If a person's data was in the generator's training set and she withdraws consent, the generator itself may carry her data; plan to regenerate from cleaned inputs.

Where synthetic data is produced strictly for research, archiving or statistical purposes, Rule 16 says the Act does not apply to processing necessary for those purposes if it follows the Second Schedule standards: lawful processing, purpose limitation, necessity, accuracy efforts, retention limits, security safeguards and accountability. Rule 16 is not a general exemption for commercial model training. See the research and statistics exemption.

Testing re-identification risk before release

Because DPDP gives no anonymisation test, borrow engineering practice and document it.

TestWhat it catchesHow
Exact and near-duplicate matchGenerator copying real rowsDistance to closest real record for every synthetic row
Membership inferenceWhether a specific person was in training dataAttack model comparing in-set vs out-of-set records
Attribute inferenceInferring a sensitive field from quasi-identifiersTrain a predictor on synthetic, test on real holdout
Rare-combination scanUnique pin code + age + job combinationsk-anonymity style counts on quasi-identifiers
Free-text leakageNames, phones, emails in generated textPII detectors on all text outputs

Practical thresholds are a judgement, but record them: for example, no synthetic row within a set distance of a real row, rare combinations suppressed below a minimum count, zero detector hits on direct identifiers.

Differentially private generation gives a mathematical bound on what any one record can contribute, at a utility cost. It is the strongest technical argument that the output is non-personal, provided the privacy budget is documented and reasonable.

Re-run tests whenever the generator, its training data or its parameters change.

Where synthetic data helps AI teams most

Synthetic data reduces DPDP exposure most where real data adds risk without adding much value.

  • Eval and golden sets. Hand-written or generated conversations replace real chats in eval spreadsheets — one of the most common unmanaged copies of customer data.
  • Test fixtures and staging. Seed staging databases and vector stores with synthetic customers so engineers never need production dumps.
  • Red-teaming and prompt-injection tests. Synthetic personas with fake PAN-format numbers and addresses let you test masking and leakage without real data.
  • Class balancing. Augment rare fraud or failure cases, after the generator itself is trained under safeguards.
  • Vendor demos and benchmarks. Share synthetic samples with vendors during evaluation instead of real records.

Where synthetic data does not replace real data — fine-tuning on genuine customer language, for instance — combine it with masking and lineage controls described in training LLMs on personal data and masking and tokenisation.

Synthetic data is not 'synthetically generated information'

The term "synthetic" appears in two unrelated Indian regulatory contexts.

  • Synthetic training data — the subject of this page — is a data-engineering technique governed by DPDP identifiability principles.
  • Synthetically generated information is the term MeitY used in its draft amendments to the IT (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, published on 22 October 2025, for AI-generated or altered content such as deepfakes. The draft proposed visible labels and metadata; final amendments were notified in February 2026 and changed several draft specifics, so check the notified text. See India's deepfake and synthetic content rules.

The two can overlap. A generator that produces realistic images of real people, or synthetic voice clips resembling a real customer, creates both a labelling question for platforms under the IT Rules and a personal data question under DPDP, because the output relates to an identifiable person.

Keep the vocabulary separate in internal policies so engineers do not assume a "synthetic data" approval covers generated media of real people.

Governance: a synthetic data policy in six lines

A short written policy prevents most problems:

  1. Approved generators and environments. Generators run only in a controlled environment with Rule 6 safeguards.
  2. Input approval. A named owner approves which real datasets and columns may train a generator, and on what lawful basis.
  3. Release gate. No synthetic dataset leaves the generation environment without passing the documented re-identification tests.
  4. Labelling. Every dataset is tagged synthetic, partially_synthetic or real_masked, with generator version and test results.
  5. Re-test on change. New generator version or new input data means new tests.
  6. Erasure path. If an input record must be erased, decide whether the generator and its outputs must be regenerated, and record the decision.

Before 13 May 2027, inventory where real personal data sits in non-production environments — staging, eval sets, notebooks, vendor demos — and replace the easy cases with synthetic data first. It shrinks the scope of erasure, breach and retention obligations that apply from that date.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Classify each dataset by identifiability

    Tag datasets as real, masked, partially synthetic or fully synthetic and treat only the last as potentially non-personal.

    DPDP Act s. 2

  2. Confirm a lawful basis for generator training

    Check consent scope or legitimate use before feeding production data into a generator.

    DPDP Act ss. 6–7

  3. Secure the generation environment

    Apply encryption, access control and logging to environments holding real input data.

    Rule 6(1)

  4. Run re-identification tests before release

    Test nearest-record distance, membership inference, rare combinations and free-text leakage.

    DPDP Act s. 8(5)

  5. Document thresholds and results

    Record test parameters, outcomes and generator versions alongside each released dataset.

    DPDP Act s. 8(4)

  6. Replace real data in staging and eval sets

    Use synthetic fixtures for staging databases, vector stores and eval spreadsheets.

    DPDP Act s. 8(7)

  7. Use Rule 16 only within its limits

    Rely on the research exemption only for genuine research, archiving or statistics following Second Schedule standards.

    Rule 16

Key dates

  1. Draft IT Rules amendment on synthetically generated information published

  2. DPDP Rules notified

  3. Rules 6 and 16 apply

Answers

Frequently asked questions

Does the DPDP Act apply to synthetic data?

Only if the synthetic data relates to an identifiable individual. Fully synthetic records that cannot be linked to any real person are outside the Act. Partially synthetic, pseudonymised or memorised outputs that can be linked back to someone remain personal data. Generating synthetic data from real records is itself processing covered by the Act.

Does India's DPDP Act define anonymisation?

No. Neither the DPDP Act, 2023 nor the DPDP Rules, 2025 define anonymisation or set a test for it. The Act turns on whether an individual is identifiable, so organisations must document their own method and evidence, such as re-identification testing or differential privacy, to support treating a dataset as non-personal.

Can I use production data to train a synthetic data generator?

Yes, if you have a lawful basis for that use and apply Rule 6 safeguards to the generation environment. Check whether your notices and consents cover model development or analytics, drop direct identifiers before training the generator, and keep the real inputs inside a controlled environment.

Is synthetic data covered by the research exemption in Rule 16?

Only when the processing is necessary for research, archiving or statistical purposes and follows the Second Schedule standards, including purpose limitation, necessity, retention limits, safeguards and accountability. Rule 16 is not a general exemption for commercial AI training, and it applies from 13 May 2027.

Are deepfakes the same as synthetic data under Indian law?

No. Synthetic training data is a privacy-engineering technique judged under DPDP identifiability. 'Synthetically generated information' is the IT Rules term for AI-generated or altered content, covered by MeitY's amendments notified in February 2026. Generated media of real people can raise both issues at once.

How do I prove a synthetic dataset is safe to share?

Run and document re-identification tests: distance from each synthetic row to the closest real record, membership-inference attacks, rare quasi-identifier combinations and PII scans on text. Record thresholds, results and generator version. Differentially private generation with a documented privacy budget gives the strongest argument.

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Sections 2 and 3 (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(1)(a) (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.