1. Meaning of AI Data Provenance

AI data provenance concerns the ability to identify and document the origin, ownership, collection method, transformations, licensing history, processing history and subsequent use of data used by an AI system.

In simple terms, provenance asks:

Where did this data come from, who supplied it, what happened to it before it entered the AI system, what transformations occurred, and on what legal basis was it used?

AI data provenance disputes can therefore concern:

• unlawfully collected training data;
• copyright-protected datasets;
• personal data used for model training;
• scraped internet material;
• confidential information;
• trade secrets;
• inaccurate datasets;
• improperly licensed databases;
• synthetic data derived from protected data;
• data obtained from third parties;
• data transferred across borders;
• inadequate records concerning training datasets;
• inability to identify the source of AI-generated information.

There is no single autonomous European cause of action called “AI data provenance dispute.” Liability is instead constructed from several areas of law, especially:

1. GDPR/data protection;
2. copyright;
3. database rights;
4. trade-secret law;
5. contract;
6. unfair competition;
7. consumer protection;
8. AI regulation;
9. confidentiality;
10. tort/delict;
11. intellectual-property law; and
12. fundamental rights.

2. Why Data Provenance Is Important for AI

Traditional software normally operates according to relatively identifiable inputs.

AI systems can involve:

• billions of data points;
• multiple datasets;
• third-party APIs;
• web scraping;
• licensed databases;
• user-generated information;
• synthetic data;
• reinforcement-learning data;
• continuously updated information.

This creates a provenance chain:

Original source → collection → extraction → transformation → labelling → storage → training → fine-tuning → deployment → output

A dispute can arise at any stage.

3. Typical AI Data Provenance Disputes

A. Ownership Disputes

The claimant may argue:

“The AI company did not own or have sufficient rights in the data it used.”

This is particularly important where datasets contain:

• photographs;
• books;
• journalism;
• music;
• software;
• databases;
• confidential material.

B. Copyright Disputes

Training datasets can contain copyright-protected works.

Questions include:

• Was copying lawful?
• Was the use covered by an exception?
• Was text-and-data mining permitted?
• Was the copyright holder able to reserve rights?
• Was the resulting AI output substantially related to protected material?

C. GDPR Disputes

AI datasets can contain personal information.

Questions include:

• Was the data lawfully collected?
• Was there an appropriate legal basis?
• Was the data used for a compatible purpose?
• Were individuals informed?
• Was data minimisation respected?
• Was inaccurate data corrected?
• Can an individual obtain information concerning processing?

4. Data Provenance and the EU AI Act

The EU AI Act significantly increases the importance of provenance for certain AI systems.

For relevant high-risk AI systems, governance can require attention to:

• data governance;
• data quality;
• data relevance;
• data representativeness;
• data-management practices;
• documentation;
• record keeping;
• traceability.

For general-purpose AI models, the regulatory architecture also places greater emphasis on documentation and information concerning training data.

Therefore, provenance is becoming not merely a technical best practice but potentially a regulatory compliance issue.

5. GDPR and Data Provenance

The GDPR does not use “AI data provenance” as a single independent legal doctrine.

Nevertheless, provenance is closely connected to several GDPR principles.

Article 5

Relevant principles include:

• lawfulness;
• fairness;
• transparency;
• purpose limitation;
• data minimisation;
• accuracy;
• storage limitation;
• integrity and confidentiality;
• accountability.

Article 15

Individuals may have rights concerning access to personal data and information about processing.

Article 16

Individuals can seek rectification of inaccurate personal data.

Article 17

The right to erasure can become important where personal data has been unlawfully processed, subject to the statutory exceptions.

Article 22

Automated decision-making can create additional rights where applicable.

6. The Provenance Problem in Machine Learning

A company may know that a model was trained on:

“Internet data.”

That description may be inadequate for serious legal compliance.

A proper provenance system may need to establish:

• source URL or database;
• source owner;
• date collected;
• collection method;
• legal basis;
• licence;
• copyright status;
• personal-data status;
• transformation history;
• filtering;
• deduplication;
• annotation;
• model-training use;
• retention period.

The absence of provenance can make litigation substantially more difficult.

7. Case Law

European courts have not yet produced a large body of cases specifically titled AI data provenance disputes. The following cases therefore include both directly relevant data/technology authorities and important analogous authorities.

Case 1 — Google Spain SL v AEPD and Mario Costeja González

C-131/12, CJEU, 2014

Facts

Mr Costeja González complained about search results linking his name to an old newspaper notice concerning property proceedings.

He sought removal of links from Google's search results.

Principle

The CJEU treated the search engine's processing of personal data as legally significant and recognised circumstances in which individuals could seek delisting.

Importance for AI provenance

The case establishes a fundamental principle:

A technology operator can have independent legal responsibilities concerning data that it processes, even though it did not originally create the underlying information.

This is highly relevant to AI developers using third-party datasets.

Governance lesson

An AI company cannot necessarily defend a provenance dispute simply by saying:

“We did not create the data.”

The company must consider its own role in processing and using the information.

Case 2 — Nowak v Data Protection Commissioner

C-434/16, CJEU, 2017

Facts

Mr Nowak sought access to information concerning an examination.

The dispute concerned whether answers and examiner comments constituted personal data.

Principle

The CJEU interpreted personal data broadly.

Information may constitute personal data where it is connected with an identifiable person, including evaluative information.

AI provenance relevance

AI systems frequently create:

• scores;
• profiles;
• predictions;
• classifications;
• evaluations.

Therefore, provenance analysis cannot be restricted to obvious identifiers.

A dataset may contain legally relevant personal data even when information has been transformed into an assessment or prediction.

Governance lesson

Companies must track not only:

“Where did the original data come from?”

but also:

“What personal information was derived from it?”

Case 3 — Wirtschaftsakademie Schleswig-Holstein GmbH

C-210/16, CJEU, 2018

Facts

The case concerned the operation of a Facebook fan page and the processing of visitor information.

Principle

The CJEU considered the circumstances in which an operator could have responsibility as a joint controller in relation to personal-data processing.

Importance for AI provenance

This is particularly important where AI systems involve multiple organisations.

For example:

Data provider → cloud provider → AI developer → AI deployer → customer

Different entities may contribute to the processing.

Governance lesson

Provenance disputes frequently require identifying who actually determined the purposes and means of processing.

Formal contractual labels may not settle the question.

Case 4 — Fashion ID GmbH & Co. KG v Verbraucherzentrale NRW e.V.

C-40/17, CJEU, 2019

Facts

Fashion ID embedded a Facebook “Like” button on its website, resulting in transmission of personal information to Facebook.

Principle

The CJEU examined joint controllership and responsibility for data transmission.

AI provenance significance

The case is relevant to AI ecosystems where data moves between several entities.

For example:

Website → analytics provider → AI provider → model developer

The organisation transmitting data cannot always assume that another entity bears the entire legal responsibility.

Governance lesson

Data provenance must be connected to responsibility mapping.

Case 5 — Orange România SA v ANSPDCP

C-61/19, CJEU, 2021

Facts

The case concerned whether consent to personal-data processing was genuinely valid.

Principle

The CJEU emphasised that valid consent requires the statutory conditions for freely given, specific, informed and unambiguous consent.

AI provenance relevance

AI companies frequently acquire data from:

• customers;
• apps;
• websites;
• partners;
• online platforms.

The fact that data was technically obtained does not establish that it was lawfully obtained.

Governance lesson

A provenance record should ideally establish:

source → legal basis → purpose → permission → subsequent AI use.

Case 6 — Planet49 GmbH

C-673/17, CJEU, 2019

Facts

The case concerned consent mechanisms for cookies and electronic communications.

Principle

The CJEU rejected the idea that consent can simply be inferred from passive or pre-selected mechanisms where active consent is legally required.

AI provenance relevance

This becomes significant when AI datasets are created through:

• online tracking;
• cookies;
• behavioural profiling;
• advertising technologies;
• user analytics.

Governance lesson

An AI company should not assume that the existence of a dataset proves lawful provenance.

The method of acquisition matters.

Case 7 — Digital Rights Ireland

Joined Cases C-293/12 and C-594/12, CJEU, 2014

Facts

The case concerned mandatory retention of electronic communications data.

Principle

The CJEU found the legislation incompatible with fundamental rights because of the breadth and seriousness of the interference and insufficient safeguards.

AI data provenance significance

Large-scale AI training may involve enormous datasets containing:

• communications;
• behavioural information;
• location data;
• metadata.

Digital Rights Ireland reinforces the importance of:

• necessity;
• proportionality;
• security;
• retention controls;
• lawful access.

Governance lesson

The mere availability of large datasets does not establish that an AI developer can lawfully acquire and retain them.

Case 8 — Tele2 Sverige AB and Watson

Joined Cases C-203/15 and C-698/15, CJEU, 2016

Facts

The cases concerned retention of electronic communications data and access by authorities.

Principle

The CJEU imposed important restrictions on general and indiscriminate retention and emphasised safeguards.

AI provenance relevance

Where AI systems use communications or behavioural data, companies need to know:

• where the data came from;
• why it was collected;
• how long it was retained;
• who could access it.

Governance lesson

Retention history is part of provenance.

Case 9 — Stichting Brein v Ziggo BV / Filmspeler line of cases

CJEU copyright jurisprudence

European copyright cases concerning online intermediaries are also important for AI training-data disputes.

The broader principle is that technological intermediaries may incur legal consequences depending upon their role in making protected works available or facilitating infringement.

AI relevance

This is analogous to disputes involving:

• AI dataset providers;
• scraping services;
• model-training platforms;
• repositories containing copyrighted material.

The legal question is not simply:

“Did the AI company create the copyrighted work?”

but:

“What did the company actually do with the protected material?”

Case 10 — SAS Institute Inc. v World Programming Ltd

C-406/10, CJEU, 2012

Facts

The case concerned software functionality, programming languages and copyright protection.

Principle

The CJEU distinguished between protected expression and aspects such as ideas, functionality and programming languages.

AI provenance relevance

The case illustrates the importance of separating:

• protected expression;
• unprotected ideas;
• functional information;
• technical concepts.

This distinction becomes important where AI systems are trained on:

• software code;
• technical documentation;
• datasets;
• programming materials.

Governance lesson

Not everything contained within a dataset necessarily receives the same form of intellectual-property protection.

8. Copyright and AI Training Data

One of the most significant provenance disputes concerns copyrighted works.

A dataset might contain:

• books;
• newspaper articles;
• photographs;
• music;
• paintings;
• films;
• computer programs.

The central questions include:

Question 1

Was the material copied?

Question 2

Was the copying legally authorised?

Question 3

Does a text-and-data-mining exception apply?

Question 4

Was a rights reservation effective?

Question 5

Was the material licensed?

Question 6

Can the claimant establish that its work was included?

Question 7

Did the AI system reproduce protected expression?

These are separate questions.

9. Text and Data Mining

European copyright law contains specific provisions concerning text and data mining.

This creates an important distinction between:

lawful analytical use of data

and

unauthorised reproduction or exploitation of protected works.

The legal assessment may depend on:

• whether the use is commercial;
• whether lawful access existed;
• whether rights were reserved;
• whether the relevant statutory exception applies;
• whether copies were retained;
• what happened to the extracted material.

10. Database Rights and AI

The EU Database Directive creates a separate sui generis database right.

A database owner may argue that an AI developer unlawfully:

• extracted substantial parts;
• repeatedly extracted portions;
• reproduced database contents;
• commercially exploited extracted material.

This can be particularly important where AI training involves structured datasets rather than individual works.

11. Trade Secrets and Confidential Data

AI provenance disputes may also involve confidential business information.

Examples include:

• proprietary datasets;
• customer databases;
• source code;
• internal reports;
• medical information;
• engineering data;
• financial information.

A company may face liability if it:

1. obtains confidential information unlawfully;
2. uses it outside the permitted purpose;
3. discloses it;
4. trains an AI system on it without authorisation.

Trade-secret law therefore becomes another layer of provenance protection.

12. Data Provenance and Contract Law

Many AI disputes will ultimately turn on contract.

A data licence may specify:

• permitted purposes;
• geographic territory;
• duration;
• sublicensing;
• AI training;
• model development;
• derivative use;
• data retention;
• deletion;
• audit rights.

A dispute can therefore arise because:

The company had access to the dataset, but did not have contractual permission to use it for AI training.

13. Data Provenance and Unlawful Scraping

Web scraping is one of the most difficult provenance issues.

A company may scrape:

• websites;
• social-media platforms;
• public databases;
• online marketplaces;
• academic repositories.

“Publicly accessible” does not automatically mean “free from all legal restrictions.”

Potential restrictions include:

• copyright;
• database rights;
• GDPR;
• contractual terms;
• trade secrets;
• unfair competition;
• confidentiality.

14. Data Accuracy and Provenance

Provenance is also closely related to accuracy.

Suppose an AI training dataset contains an incorrect allegation that a particular person:

committed a crime.

If the model repeatedly reproduces that information, potential legal problems can involve:

• GDPR accuracy requirements;
• defamation;
• privacy;
• consumer protection;
• professional negligence.

The company needs to know:

Who supplied the information?

When was it collected?

Was it verified?

Was it subsequently corrected?

15. Synthetic Data and Provenance

Synthetic data does not necessarily eliminate provenance issues.

Suppose:

Protected personal dataset → AI generation process → synthetic dataset → second AI model

Questions remain:

• Was the original dataset lawfully obtained?
• Does the synthetic dataset retain personal information?
• Can individuals be re-identified?
• Was the synthetic generation itself lawful?
• Does the synthetic data reproduce protected material?

Thus:

Transformation does not automatically erase legal provenance.

16. AI-Generated Outputs and Provenance

A second type of dispute concerns the provenance of AI outputs.

A claimant may ask:

“What source material caused this AI output?”

This can raise questions about:

• copyright;
• plagiarism;
• confidential information;
• personal data;
• defamation;
• trade secrets.

The evidential difficulty is that modern AI systems do not necessarily preserve a simple source-to-output chain.

17. The Black-Box Problem

AI provenance litigation can be difficult because a model may not provide a straightforward explanation such as:

“This sentence came from document X.”

Instead, the model may have learned statistical patterns from enormous numbers of documents.

Therefore, litigation may require:

• training-data records;
• dataset manifests;
• model cards;
• audit logs;
• source lists;
• sampling;
• technical expert evidence;
• metadata;
• data lineage records.

18. Burden of Proof

The claimant may have difficulty proving that its material was included in an AI training dataset.

Possible evidence includes:

• leaked datasets;
• technical documentation;
• developer disclosures;
• dataset inventories;
• source-code evidence;
• metadata;
• matching output;
• statistical similarity;
• internal correspondence.

Courts may need to balance:

claimant's need for evidence

against:

AI developer's trade secrets and confidential information.

19. Remedies

Depending on the applicable legal basis, remedies may include:

Data-protection remedies

• access;
• rectification;
• erasure;
• restriction;
• objection;
• compensation.

Copyright remedies

• injunction;
• damages;
• account of profits where available;
• destruction or removal of infringing copies;
• corrective measures.

Database-right remedies

• injunction;
• damages;
• cessation of extraction;
• other national remedies.

Contract remedies

• damages;
• termination;
• restitution;
• injunction;
• contractual audit.

Trade-secret remedies

• injunction;
• damages;
• prohibition of unlawful use;
• corrective measures.

Regulatory remedies

• administrative fines;
• corrective orders;
• processing restrictions;
• AI-system restrictions.

20. Defences

An AI company may argue:

1. Lawful licence

The company obtained valid permission.

2. Statutory exception

The use falls within a copyright or data-processing exception.

3. No personal data

The relevant information is genuinely anonymised or otherwise outside the applicable data-protection rule.

4. No protected work

The material lacks the necessary intellectual-property protection.

5. No substantial extraction

A database-right claim may fail if the required extraction threshold is not met.

6. Independent generation

The output was independently generated rather than copied.

7. Lack of causation

Even if provenance was defective, the claimant cannot establish legally recoverable harm.

21. Corporate AI Provenance Governance

A sophisticated AI company should maintain a data lineage register containing, where appropriate:

22. Six-Stage Legal Test for AI Data Provenance Disputes

A useful analytical framework is:

Stage 1 — Identify the data

What precisely was used?

Stage 2 — Identify the source

Who created, supplied or controlled it?

Stage 3 — Identify the legal status

Is it:

• personal data?
• copyrighted material?
• database material?
• confidential information?
• trade-secret information?
• public information?

Stage 4 — Identify the legal permission

What authorised the AI company to acquire and use it?

Stage 5 — Identify the AI use

Was it:

• training;
• fine-tuning;
• retrieval;
• testing;
• validation;
• deployment?

Stage 6 — Identify the consequence

What legally protected interest was harmed?

This produces:

Data source → legal status → permission → AI use → breach → causation → damage/remedy

23. Comparison of the Major Authorities

24. Important Distinction: Provenance vs Ownership

These concepts should not be confused.

Provenance asks:

Where did the data come from?

Ownership asks:

Who legally owns or controls the relevant rights?

Lawful-use analysis asks:

Was the AI company permitted to use it?

A company may have:

• lawful possession but no copyright licence;
• a copyright licence but no GDPR legal basis;
• access to personal data but no right to reuse it for AI training;
• a database licence but restrictions on extraction.

Therefore:

Access ≠ ownership ≠ permission ≠ lawful AI use.

25. Important Distinction: Public Data vs Free Data

Another important principle is:

Publicly accessible information is not automatically legally unrestricted information.

A website may be publicly accessible while its contents remain subject to:

• copyright;
• database rights;
• personal-data protection;
• contractual restrictions;
• confidentiality;
• other legal limitations.

Consequently, AI developers need a provenance analysis even for publicly available datasets.

26. Practical Example

Assume an AI company creates a medical model using:

5 million online medical articles + hospital records + patient data + public websites.

A patient later discovers that the model contains information about them.

The legal analysis might involve:

GDPR

Was the patient's data lawfully processed?

Copyright

Were protected articles copied?

Contract

Did the hospital permit the data to be used for AI training?

Confidentiality

Was medical information used outside its permitted purpose?

AI regulation

Were appropriate data-governance controls implemented?

Liability

Did the unlawful processing cause legally recognised harm?

This demonstrates why AI provenance is a multi-layer legal problem.

27. Conclusion

AI data provenance disputes are becoming a major European legal issue because modern AI systems depend on enormous quantities of data whose origins, legal status and permitted uses may be difficult to establish.

There is presently no single European cause of action specifically called AI data provenance liability. Instead, disputes arise through the interaction of:

• GDPR;
• copyright;
• database rights;
• trade-secret law;
• contract;
• consumer law;
• AI regulation;
• confidentiality;
• tort/delict;
• fundamental rights.

The cases of Google Spain, Nowak, Wirtschaftsakademie, Fashion ID, Orange România, Planet49, Digital Rights Ireland, Tele2 Sverige/Watson and SAS Institute demonstrate the developing legal principles.

The central legal chain is:

Source → collection → legal status → permission → transformation → AI training/use → output → harm → remedy.

For companies, the most important governance principle is therefore:

If an organisation cannot reliably explain where its AI data came from, what legal rights attach to it, and why the organisation was entitled to use it, it may face substantial regulatory, intellectual-property, data-protection and civil-liability risks.

The strongest future AI-provenance systems will consequently combine technical data lineage with legal provenance, allowing an organisation to demonstrate not merely what data entered an AI system, but also where it came from, what rights governed it, how it was transformed, who used it, and whether that use remained lawful throughout the AI lifecycle.