DPDP Act and AI Training Data: What Data Buyers and Vendors Must Do in 2026–27

Last updat­ed: Octo­ber 2026

The DPDP Act AI train­ing data ques­tion has moved from legal the­o­ry to pro­cure­ment check­lists. Indi­a’s Dig­i­tal Per­son­al Data Pro­tec­tion Act, 2023 and the DPDP Rules, 2025 now gov­ern how teams col­lect, label, store and delete per­son­al data for mod­el train­ing. Most core duties start in May 2027, and datasets bought this quar­ter will still be in use then.

This guide is for AI teams that buy speech, video or text data, and for the ven­dors that col­lect and anno­tate it. It cov­ers scope, the ver­i­fied rules time­line, con­sent and proces­sor duties. It ends with an 8‑point buy­er check­list for your next ven­dor review. If you track India data pro­tec­tion and AI pol­i­cy, the time­line sec­tion is the one to book­mark.

This arti­cle is gen­er­al infor­ma­tion, not legal advice. Check the Act, the Rules and your own coun­sel before you act on it.

Sourc­ing Indi­an data for a 2027 mod­el? Ask us for our con­sent & data-han­dling pack before you sign a ven­dor.

What does the DPDP Act cover for AI training data?

The DPDP Act cov­ers the pro­cess­ing of dig­i­tal per­son­al data in India. Per­son­al data means any data about an indi­vid­ual who is iden­ti­fi­able by or in rela­tion to that data. Pro­cess­ing includes col­lec­tion, record­ing, struc­tur­ing, stor­age, use, shar­ing and era­sure. That maps direct­ly onto col­lec­tion, anno­ta­tion and mod­el train­ing. In short: if a per­son can be iden­ti­fied, the Act applies.

Sec­tion 3 of the Act applies to per­son­al data col­lect­ed in dig­i­tal form in India, or col­lect­ed on paper and dig­i­tized lat­er. It also reach­es pro­cess­ing out­side India when that pro­cess­ing is con­nect­ed with offer­ing goods or ser­vices to peo­ple in India [1].

Two exclu­sions mat­ter for AI teams. The Act does not apply to data processed for pure­ly per­son­al or domes­tic pur­pos­es. It also does not apply to per­son­al data made pub­lic by the per­son it relates to, or by some­one legal­ly required to pub­lish it. We cov­er that sec­ond exclu­sion below, because it is often mis­read as a scrap­ing license.

Four roles to map before any data moves

Every oblig­a­tion in the Act attach­es to a role. Map these on day one of a project, in writ­ing.

Role	Mean­ing in the Act	Typ­i­cal par­ty in an AI data project
Data Prin­ci­pal	The indi­vid­ual the data relates to. For a child, it includes the par­ent or law­ful guardian.	Speak­ers, cam­era wear­ers, bystanders, anno­ta­tors
Data Fidu­cia­ry	Decides the pur­pose and means of pro­cess­ing, alone or with oth­ers.	The AI com­pa­ny com­mis­sion­ing or buy­ing the dataset
Data Proces­sor	Process­es per­son­al data on behalf of a Data Fidu­cia­ry.	Col­lec­tion, tran­scrip­tion and anno­ta­tion ven­dors
Con­sent Man­ag­er	Reg­is­tered with the Board. Lets a per­son give, man­age, review and with­draw con­sent on one plat­form.	An option­al con­sent plat­form cho­sen by the indi­vid­ual

Roles fol­low deci­sions, not job titles. A ven­dor can be your proces­sor for the dataset and also a fidu­cia­ry for its own par­tic­i­pant onboard­ing and pay­ment records. Agree that split before kick­off.

When do the DPDP Rules 2025 take effect?

The DPDP Rules, 2025 were pub­lished in the Gazette as G.S.R. 846(E), dat­ed 13 Novem­ber 2025, and announced by the gov­ern­ment on 14 Novem­ber 2025. They com­mence in three phas­es. Board rules applied at once. Con­sent man­ag­er reg­is­tra­tion starts one year lat­er. Notice, secu­ri­ty, breach, chil­dren’s data and rights duties start eigh­teen months lat­er, around 13 May 2027.

Rule 1 of the Rules sets the phas­es [2]. The gov­ern­men­t’s own sum­ma­ry describes “an eigh­teen-month peri­od for phased com­pli­ance” [3]. Here is what each phase means for AI data work.

Phase	Rules that com­mence	Approx­i­mate date	What it means for AI data
On noti­fi­ca­tion	Rules 1, 2 and 17 to 21: def­i­n­i­tions and the Data Pro­tec­tion Board	13 Novem­ber 2025	Board frame­work exists; oper­at­ing duties not yet in force
+12 months	Rule 4: reg­is­tra­tion and oblig­a­tions of Con­sent Man­agers	Around 13 Novem­ber 2026	Reg­is­tered con­sent plat­forms can begin oper­at­ing
+18 months	Rules 3, 5 to 16, 22 and 23: notice, secu­ri­ty, breach, reten­tion, chil­dren, Sig­nif­i­cant Data Fidu­cia­ries, rights, trans­fers, research stan­dards	Around 13 May 2027	Full duties for buy­ers and their ven­dors

Some advis­ers count each date one day lat­er, from the 14 Novem­ber announce­ment. Con­firm exact dates with coun­sel before you plan around them.

Has the 18-month timeline been shortened?

Not as of mid-Octo­ber 2026. In Jan­u­ary 2026, Busi­ness Stan­dard report­ed a MeitY pro­pos­al to cut the win­dow to 12 months for Sig­nif­i­cant Data Fidu­cia­ries [4]. Some pro­vi­sions would also start soon­er. We found no Gazette noti­fi­ca­tion mak­ing that change as of this update.

One small change did land. On 5 Octo­ber 2026, MeitY issued the DPDP (Removal of Dif­fi­cul­ties) Order, 2026, S.O. 5458(E), under Sec­tion 43 of the Act. It adds the word “of” in Sec­tion 9(1) and changes “audit” to “data audit” in Sec­tion 10(2)©(ii). Law firm reviews describe both edits as edi­to­r­i­al. The order moves no dead­line.

Our advice: plan to May 2027, but build your con­sent and breach process­es as if the date could move ear­li­er.

What counts as personal data in AI training datasets?

Per­son­al data in train­ing datasets is any data about an indi­vid­ual who can be iden­ti­fied from it, direct­ly or in com­bi­na­tion. In AI work, that includes voice record­ings, faces in video, names and phone num­bers in tran­scripts, speak­er meta­da­ta such as age band and dis­trict, and anno­ta­tor IDs. Remov­ing names alone rarely makes a dataset non-iden­ti­fi­able.

Here is where per­son­al data hides, by data type.

• Speech. A voice record­ing is data about the speak­er. The spo­ken con­tent may name oth­er peo­ple. Meta­da­ta such as gen­der, age band, dis­trict and dialect nar­rows iden­ti­ty fur­ther. Our guide to Indic speech data col­lec­tion cov­ers how these fields are spec­i­fied.
• Ego­cen­tric and oth­er video. Faces, hands with jew­el­ry or tat­toos, num­ber plates, screens, doc­u­ments, house num­bers and bystanders who nev­er signed any­thing. See what to ask an ego­cen­tric data col­lec­tion ven­dor about bystanders.
• Text and chat logs. Names, phone num­bers, address­es and account num­bers inside free text. Code-mixed Indic text is hard­er to redact auto­mat­i­cal­ly than Eng­lish.
• Work­force data. Anno­ta­tor IDs, pro­duc­tiv­i­ty logs and pay­ment details. Anno­ta­tors are Data Prin­ci­pals too.

Is publicly available data exempt?

Only in part. Sec­tion 3©(ii) excludes per­son­al data made pub­lic by the Data Prin­ci­pal her­self, or by some­one legal­ly required to pub­lish it [1]. The Act’s own exam­ple is a per­son who blogs her views on social media.

The exemp­tion turns on who made the data pub­lic. A scraped cor­pus mix­es self-pub­lished posts with reposts, leaks, aggre­ga­tor list­ings and screen­shots that the per­son nev­er pub­lished. Law firm com­men­tary in 2026 notes that large-scale com­mer­cial reuse for AI train­ing remains unset­tled. Doc­u­ment prove­nance per source, and do not assume a scrape sits out­side the Act.

Does the research exemption cover model training?

Rarely for com­mer­cial prod­ucts. Sec­tion 17(2)(b) says the Act does not apply to pro­cess­ing nec­es­sary for research, archiv­ing or sta­tis­ti­cal pur­pos­es. Two con­di­tions apply. The data must not be used to take any deci­sion spe­cif­ic to a Data Prin­ci­pal. The pro­cess­ing must fol­low pre­scribed stan­dards [1].

Rule 16 and the Sec­ond Sched­ule set those stan­dards [2]. Pro­cess­ing must be law­ful and lim­it­ed to nec­es­sary data. It must aim for accu­ra­cy, keep data only as long as need­ed, apply secu­ri­ty safe­guards and have clear account­abil­i­ty. Rule 16 sits in the 18-month phase. A mod­el that lat­er dri­ves deci­sions about indi­vid­u­als, such as cred­it scor­ing or per­son­al­ized offers, is a weak fit. Treat the exemp­tion as nar­row and take advice before rely­ing on it.

What about anonymized data?

The Act does not define anonymiza­tion. Data that is tru­ly not about an iden­ti­fi­able per­son falls out­side the def­i­n­i­tion of per­son­al data. In prac­tice, voic­es and faces are hard to make non-iden­ti­fi­able, so most raw speech and video datasets stay in scope.

What consent does the DPDP Act require for AI training data?

Con­sent must be free, spe­cif­ic, informed, uncon­di­tion­al and unam­bigu­ous, giv­en by a clear affir­ma­tive action. It cov­ers only the data need­ed for the spec­i­fied pur­pose. It must fol­low an item­ized, plain-lan­guage notice, offered in Eng­lish or any Eighth Sched­ule lan­guage. With­draw­al must be as easy as giv­ing con­sent, and the fidu­cia­ry must prove it all.

Those words come from Sec­tions 5 and 6 of the Act [1]. The DPDP Rules 2025 con­sent notice stan­dard sits in Rule 3 [2]. The notice must stand on its own and item­ize the per­son­al data. It must state the spec­i­fied pur­pose and the uses it enables. It must also link to ways to with­draw con­sent, exer­cise rights and com­plain to the Board.

For AI data projects, four con­se­quences fol­low.

• Name AI train­ing as its own pur­pose. Con­sent giv­en for a ser­vice does not cov­er mod­el train­ing. The Act’s telemed­i­cine exam­ple shows con­sent shrink­ing to what the stat­ed pur­pose needs.
• Offer the notice in the par­tic­i­pan­t’s lan­guage. Sec­tions 5(3) and 6(3) require the option of Eng­lish or any lan­guage in the Eighth Sched­ule to the Con­sti­tu­tion. For Indic speech col­lec­tion, that usu­al­ly means the lan­guage being record­ed.
• Make with­draw­al real. Under Sec­tion 6(6), the fidu­cia­ry must stop pro­cess­ing with­in a rea­son­able time and make its proces­sors stop too. Decide ear­ly which arte­facts get delet­ed: raw files, tran­scripts, labels and derived sub­sets.
• Keep proof. Sec­tion 6(10) puts the bur­den on the fidu­cia­ry to prove that notice was giv­en and con­sent was obtained.

Con­sent is not the only law­ful ground. Sec­tion 7 lists “cer­tain legit­i­mate uses”, such as legal oblig­a­tions, med­ical emer­gen­cies and employ­ment. There is no gen­er­al “legit­i­mate inter­ests” ground of the GDPR kind. For com­mis­sioned data col­lec­tion, con­sent is usu­al­ly the prac­ti­cal basis. For voice projects, ask us for our speak­er con­sent tem­plate, which is built around these rules.

Consent manager under the DPDP Act

A Data Prin­ci­pal may give, man­age, review or with­draw con­sent through a Con­sent Man­ag­er reg­is­tered with the Board (Sec­tion 6(7) to 6(9)). Rule 4 on reg­is­tra­tion com­mences one year after the Rules were pub­lished, around 13 Novem­ber 2026 [2].

The First Sched­ule sets the bar. A Con­sent Man­ag­er must be a com­pa­ny incor­po­rat­ed in India with a net worth of at least ₹2 crore. It must not be able to read the per­son­al data it helps share. It must keep con­sent records for at least sev­en years and avoid con­flicts of inter­est with Data Fidu­cia­ries. For buy­ers, the prac­ti­cal step is sim­ple. Make sure your con­sent records can accept and hon­or sig­nals from a reg­is­tered plat­form if a par­tic­i­pant uses one.

Children and guardians

A child is any­one under 18 (Sec­tion 2(f)). Before pro­cess­ing a child’s data, the fidu­cia­ry must obtain ver­i­fi­able con­sent from a par­ent (Sec­tion 9(1)). Rule 10 requires due dili­gence that the per­son con­sent­ing is an iden­ti­fi­able adult. The check can use reli­able iden­ti­ty and age details already held, or details the per­son pro­vides [2]. It can also use a vir­tu­al token from an autho­rized enti­ty, such as a Dig­iLock­er ser­vice.

Sec­tion 9 also bars pro­cess­ing like­ly to harm a child’s well-being, and track­ing, behav­ioral mon­i­tor­ing or tar­get­ed adver­tis­ing direct­ed at chil­dren. Per­sons with dis­abil­i­ty who have a law­ful guardian need the guardian’s ver­i­fi­able con­sent under Rule 11. For speech or house­hold video projects, choose one path per project: exclude minors, blur them, or run a ver­i­fied parental con­sent work­flow.

What must a DPDP compliant data annotation vendor do?

The Act places legal duties on the Data Fidu­cia­ry, who stays respon­si­ble for every­thing its Data Proces­sor does. A DPDP com­pli­ant data anno­ta­tion ven­dor works only under a valid con­tract and on the buy­er’s instruc­tions. It applies the Rule 6 secu­ri­ty safe­guards, reports breach­es fast enough for the 72-hour Board report, and stops pro­cess­ing and eras­es data when told.

The ven­dor duties come most­ly through the fidu­cia­ry. Here is the chain, clause by clause.

• Respon­si­bil­i­ty stays with the buy­er. Sec­tion 8(1) makes the fidu­cia­ry respon­si­ble for pro­cess­ing done on its behalf, what­ev­er the con­tract says.
• A valid con­tract is manda­to­ry. Sec­tion 8(2) allows a proces­sor to be engaged only under a valid con­tract.
• Secu­ri­ty safe­guards cov­er the ven­dor. Rule 6 sets min­i­mum mea­sures [2]. They include encryp­tion, mask­ing or vir­tu­al tokens, access con­trol, log­ging and mon­i­tor­ing, and back­ups. Logs and relat­ed per­son­al data must be kept for one year to inves­ti­gate inci­dents, unless anoth­er law says oth­er­wise. The fidu­cia­ry-proces­sor con­tract must include secu­ri­ty terms.
• Breach report­ing runs on two clocks. Under Rule 7, the fidu­cia­ry must tell each affect­ed per­son and the Board with­out delay. A detailed Board report is due with­in 72 hours, unless the Board allows longer. Your ven­dor’s alert clause has to be much short­er.
• With­draw­al and era­sure reach the ven­dor. Sec­tions 6(6) and 8(7)(b) require the fidu­cia­ry to make its proces­sors stop pro­cess­ing and erase data shared with them.
• Know your sub-proces­sors. Under Sec­tion 11(1)(b), a per­son can ask the fidu­cia­ry to name every fidu­cia­ry and proces­sor their data was shared with. Keep a cur­rent list.

One clause helps Indi­an ven­dors with over­seas clients. Sec­tion 17(1)(d) exempts much of the Act for some off­shore work. It applies when an India-based enti­ty process­es data of peo­ple out­side India under a con­tract with a for­eign par­ty. Sec­tion 8(5) secu­ri­ty safe­guards and Sec­tion 8(1) respon­si­bil­i­ty still apply [1].

Who does what: buyer and vendor

Oblig­a­tion	Buy­er (Data Fidu­cia­ry)	Ven­dor (Data Proces­sor)
Notice and con­sent	Owns notice word­ing and pur­pos­es; must prove con­sent	Runs the notice and cap­ture flow as instruct­ed; keeps evi­dence
Secu­ri­ty safe­guards	Legal­ly respon­si­ble, includ­ing for ven­dor pro­cess­ing	Encryp­tion or mask­ing, access con­trol, logs, as con­tract­ed
Breach	Informs the Board and affect­ed peo­ple; files the 72-hour report	Alerts the buy­er at once with known facts
With­draw­al and era­sure	Stops pro­cess­ing; directs era­sure	Stops pro­cess­ing; deletes copies; con­firms in writ­ing
Rights requests	Answers access, cor­rec­tion and era­sure requests	Finds records by con­sent ID; sup­ports the response

Penalties under the DPDP Act

The Sched­ule to the Act sets max­i­mum penal­ties [1]. A fail­ure to take rea­son­able secu­ri­ty safe­guards can reach ₹250 crore. A fail­ure to noti­fy a breach can reach ₹200 crore, as can a breach of the chil­dren’s data duties. Sig­nif­i­cant Data Fidu­cia­ry breach­es can reach ₹150 crore, and oth­er breach­es ₹50 crore. The Board weighs grav­i­ty, dura­tion, rep­e­ti­tion and mit­i­ga­tion.

These penal­ties fall on the fidu­cia­ry. A ven­dor’s expo­sure is most­ly con­trac­tu­al, plus direct lia­bil­i­ty for any data it han­dles as a fidu­cia­ry itself.

Significant Data Fiduciaries

The gov­ern­ment may noti­fy some Data Fidu­cia­ries as Sig­nif­i­cant Data Fidu­cia­ries. Fac­tors include data vol­ume and sen­si­tiv­i­ty, and risk to peo­ple’s rights. They must appoint a Data Pro­tec­tion Offi­cer based in India and an inde­pen­dent data audi­tor. Under Rule 13, they must run a Data Pro­tec­tion Impact Assess­ment and audit every 12 months and report key find­ings to the Board. They must also check that algo­rith­mic soft­ware they use is not like­ly to pose a risk to peo­ple’s rights. Spec­i­fied data may face lim­its on trans­fer out­side India.

DPDP Act AI training data checklist: the 8‑point dataset diligence check

Use this check before a dataset enters your train­ing pipeline. It works for new col­lec­tions and for datasets you already hold. Each point maps to a clause above, so a “no” tells you exact­ly which gap to close.

1. Roles mapped in writ­ing. Name the fidu­cia­ry, the proces­sor and every sub-proces­sor for this dataset.
2. Item­ized notice per Rule 3. Lists the data col­lect­ed and the AI train­ing pur­pose, in the par­tic­i­pan­t’s lan­guage.
3. AI train­ing as a sep­a­rate pur­pose. A dis­tinct opt-in, not a clause buried in ser­vice terms.
4. Con­sent ref­er­ence on every file. Each record­ing, clip or doc­u­ment links to a con­sent record and notice ver­sion.
5. With­draw­al-to-era­sure path test­ed. Run one test dele­tion across your copies and the ven­dor’s copies.
6. Minors and guardians han­dled. Exclud­ed, blurred, or cov­ered by ver­i­fi­able parental or guardian con­sent.
7. Secu­ri­ty terms in the con­tract. Encryp­tion or mask­ing, access con­trol, and one-year logs per Rule 6.
8. Breach clock that fits 72 hours. The ven­dor alerts you fast enough for your detailed Board report.

For scraped or third-par­ty data, add two checks: prove­nance per source, and a writ­ten basis for any reliance on the pub­lic-data or research exemp­tions.

Want to run this check on a live project? Ask us for our con­sent & data-han­dling pack and com­pare it against your cur­rent ven­dor terms.

Common mistakes with personal data in training datasets

Most DPDP prob­lems in AI data projects trace back to a few habits. Each one is cheap to fix before col­lec­tion and expen­sive after.

• Reusing ser­vice con­sent for train­ing. Con­sent cov­ers the spec­i­fied pur­pose only. A new mod­el pur­pose needs its own notice and opt-in.
• Treat­ing “pub­lic” as “free to use”. The exemp­tion cov­ers data the per­son made pub­lic, not every­thing a crawler can reach.
• Call­ing com­mer­cial train­ing “research”. The research exemp­tion fails if the data dri­ves deci­sions about indi­vid­u­als.
• Strip­ping names and call­ing it anonymized. Voic­es, faces and rich meta­da­ta can still iden­ti­fy peo­ple.
• For­get­ting bystanders and anno­ta­tors. Both are Data Prin­ci­pals. Plan blur­ring or exclu­sion rules, and pro­tect work­force records.
• Los­ing track of off­shore access. Sec­tion 16 lets the gov­ern­ment restrict trans­fers to noti­fied coun­tries. Keep a map of where data and anno­ta­tors sit.
• No breach clock in the ven­dor con­tract. With­out one, the 72-hour Board report is hard to meet.

Teams that already run an AI risk pro­gram can add these items to the data sec­tion of their reg­is­ter. A data val­i­da­tion pass before train­ing is a nat­ur­al place to catch them.

How Graveiens AI handles consent

We run explic­it-con­sent data col­lec­tion on every project. The work­flow has four stages: par­tic­i­pant onboard­ing, record­ing or cap­ture man­age­ment, meta­da­ta tag­ging and QA. Con­sent sits in the first stage, and noth­ing moves to cap­ture with­out it.

• Con­sent first. Our con­sent-based data col­lec­tion starts with par­tic­i­pant onboard­ing, before any record­ing begins.
• Lan­guage cov­er­age. We work in 25+ lan­guages with deep Indic cov­er­age. That mat­ters when notices must be avail­able in the lan­guage a par­tic­i­pant speaks. See our voice and speech data ser­vices.
• Pay for approved deliv­er­ables. You review deliv­er­ables against agreed accep­tance cri­te­ria and pay for what you approve.
• Qual­i­ty sys­tem. Graveiens AI is ISO 9001 cer­ti­fied.
• Your terms, scoped togeth­er. Roles, notice word­ing, con­sent ref­er­ences in meta­da­ta, reten­tion and dele­tion are project-spe­cif­ic. We scope them with you at kick­off.

If you want to see the out­put before you com­mit, our 30-minute speech sam­ple set is free. For DPDP Act AI train­ing data projects, we scope the role map and con­sent flow with you first.

Frequently asked questions

Sources

1. Gov­ern­ment of India, “The Dig­i­tal Per­son­al Data Pro­tec­tion Act, 2023 (No. 22 of 2023),” Gazette of India, 11 August 2023, via MeitY. meity.gov.in
2. Min­istry of Elec­tron­ics and Infor­ma­tion Tech­nol­o­gy, “Dig­i­tal Per­son­al Data Pro­tec­tion Rules, 2025,” G.S.R. 846(E), 13 Novem­ber 2025. meity.gov.in
3. Press Infor­ma­tion Bureau, “DPDP Rules, 2025 Noti­fied: A Cit­i­zen-Cen­tric Frame­work for Pri­va­cy Pro­tec­tion and Respon­si­ble Data Use,” 17 Novem­ber 2025. pib.gov.in
4. Aashish Aryan, “Meity may cut com­pli­ance time­line for key DPDP rules to 12 months,” Busi­ness Stan­dard, 22 Jan­u­ary 2026. business-standard.com

About the author

Graveiens AI Edi­to­r­i­al Team. The Graveiens AI edi­to­r­i­al team writes about human data for AI: speech and ego­cen­tric data col­lec­tion, anno­ta­tion, val­i­da­tion and eval­u­a­tion. Our guides draw on how we scope, cap­ture and review con­sent-based data pro­grams for AI teams.