EU AI Act Training Data Requirements: What Article 10 Means for Data Vendors

Last updat­ed: Novem­ber 2026

Key takeaways

• EU AI Act train­ing data require­ments sit main­ly in Arti­cle 10. They apply to high-risk AI sys­tems and cov­er train­ing, val­i­da­tion and test­ing data sets.
• After the Dig­i­tal Omnibus on AI, Arti­cle 10 applies to Annex III high-risk sys­tems from 2 Decem­ber 2027 and to Annex I prod­uct-embed­ded sys­tems from 2 August 2028.
• The provider owns the legal duty. Data ven­dors sup­ply the evi­dence, such as con­sent records, prove­nance, QA reports and bias reviews.
• The old Arti­cle 10(5) rule on spe­cial-cat­e­go­ry data for bias detec­tion now sits in a new Arti­cle 4a, with six strict con­di­tions.
• GDPR still applies in par­al­lel. An AI Act-ready dataset also needs a law­ful basis and a clean con­sent trail.

Buy­ers and audi­tors now ask high-risk AI teams one hard ques­tion. Can you prove where your data came from, and that it fits the job? EU AI Act train­ing data require­ments turn that ques­tion into a legal duty.

Arti­cle 10 sets out how train­ing, val­i­da­tion and test­ing data must be gov­erned, checked for bias and doc­u­ment­ed. The Dig­i­tal Omnibus on AI moved the dates, but the core duties stayed. This guide explains what Arti­cle 10 says, who owns each task, and the evi­dence pack your data ven­dor should hand over.

Not legal advice: this arti­cle is gen­er­al infor­ma­tion. Con­firm how the AI Act applies to your sys­tem with qual­i­fied EU coun­sel.

Build­ing a high-risk sys­tem for the EU mar­ket? Book a 20-minute scop­ing call and we will map the data evi­dence your project still needs.

What are the EU AI Act training data requirements?

EU AI Act train­ing data require­ments are the Arti­cle 10 rules for high-risk AI sys­tems. Train­ing, val­i­da­tion and test­ing data must fol­low doc­u­ment­ed gov­er­nance prac­tices and meet qual­i­ty cri­te­ria. They must also reflect the set­ting where the sys­tem will run and be exam­ined for bias. The provider must be able to show all of this.

In short: Arti­cle 10 does not ban any data source. It asks you to prove your data choic­es.

Which systems are covered?

Arti­cle 10 applies to high-risk AI sys­tems. These include Annex III uses such as employ­ment, edu­ca­tion and access to essen­tial ser­vices, plus AI built into prod­ucts cov­ered by Annex I. The Act reach­es providers that place such sys­tems on the EU mar­ket, wher­ev­er they are based.

For high-risk sys­tems that do not train mod­els, Arti­cle 10(6) applies the rules only to test­ing data sets.

When does Article 10 apply after the Digital Omnibus?

The Dig­i­tal Omnibus on AI, Reg­u­la­tion (EU) 2026/1744, was pub­lished in the Offi­cial Jour­nal on 24 July 2026. It entered into force on 27 July 2026. As of Novem­ber 2026, the Euro­pean Com­mis­sion’s AI Act Ser­vice Desk time­line lists these dates:

• 2 August 2025: oblig­a­tions for providers of gen­er­al-pur­pose AI (GPAI) mod­els apply.
• 2 August 2026: most rules apply, and enforce­ment starts for GPAI mod­els, pro­hi­bi­tions, trans­paren­cy rules and AI lit­er­a­cy.
• 2 Decem­ber 2027: rules for high-risk AI sys­tems in Annex III apply, includ­ing Arti­cle 10.
• 2 August 2028: rules for high-risk AI embed­ded in prod­ucts cov­ered by Annex I apply.

The delay buys time, not an exemp­tion. Data col­lect­ed now will often train the sys­tems that must com­ply in late 2027, so build the evi­dence dur­ing col­lec­tion.

What does Article 10 require, paragraph by paragraph?

Arti­cle 10 has four work­ing parts. Para­graph 2 lists the gov­er­nance prac­tices your data must fol­low. Para­graph 3 sets qual­i­ty cri­te­ria. Para­graph 4 asks data to match the real-world set­ting. A new Arti­cle 4a, added by the Dig­i­tal Omnibus, cov­ers sen­si­tive data used for bias detec­tion.

Article 10(2): data governance and management practices

Arti­cle 10(2) says data sets “shall be sub­ject to data gov­er­nance and man­age­ment prac­tices appro­pri­ate for the intend­ed pur­pose” of the sys­tem. It then lists what those prac­tices must cov­er. Beyond the rel­e­vant design choic­es (point a), they fall into six evi­dence areas:

1. Ori­gin (point b): data col­lec­tion process­es and where the data came from. For per­son­al data, record the orig­i­nal pur­pose of col­lec­tion.
2. Prepa­ra­tion (point c): anno­ta­tion, label­ing, clean­ing, updat­ing, enrich­ment and aggre­ga­tion steps.
3. Assump­tions (point d): what the data is sup­posed to mea­sure and rep­re­sent.
4. Suit­abil­i­ty (point e): an assess­ment of the avail­abil­i­ty, quan­ti­ty and suit­abil­i­ty of the data sets need­ed.
5. Bias checks (points f and g): exam­i­na­tion for bias­es that could affect health, safe­ty or fun­da­men­tal rights, and mea­sures to detect, pre­vent and mit­i­gate them.
6. Gaps (point h): data gaps or short­com­ings that block com­pli­ance, and how you will address them.

Article 10(3): the quality criteria

Arti­cle 10(3) is the line most teams quote. Data sets “shall be rel­e­vant, suf­fi­cient­ly rep­re­sen­ta­tive, and to the best extent pos­si­ble, free of errors and com­plete in view of the intend­ed pur­pose.” They must also have appro­pri­ate sta­tis­ti­cal prop­er­ties for the peo­ple or groups the sys­tem will be used on.

Note the words “to the best extent pos­si­ble.” The law does not demand per­fect data. It demands a doc­u­ment­ed effort, judged against the intend­ed pur­pose.

Article 10(4): context of use

Data sets must take into account the “geo­graph­i­cal, con­tex­tu­al, behav­iour­al or func­tion­al set­ting” where the sys­tem will be used. For a voice sys­tem, that can mean region­al accents, phone-line audio and real back­ground noise.

Former Article 10(5): now Article 4a

The Dig­i­tal Omnibus delet­ed Arti­cle 10(5) and moved its con­tent into a new Arti­cle 4a. Providers of high-risk sys­tems may “excep­tion­al­ly” process spe­cial cat­e­gories of per­son­al data to detect and cor­rect bias. This is allowed only where strict­ly nec­es­sary and when six con­di­tions are all met. Key ones:

• oth­er data, includ­ing syn­thet­ic or anonymized data, can­not do the job;
• pseu­do­nymiza­tion and strict access con­trols apply;
• the data is not trans­mit­ted to or accessed by oth­er par­ties;
• the data is delet­ed once the bias is cor­rect­ed or reten­tion ends.

Arti­cle 4a applies in addi­tion to GDPR, not instead of it.

The Article 10 Evidence Pack: a six-file checklist

Arti­cle 10 says what to gov­ern, not which files to keep. We use a sim­ple mod­el with data buy­ers: the Arti­cle 10 Evi­dence Pack. It is six files every train­ing, val­i­da­tion or test deliv­ery should car­ry, each mapped to Arti­cle 10.

1. Datasheet. What the data set con­tains, its intend­ed pur­pose, vol­umes, splits and known lim­its. Sup­ports design choic­es, assump­tions and suit­abil­i­ty (10(2)(a), (d), (e)).
2. Con­sent log. Who agreed to what, when, and for which use, with with­draw­al han­dling. Sup­ports the per­son­al-data pur­pose record in 10(2)(b) and your GDPR basis.
3. Prove­nance record. Col­lec­tion method, sources, dates, devices and any third-par­ty inputs. Sup­ports ori­gin (10(2)(b)).
4. Guide­lines. Ver­sioned anno­ta­tion and label­ing rules, with change logs. Sup­ports prepa­ra­tion (10(2)©).
5. QA report. Sam­pling method, error rates found, rework and accep­tance results. Sup­ports the “free of errors and com­plete” cri­te­ri­on (10(3)).
6. Bias review. Cov­er­age tables by rel­e­vant groups and set­tings, gaps found and actions tak­en. Sup­ports bias checks, gaps and con­text (10(2)(f)-(h), 10(4)).

The provider folds these files into its tech­ni­cal doc­u­men­ta­tion. If a file is miss­ing, log it as a known gap under 10(2)(h), with a fix plan.

Want to see how con­sent­ed, meta­da­ta-tagged data looks before you com­mit? Book a 20-minute scop­ing call, or scope a pilot billed only on approved files.

Who is responsible for training data: the provider or the data vendor?

The provider of the high-risk sys­tem car­ries the Arti­cle 10 duty. Sup­ply­ing data does not make a ven­dor the provider. But Arti­cle 25(4) requires the provider and sup­pli­ers of tools, ser­vices, com­po­nents or process­es to agree in writ­ing on the infor­ma­tion and help the provider needs. In prac­tice, the ven­dor sup­plies evi­dence.

Task	Provider	Data ven­dor
Intend­ed pur­pose and data needs	Defines and signs off	Turns needs into a col­lec­tion spec
Risk man­age­ment (Arti­cle 9)	Owns the risk sys­tem	Flags data risks it sees
Col­lec­tion and con­sent	Sets law­ful basis and rules	Runs col­lec­tion, keeps con­sent records
Anno­ta­tion and label­ing	Approves guide­lines	Labels to ver­sioned guide­lines
Qual­i­ty checks	Sets accep­tance cri­te­ria	Runs QA, shares reports
Bias exam­i­na­tion	Owns analy­sis and fix­es	Sup­plies cov­er­age data and meta­da­ta
Spe­cial-cat­e­go­ry data (Arti­cle 4a)	Process­es only under the six con­di­tions	Should not receive it unless coun­sel con­firms
Tech­ni­cal doc­u­men­ta­tion	Com­piles and keeps it	Deliv­ers the Evi­dence Pack files

The Arti­cle 4a row mat­ters most for plan­ning. Because spe­cial-cat­e­go­ry data must not reach oth­er par­ties, decide ear­ly where bias test­ing on sen­si­tive attrib­ut­es will hap­pen. Often that is inside the provider’s own envi­ron­ment.

Write the split into your con­tract, nam­ing each Evi­dence Pack file and its for­mat. A risk man­age­ment frame­work helps you con­nect data evi­dence to wider con­trols.

How does GDPR overlap with AI Act training data rules?

GDPR and the AI Act apply side by side. GDPR asks whether you may process per­son­al data at all, and how you respect peo­ple’s rights. The AI Act asks whether the data is good enough and fair enough for a high-risk use. Doc­u­men­ta­tion and spe­cial-cat­e­go­ry safe­guards sit in both.

The Euro­pean Data Pro­tec­tion Board (EDPB) adopt­ed Opin­ion 28/2024 on AI mod­els on 18 Decem­ber 2024. It said legit­i­mate inter­est can work as a legal basis only if the pro­cess­ing is strict­ly nec­es­sary and the bal­anc­ing of rights is respect­ed. It also warned that unlaw­ful­ly processed train­ing data could affect the law­ful­ness of a mod­el’s deploy­ment, unless the mod­el is duly anonymized. So the con­sent trail is a deploy­ment risk, not just paper­work.

For speech, image and video col­lec­tion, explic­it par­tic­i­pant con­sent is usu­al­ly the clean­est route. Our speak­er con­sent form tem­plate shows the fields a con­sent record should cap­ture.

What about general-purpose AI models?

GPAI providers face a dif­fer­ent data duty. Arti­cle 53(1)(d) requires a pub­lic sum­ma­ry of the con­tent used to train the mod­el, using a tem­plate from the AI Office. The Com­mis­sion pub­lished that tem­plate on 24 July 2025. GPAI oblig­a­tions have applied since 2 August 2025.

As of Novem­ber 2026, the Com­mis­sion time­line shows GPAI enforce­ment start­ed on 2 August 2026. Data sup­pli­ers to GPAI devel­op­ers should expect ques­tions on source types, col­lec­tion peri­ods and licens­ing, so keep them in the prove­nance record.

Common mistakes when evidencing Article 10

• Treat­ing the Omnibus as a pause. Data col­lect­ed today still feeds the sys­tems due in 2027.
• Assum­ing the ven­dor car­ries com­pli­ance. The provider owns Arti­cle 10. Con­tracts should define what evi­dence the ven­dor must deliv­er.
• Los­ing the orig­i­nal pur­pose. Reused per­son­al data with­out its orig­i­nal col­lec­tion pur­pose is hard to defend under 10(2)(b).
• Unlogged guide­line changes. If label­ing rules changed mid-project with­out a log, your prepa­ra­tion record breaks.
• One-off bias checks. Bias exam­i­na­tion should be repeat­ed when data, labels or the intend­ed pur­pose change.
• Send­ing sen­si­tive attrib­ut­es to ven­dors. Under Arti­cle 4a, spe­cial-cat­e­go­ry data used for bias work should not trav­el to oth­er par­ties.

A quick refresh­er on what train­ing data is, and how it dif­fers from val­i­da­tion and test data, helps teams avoid mix­ing these splits.

How Graveiens AI approaches Article 10-ready data

Graveiens AI col­lects and pre­pares train­ing data for AI teams in the US and EU. We do not give legal sign-off. We own the part a data part­ner can: con­sent­ed data with the evi­dence attached.

• Explic­it-con­sent col­lec­tion. Our data col­lec­tion work­flow runs par­tic­i­pant onboard­ing, record­ing and cap­ture man­age­ment, meta­da­ta tag­ging and QA, with con­sent cap­tured up front.
• Ver­sioned label­ing. Our data anno­ta­tion teams work to writ­ten guide­lines with logged changes.
• Inde­pen­dent checks. Data val­i­da­tion adds a sec­ond QA lay­er and a report you can file.
• Cov­er­age. We work in 25+ lan­guages, includ­ing Indi­an and Euro­pean lan­guages such as Ital­ian.
• Process dis­ci­pline. Graveiens AI is ISO 9001 cer­ti­fied and bills on approved deliv­er­ables.

Frequently asked questions

Get Article 10-ready training data

Start small and judge the files your­self. Book a 20-minute scop­ing call or ask us for the Arti­cle 10 Evi­dence Pack check­list. Pilots are billed only on approved files.

Sources

1. EUR-Lex: AI Act (Reg­u­la­tion (EU) 2024/1689), con­sol­i­dat­ed text as at 27 July 2026
2. Euro­pean Com­mis­sion AI Act Ser­vice Desk: imple­men­ta­tion time­line
3. Euro­pean Com­mis­sion: GPAI train­ing con­tent sum­ma­ry tem­plate (24 July 2025)
4. EDPB: Opin­ion 28/2024 on AI mod­els (18 Decem­ber 2024)

About the author

Graveiens AI Edi­to­r­i­al Team. Our edi­to­r­i­al team writes about AI train­ing data, speech data and data com­pli­ance for AI builders in the US and EU. The team works along­side our data col­lec­tion, anno­ta­tion and QA leads, so each guide reflects how real projects run.