AI & personal dataUpdated 22 November 20269 min read

AI Personalisation and Recommendation Engines Under DPDP

How recommendation and personalisation systems must handle consent, purpose limits, withdrawal, children's data and erasure under the DPDP Act and Rules from 13 May 2027.

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

Legal basis
Consent (s. 6) for most personalisation
Notice requirement
Rule 3 itemised, purpose-specific notice
Children (under 18)
No tracking, behavioural monitoring or targeted ads (s. 9(3))
Children's obligations penalty
Up to ₹200 crore
SDF algorithmic check
Rule 13(3), once notified
Rules apply from
13 May 2027

Quick answer

Personalisation under the DPDP Act generally needs specific, informed consent for that purpose, with an itemised Rule 3 notice and withdrawal as easy as opting in. For users under 18, section 9(3) bars tracking, behavioural monitoring and targeted advertising directed at children. Engineering must make consent state a feature-store input, and erase user embeddings when consent is withdrawn.

  • Behavioural personalisation is usually a separate purpose from delivering the core service, so it needs its own consent.
  • Rule 3 notices must itemise the data and describe the purpose; 'to improve your experience' is not specific enough.
  • Section 9(3) prohibits tracking, behavioural monitoring and targeted advertising directed at children, subject to Fourth Schedule exemptions.
  • Withdrawing consent must stop personalisation within a reasonable time and remove the user's profile features and embeddings.
  • Significant Data Fiduciaries, once notified, must verify their algorithmic software does not pose risks to Data Principals' rights.

Who this is for: Recommendation and ranking engineers, Product managers for personalisation, Growth and marketing teams, E-commerce, media and OTT platforms, Data Protection Officers

The DPDP Act allows processing on two grounds: consent under section 6, or a legitimate use under section 7. The legitimate use most relevant to commercial apps is section 7(a) — personal data voluntarily provided for a specified purpose where the person has not indicated she does not consent.

Section 7(a) fits narrow cases: using a delivery address to deliver an order, or a typed search query to return results. It is a poor fit for behavioural personalisation built from clickstreams, dwell time, purchase history and inferred interests across sessions, because that data is collected passively and used for a purpose the user did not specifically provide it for.

For most recommendation engines, treat consent as the basis and design for it:

  • Consent must be free, specific, informed, unconditional and unambiguous, given by clear affirmative action (section 6(1)).
  • It must be limited to data necessary for the specified purpose.
  • It cannot be bundled as a condition for a service that does not need personalisation.
  • Withdrawal must be as easy as giving consent (section 6(4)), and processing must stop within a reasonable time after withdrawal.

The Rules that operationalise notices and rights apply from 13 May 2027.

Writing a Rule 3 notice for personalisation

Rule 3 requires a notice that stands on its own, uses clear and plain language and includes, at minimum, an itemised description of the personal data and the specified purpose with a specific description of the goods, services or uses enabled.

For personalisation, itemise the signals you actually use:

SignalExample notice wording
Browsing and viewing historyProducts and pages you view, and how long you view them
Purchase historyItems you buy and return
Search queriesTerms you search for in the app
Device and coarse locationDevice type and city-level location
Inferred interestsInterest categories we derive from the above

Then state the use specifically: "to rank products and content on your home feed and send you recommendations by email", rather than "to improve your experience".

Rule 3(c) also requires a link to withdraw consent with ease comparable to giving it, to exercise rights, and to complain to the Board. A toggle in settings that mirrors the opt-in screen is the cleanest implementation. See Rule 3 notice requirements and consent and withdrawal UX.

Consent compliance fails in recommendation systems when consent lives in a CMP database that the ML pipeline never reads. Wire it in end to end:

  1. Consent service. Store consent per user, per purpose (personalisation, email_recs, ads), with version of the notice shown and timestamp.
  2. Event gating. At the event collector, drop or tag behavioural events from users without personalisation consent. Do not collect first and filter later.
  3. Feature store filter. Join consent state on every feature materialisation job; users without consent get only non-personal or session-only features.
  4. Serving fallback. At inference, the ranking service checks consent and falls back to popularity or contextual ranking (current page, current query) for non-consenting users.
  5. Training filter. Exclude non-consenting users' interactions from training data for personalisation models.
  6. Withdrawal propagation. A withdrawal event triggers deletion of the user's profile features, user embeddings and cached recommendation lists, and removes them from the next training snapshot.

Contextual recommendations based only on what the user is doing right now carry much less risk than cross-session profiling, and are a good default for logged-out or non-consenting users.

Children: the section 9(3) wall

Section 9(3) of the DPDP Act says a Data Fiduciary shall not undertake tracking or behavioural monitoring of children or targeted advertising directed at children. A child is anyone under 18. Breach of children's obligations carries a penalty of up to ₹200 crore.

For recommendation engines this is close to a hard stop. A feed ranked on a child's viewing history is behavioural monitoring by any reasonable reading.

What remains possible:

  • Contextual ranking — recommendations based on the current item or search, without a persistent behavioural profile.
  • Editorial or age-appropriate curation.
  • Fourth Schedule Part B item 5 — processing to ensure information, services or advertisements likely to cause a detrimental effect on a child's well-being are not accessible to her, restricted to the extent necessary.
  • Fourth Schedule Part B item 6 — processing to confirm a user is not a child, restricted to that purpose.
  • Educational institutions (Part A item 3) — tracking and behavioural monitoring restricted to educational activities or children's safety.

Section 9(2) — no processing likely to cause a detrimental effect on a child's well-being — is not exempted for anyone. Age assurance and parental consent under Rule 10 are covered in AI products for children.

Accuracy, profiling decisions and SDF duties

Section 8(3) of the Act requires a fiduciary to ensure completeness, accuracy and consistency of personal data where it is likely to be used to make a decision that affects the Data Principal, or disclosed to another fiduciary. Most product recommendations are not decisions of that kind. But personalisation that drives pricing, credit offers, eligibility or which loan products a user sees moves closer to a decision, and the accuracy duty bites.

Practical controls:

  • Let users see and correct inferred interests where they materially shape what they are offered.
  • Document which models feed decisions versus suggestions.
  • Monitor for proxy discrimination in offers — the MeitY AI Governance Guidelines list fairness and equity among their seven sutras.

If your platform is notified as a Significant Data Fiduciary, Rule 13(3) adds a specific duty: observe due diligence to verify that technical measures including algorithmic software adopted for hosting, display, uploading, modification, publishing, transmission, storage, updating or sharing of personal data are not likely to pose a risk to the rights of Data Principals. Recommendation and ranking systems are squarely within "display" and "sharing". Rule 13(1) also requires an annual DPIA and audit. See algorithmic due diligence.

Common mistakes in personalisation compliance

  • Bundling. Making personalisation consent a condition of account creation when the service works without it.
  • Collect-then-filter. Logging every event and filtering non-consenting users downstream; the collection itself is processing.
  • Stale embeddings. Withdrawing consent stops new events but the user's embedding keeps driving the feed.
  • Third-party SDKs. Analytics and ad SDKs that build profiles independently of your consent state.
  • Age-blind models. No age signal at all, so children are profiled by default.
  • Notice drift. Adding new signals (location, contacts) to the model without updating the itemised notice.
  • Ignoring retention. Large e-commerce and social media platforms (at least 2 crore registered users in India) face Third Schedule erasure after three years of inactivity, so dormant profiles must expire too. See Rule 8 retention.

Fixing these before 13 May 2027 is mostly plumbing: a consent service the ML stack can read, and withdrawal events that reach every store holding a user vector.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Separate personalisation consent from service consent

    Ask for personalisation as its own purpose and do not make it a condition of using the core service.

    DPDP Act s. 6(1)

  2. Itemise personalisation signals in the notice

    List browsing, purchase, search, location and inferred interests, and describe the specific use.

    Rule 3(b)

  3. Gate event collection on consent

    Drop behavioural events from non-consenting users at the collector, not downstream.

    DPDP Act s. 6(1)

  4. Join consent state in feature pipelines

    Filter feature materialisation and training data by current consent status.

    DPDP Act s. 6(1)

  5. Provide a contextual fallback ranking

    Serve non-personal recommendations to non-consenting users and children.

    DPDP Act s. 9(3)

  6. Propagate withdrawal to embeddings and caches

    Delete profile features, user vectors and cached lists when consent is withdrawn.

    DPDP Act s. 6(6); s. 8(7)

  7. Disable behavioural profiling for under-18s

    Use age assurance and route children to contextual or curated feeds.

    DPDP Act s. 9(3)

  8. Audit third-party SDKs

    Ensure analytics and ad SDKs respect your consent and children's flags.

    DPDP Act s. 8(1)

Key dates

  1. DPDP Rules notified

  2. Rule 3 notices, Rule 10 parental consent and Rule 12 exemptions apply

Answers

Frequently asked questions

Do I need consent for product recommendations under DPDP?

For behavioural recommendations built from browsing and purchase history across sessions, consent is the safer basis. The section 7(a) legitimate use covers data a user voluntarily provides for a specified purpose, which suits narrow uses like returning search results, but not passive profiling for a purpose the user never specified.

Can I personalise content for users under 18 in India?

Not through tracking or behavioural monitoring. Section 9(3) of the DPDP Act prohibits tracking, behavioural monitoring and targeted advertising directed at children, meaning anyone under 18. Contextual recommendations based on the current item, curated feeds and processing to keep harmful content away from children under the Fourth Schedule remain possible.

What happens to a user's recommendation profile when they withdraw consent?

Processing for that purpose must stop within a reasonable time, and the personal data should be erased unless retention is required by law. In practice, delete the user's profile features, embeddings and cached recommendation lists, exclude them from future training, and switch them to non-personal recommendations. Rule 8(3) one-year log retention still applies to processing logs.

Is a recommendation algorithm subject to audit under DPDP?

Only for Significant Data Fiduciaries, once notified. Rule 13(1) requires them to conduct a DPIA and audit every twelve months, and Rule 13(3) requires due diligence that algorithmic software used for displaying or sharing personal data does not pose a risk to Data Principals' rights. No SDF notification is assumed here.

Is 'to improve your experience' a valid purpose in a DPDP notice?

It is unlikely to meet Rule 3, which requires an itemised description of the personal data and the specified purpose with a specific description of the goods, services or uses enabled. Name the signals and the use, for example ranking your home feed and sending recommendation emails.

Does targeted advertising need separate consent from personalisation?

Treat it as a separate purpose. Consent must be specific to each purpose, so ranking a feed and using the same profile to target ads, including through third-party ad networks, should be presented and recorded separately, with independent withdrawal. For children, targeted advertising is prohibited outright by section 9(3).

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Sections 5, 6 and 9 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 3 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 12 and Fourth Schedule (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 13(3) (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.