DPDP and AI: the English text never names it, and public data is not a free pass
By Abhijeet Singh · Primary sources verified by dpdprules.orgPublished · Last reviewed · Updated 16 min read
Jump to sourcesDoes the DPDP Act apply to AI training data?
The short answer
Not yet under the main processing provisions, but AI training data can fall within the Act once they commence. The English text of the Act and Rules does not use the phrases artificial intelligence, machine learning or profiling. Section 2(b) defines automated broadly, and section 2(x) includes collection, recording, organisation, storage, retrieval, use, combination and indexing in processing. Training is not itself a listed term, but those operations are processing when performed on digital personal data, whatever the system is called. Public availability alone is insufficient for section 3(c)(ii). The data must have been made or caused to be made public by the Data Principal, or made public by another person under an obligation in a law in force in India. A third party post qualifies only if one of those limbs is met. Section 17(2)(b) separately covers processing necessary for research, archiving or statistical purposes only where the data is not to be used to take a decision specific to a Data Principal and the prescribed standards are met. Rule 13(3) expressly addresses algorithmic software but applies only to a Significant Data Fiduciary. Section 9(3) separately addresses children's tracking, behavioural monitoring and targeted advertising. The Act contains no express right against automated decision making. Section 2 is in force. Sections 3, 9 and 17 and Rules 13 and 16 are in 18 month groups, computed as 13 May 2027, interpretation until officially confirmed.
The phrase "artificial intelligence" appears 0 times in the English text of the Digital Personal Data Protection Act, 2023. It appears 0 times in the English text of the Digital Personal Data Protection Rules, 2025. So do "machine learning", "profiling", "inference" and "large language".
We measured those English counts across the Gazette text of the Act, the India Code consolidated print, the Bill as passed by both Houses, the English Rules and the English section of the bilingual Rules file, with the corrigendum checked separately. The word "consent" was counted in the same pass as a control and returned 52 in the Act Gazette text, 53 in the India Code print, 53 in the Bill and 59 in each Rules text layer, which is how we know the search was reading the documents rather than failing quietly. The corrigendum is a short list of typographical corrections and contains "consent" 0 times, so it carries no control of its own. This measurement makes no claim about the Hindi wording, whose text layer is corrupt.
None of that means AI sits outside the framework. It means the analysis begins with what is being processed rather than with the name of the technique. Once the main processing provisions commence, a particular activity must still fall within the territorial scope in section 3(a) or 3(b) and remain outside section 3(c)'s exclusions.
Open the DPDP and AI diagram at full size.
The Act reaches AI without naming it
Two definitions do the work, and both are already in force.
"“automated” means any digital process capable of operating automatically in response to instructions given or otherwise for the purpose of processing data"
Section 2(x):
"“processing” in relation to personal data, means a wholly or partly automated operation or set of operations performed on digital personal data, and includes operations such as collection, recording, organisation, structuring, storage, adaptation, retrieval, use, alignment or combination, indexing, sharing, disclosure by transmission, dissemination or otherwise making available, restriction, erasure or destruction"
Read the operative list against what building a model actually involves. Collection, recording, organisation, structuring, storage, indexing, combination, retrieval and use are each named. Assembling a training corpus is collection and combination. Holding it is storage. Building a vector index over it is indexing. Querying it at inference time is retrieval and use. On our reading every one of those steps is processing under section 2(x) when the corpus contains digital personal data, and nothing in the definition turns on the technique, the model architecture or whether anyone in the room calls the system AI.
Nothing in section 2(b) turns on how the instructions were produced either. A rules engine written by hand and a model whose weights were learned from data are both digital processes capable of operating automatically in response to instructions. The distinction that matters to engineers does not appear in the definition.
This is the same structure as the cookie question. On this site's reading, scope starts with the digital personal data and the processing operation, not with the storage mechanism or the name of the technique.
The English text names algorithmic software once
There is exactly 1 provision in the English text of either instrument that expressly names algorithmic software. It is Rule 13(3):
"A Significant Data Fiduciary shall observe due diligence to verify that technical measures including algorithmic software adopted by it for hosting, display, uploading, modification, publishing, transmission, storage, updating or sharing of personal data processed by it are not likely to pose a risk to the rights of Data Principals."
Three limits are worth stating plainly, because guidance that quotes this rule often carries none of them.
It binds a Significant Data Fiduciary and nobody else. That status exists only where the Central Government notifies a Data Fiduciary or a class of them under section 10. No such notification was located in MeitY's published document library or this site's source registry as at 24 October 2026. That is a bounded finding and does not include an eGazette check. Rule 13 is in the Rules' 18 month group under Rule 1(4), commencing on a computed 13 May 2027, which is interpretation until officially confirmed.
The duty is due diligence to verify that the measures are not likely to pose a risk to the rights of Data Principals. It does not expressly prescribe a bias audit or a duty to explain automated logic; how its broader due diligence standard applies is interpretation. It is separate from Rule 13(1)'s annual data protection impact assessment and audit cycle, and it is not the framework's accuracy standard either. That sits elsewhere: section 8(3) of the Act requires a Data Fiduciary to ensure completeness, accuracy and consistency where personal data is likely to be used to make a decision that affects the Data Principal or to be disclosed to another Data Fiduciary, and item (d) of the Second Schedule to the Rules carries a reasonable efforts version of the same standard for the processing that Schedule governs.
The full duty set attaching to that status lives on the Significant Data Fiduciary obligations guide, which is where the Rule 13 cycle, the reporting of significant observations and the localisation restriction are worked through. This article does not rebuild them.
Public data is not a free pass
This is where most of the published guidance goes wrong, and the error is worth isolating because it is an error about a single word.
Section 3(c)(ii) takes certain personal data outside the Act:
"personal data that is made or caused to be made publicly available by — ( A ) the Data Principal to whom such personal data relates; or ( B ) any other person who is under an obligation under any law for the time being in force in India to make such personal data publicly available."
Read what both limbs identify. Neither describes a state of the data. Both identify who made or caused the data to be made public and the relevant legal capacity. The question the exclusion asks is not "is this data public" but "who made it public, and in which capacity".
That gives 2 routes out:
- The Data Principal made or caused her personal data to be made public.
- Some other person published it while under an obligation, under a law in force in India, to do so.
A dataset scraped from the open web does not qualify merely by being on the open web. A third party post can satisfy limb (A) if the Data Principal caused that publication, or limb (B) if the publisher acted under the required obligation in Indian law. If neither route is met, public accessibility alone does not engage the exclusion.
Limb (A) has a wider reach than a first reading suggests, because the words are "made or caused to be made" publicly available, which extends to a publication the Data Principal brought about without performing it herself. That reading, and how far it stretches over compiled and enriched records, is worked through on the startups applicability guide and is not rebuilt here.
Limb (B) carries a second constraint that is easy to read past. The obligation must arise under a law in force in India. A disclosure duty owed under a foreign statute does not satisfy the limb on the text, however genuine the obligation is in its own jurisdiction. The official materials reviewed for this article on 26 October 2026 contained no separate answer on that point, so treat this as our reading of the words rather than as settled law.
Section 3(c)(ii) is also not the only route out, and an article about training data that stops there is incomplete. Section 17(2)(b) disapplies the Act to processing "necessary for research, archiving or statistical purposes if the personal data is not to be used to take any decision specific to a Data Principal and such processing is carried on in accordance with such standards as may be prescribed", and Rule 16 supplies those standards by pointing at the Second Schedule. Two things follow on the text. The exemption is conditioned on the data not to be used to take a decision specific to a Data Principal, which is a real constraint on a model that will later be pointed at individuals. And the Second Schedule standards it imports include lawfulness, purpose limitation, data minimisation, retention limits, security safeguards and, at item (d), reasonable efforts to ensure completeness, accuracy and consistency. The official materials reviewed for this article on 26 October 2026 did not resolve whether a given training run is research within section 17(2)(b), and we do not treat it as resolved here. Section 17 is in the same 18 month commencement group as section 3, and Rule 16 is in the Rules' 18 month group.
The 2 exclusions also operate differently, and the distinction matters to the downstream question. Section 3(c)(ii) is worded as taking a category of data outside the Act, because what section 3(c) does not apply to is "personal data that is made or caused to be made publicly available". Section 17(2)(b) is worded as disapplying the Act to a category of processing, because what section 17(2) does not apply to is "the processing of personal data" that is necessary for those purposes. On our reading it follows that an argument starting from the data has to be made about who published each record, while an argument starting from the purpose has to be made about the processing and has to satisfy the decision specific condition. Neither instrument states how long that condition must keep being satisfied or what follows if it stops being satisfied, and the official materials reviewed on 26 October 2026 contained no separate answer.
What the Illustration settles, and what it does not
The Act supplies its own worked example, immediately after the exclusion:
"Illustration. X, an individual, while blogging her views, has publicly made available her personal data on social media. In such case, the provisions of this Act shall not apply."
The Illustration is statutory text, not commentary, and it is the clearest thing in the framework on this question. It also has a narrow shape that is worth noticing: the person publishing and the person the data is about are the same person, and the consequence stated is about that data.
What it settles is limb (A) in its simplest form. X blogs her own views, and the Act does not apply to the personal data she thereby made available.
What it does not address is the question an AI training set actually raises, which is what a different person may then do with that data, at what scale, and for what purpose. The Illustration describes the publication. It says nothing about downstream reuse, and neither does the rest of section 3(c)(ii). On our reading the exclusion, once it bites, removes that data from the Act's application rather than licensing a particular use of it, which means the interesting arguments run to whether the exclusion was ever engaged for a given record rather than to what may be done afterwards. The official materials reviewed on 26 October 2026 contained no separate answer on downstream reuse, so claims that it is settled in either direction go beyond those materials.
On our reading the practical consequence is that the test runs per record, not per dataset. A corpus of 10 million records is not inside or outside the Act as a block. Each record has a publication path, and limb (A) or limb (B) either reaches that path or it does not. A dataset assembled from many sources may contain a mixture. The Act does not state this record level reading or allocate an evidentiary burden, and the official materials reviewed on 26 October 2026 contained no separate answer on either point.
There is no counterpart to a right against automated decisions
The word "profiling" appears 0 times in the English text of the Act and 0 times in the English text of the Rules. The instruments contain no express provision giving a Data Principal a right not to be subject to a decision based solely on automated processing, and no express requirement to explain automated logic.
Chapter III names the right to access information about personal data in section 11, the right to correction and erasure in section 12, the right of grievance redressal in section 13 and the right to nominate in section 14. Outside that chapter, section 6(4) expressly gives a right to withdraw consent and section 29(1) expressly gives a right to appeal to the Appellate Tribunal. None of those named rights is an express automated decision right, and none turns on how a decision was reached.
A reader arriving from a GDPR background is usually looking for Article 22 at this point. That Article is the right place to look in the European instrument, and it should be read there rather than through this page. We do not restate GDPR requirements as claims to rely on. The claim being made here is about the Indian text, and it is a negative: the counterpart is absent.
The single provision of the Act that addresses behavioural monitoring at all is section 9(3), which is confined to children:
"A Data Fiduciary shall not undertake tracking or behavioural monitoring of children or targeted advertising directed at children."
That is a prohibition rather than a consent question, so consent alone does not switch it off. Routes that can lift or disapply it include section 9(4), implemented by Rule 12 and the Fourth Schedule for specified classes and purposes subject to their conditions, section 9(5), and the applicable parts of section 17. It concerns children rather than automated processing generally. The full children analysis belongs to the children's data guide and is not developed here.
A research trap worth knowing
Search the English text of both instruments for the word "training" and you will get exactly 1 hit, and it is in the Rules:
"the Ministry of Personnel, Public Grievances and Pensions, Department of Personnel and Training"
It appears in the Sixth Schedule, printed under the words "See rule 21(2)", which sets the terms and conditions of appointment and service of officers and employees of the Board. It has nothing to do with training a model, and an argument built on that hit would be an argument about civil service staffing.
The same instrument holds a second near miss, and this one is closer to home. Search the English stem "profil" and the Rules return exactly 1 hit, in the definition of "user account", which "includes any profiles, pages, handles, email address, mobile number and other similar presences". That is a noun about a social media presence. The word "profiling", the thing a reader searching that stem is actually looking for, still appears 0 times in the English text of either instrument.
This matters because these are the second and third traps of this shape we have documented. The word "succession" also appears exactly once in the Act, in section 18(2), where it gives the Board perpetual succession as a body corporate, and a search that finds it and builds an inheritance argument has cited the wrong provision entirely. A single hit on a promising keyword is a reason to read the surrounding clause, not a finding.
What this means in practice
None of the following is legal advice, and none of it is a requirement stated by the framework. It is what follows from the text above.
Record provenance at collection time, not later. Because the section 3(c)(ii) test turns on the publication path, a dataset whose provenance was never recorded may be difficult to assess against the exclusion. Capturing provenance at collection time is usually easier than reconstructing it later.
Do not treat "publicly accessible" and "publicly available by the Data Principal" as the same category. They are different tests and only the second one is in the Act.
Check whether the Significant Data Fiduciary question is live for you. If it is, Rule 13(3) is the provision that reaches your algorithmic software, and the wider duty set comes with it.
Children are a separate question with a different answer. Section 9(3) is a prohibition, not a consent gate, and sections 9(4), 9(5) and 17 supply distinct routes that may lift or disapply it.
Watch official developments, not commentary alone. Developments that could change this analysis include a commencement notification, a section 10 notification, an amendment, new rules, a corrigendum, a court ruling or official guidance. Trade body positions and consultation papers do not themselves change the instruments.
When this starts to apply
Section 2 is already in force, which is why the definitions above can be stated in the present tense.
Section 3 is not. It sits in the group that paragraph (c) of commencement notification G.S.R. 843(E) places 18 months after publication, which this site computes as 13 May 2027 and labels interpretation until officially confirmed, because the notification states a period rather than a date and does not say how the period is counted. Section 9 is in the same paragraph, and Rule 1(4) of the Rules puts Rule 13 in the Rules' own 18 month group. Rule 1(3) is the one year group and holds only Rule 4.
There is a further wrinkle on the start date that this site tracks separately. The notification's masthead gives 13 November 2025, the eGazette file code embeds 14112025, and the notification does not state how the period is counted. Together those uncertainties could move the computed date by up to 2 days, to 15 May 2027 at the latest. The issue is explained in its own article.
So the applicability analysis in this article, including the exclusion everyone is arguing about, describes a provision that does not yet bite. That is not a reason to defer the provenance work, because the corpus being assembled today may still be in place when section 3 commences, and missing provenance can be difficult or impossible to reconstruct later. It is a reason to be precise about tense, and to be sceptical of any guidance telling you that you are in breach today.
What Rule 13 adds for a Significant Data Fiduciary →The section 3 applicability test →Why the duty follows the data, not the technology →Section 3, official text with sources →
Artificial intelligenceTraining dataSection 3Myth correction