AI training data licensing across the product supply chain

Use an asset schedule covering five layers. Record the supplier, version, relevant countries and governing documents for each.

1. Training, fine-tuning and evaluation datasets. Identify permission to acquire, copy, transform and retain the material for each development activity. Include benchmark terms and restrictions on publishing test material.
2. Base models, modified weights and software. Separate the model licence, code licences and hosted API terms. Check commercial use, modification, redistribution, service access, attribution and any restrictions on competing models.
3. Customer information. Cover prompts, uploads, feedback, outputs retained as logs and support records. State which uses serve the customer and which would improve the supplier’s wider product.
4. Retrieval sources and indexes. Identify rights to ingest documents, create embeddings, cache passages and disclose retrieved content to users. A source subscription may limit users, territories and onward delivery.
5. Outputs and synthetic datasets. Record the generation inputs, applicable model terms, human contributions and proposed recipients. Reuse requires a separate check of protected expression, personal data and confidential information.

Copyright permission concerns copying protected works and other restricted acts. Privacy requires a lawful basis and applicable safeguards for personal-data processing; consent is only one possible basis. Confidentiality agreements restrict disclosure and use, including information that contains no personal data. One permission cannot establish the other two. (Directive 2001/29/EC, Article 2; GDPR, Articles 5, 6 and 9.)

Training and fine-tuning use data to develop or change a model. Evaluation tests it and can require separate copies or disclosures. Retrieval-augmented generation, or RAG, supplies selected source material when answering a query. Its legal review must cover ingestion, indexing, retrieval and the passages delivered in answers. An agreement covering training alone may leave those acts outside its grant.

For EU text and data mining, Article 4 of Directive (EU) 2019/790 requires lawful access and no effective reservation by the rights holder. Online reservations must be expressed appropriately, including by machine-readable means. The separate Article 3 route concerns scientific research by qualifying research organisations and cultural heritage institutions. A commercial developer cannot assume that it qualifies merely because its engineers conduct research. Before mining, check the applicable national provisions and their conditions for retaining copies. (Directive (EU) 2019/790, Articles 2–4 and 7.)

Germany’s implementation permits qualifying mining of lawfully accessible works unless rights are reserved. Online reservations must be machine-readable, and copies must be deleted when no longer necessary for mining. That route does not authorise every later disclosure of source passages or reproduction in model outputs. Review those acts separately. (Urheberrechtsgesetz, sections 16, 19a, 44b and 87b.)

UK commercial development cannot rely on the computational-analysis exception in section 29A, which requires the sole purpose of non-commercial research. The developer needs a licence or another applicable exception for restricted copying. US fair use requires an assessment of the actual use under four statutory factors, including market effects. A US conclusion does not supply permission for copying in another country. (Copyright, Designs and Patents Act 1988, sections 16 and 29A; 17 U.S.C. sections 106–107.)

Database rights need their own check. Extracting substantial contents, or repeatedly taking small parts in the prohibited circumstances, can engage rights beyond copyright in individual records. The review should establish the protected database, qualifying maker and relevant acts. (Directive 96/9/EC, Article 7; Copyright and Rights in Databases Regulations 1997, regulations 13 and 16–18.)

Before ingestion, the vendor’s legal and engineering teams should agree an activity schedule and permission basis for each corpus. Where the proposed use exceeds the licence or an exception’s conditions, obtain the missing grant, remove the material or change the operation. Record collection, copying, training, hosting and support locations so origin-country restrictions and destination duties receive separate review.

Which copyright route supports the planned training?

Compare training permission and transparency duties. Read the jurisdictional analysis when deciding between licensed material and an applicable statutory exception.

A no-training promise needs an operational definition. The customer schedule should distinguish service delivery, customer-specific evaluation, security monitoring and support from general training, fine-tuning, shared benchmarking and product improvement. Include prompts, retrieved documents, embeddings, feedback and logs. Obtain matching commitments from model and hosting suppliers, and configure retention and reuse settings accordingly.

Where EU or UK GDPR applies, processing personal data on the customer’s behalf requires Article 28 terms. If the vendor determines a separate improvement purpose, assess its controller responsibilities for that processing. Establish the lawful basis, purpose compatibility, notices, sensitive-data conditions and any required impact assessment before reuse. Assess EU and UK transfer requirements separately, including qualifying overseas support access. (EU GDPR and UK GDPR, Articles 3, 5, 6, 9, 13–14, 28, 35 and Chapter V; UK GDPR, Article 8A.)

Synthetic data and embeddings require evidence about their actual content. Removing names or generating new records does not establish anonymity. Assess extraction and identification risks for the released artefact and intended recipients. Even established anonymity leaves applicable copyright and contractual restrictions to resolve. (GDPR, recital 26; EDPB Opinion 28/2024, paragraphs 38–48.)

Can customer information be reused for improvement?

Review data protection for AI. Identify the processing purpose, lawful basis, data-processing terms and safeguards needed before reuse.

Can synthetic data leave the customer environment?

Read the synthetic-data supply analysis. Follow the separate decisions on generation, supplier access, anonymity and reuse by other customers.

The permission chain starts with the relevant rights holders. An aggregator must be authorised to grant the promised uses; the model supplier and reseller must preserve upstream limits. Request the actual grants and any necessary sublicensing authority. Compilation copyright does not establish rights in every underlying item. (17 U.S.C. sections 103(b), 106 and 201(c).)

Openly downloadable weights require a version-specific licence review. Apache License 2.0 permits specified uses and redistribution subject to its conditions, while generally disclaiming title and non-infringement warranties. It does not establish rights in every training item. Other model licences can restrict branding, recipients or uses. (Apache License 2.0, sections 2–4 and 7–9.)

Ask each supplier for:

• The asset inventory, source identifiers, acquisition dates and links between datasets, transformations and model versions.
• The executed licences, contributor permissions, accepted terms and evidence of authority to sublicense or permit downstream use.
• The permitted activities, territories, customer categories, attribution duties and recorded rights reservations.
• The customer-data flow, processor terms, reuse settings, retention periods and subprocessor arrangements.
• The relevant testing, complaints and removal records, together with available statutory documentation and the proposed response to a defective source.

For providers placing general-purpose AI models on the EU market, copyright-policy and public training-summary duties remain even for qualifying open-source releases. They apply from 2 August 2025; models placed on the market before that date have until 2 August 2027 to comply. The public summary differs from a buyer’s negotiated evidence request and the authority-facing technical file. (Regulation (EU) 2024/1689, Articles 2(1)(a), 53, 111(3) and 113(b), Annexes XI–XII, as amended by Regulation (EU) 2026/1744.)

Procurement should record evidence gaps and their consequences. Where essential rights cannot be established, approval should depend on replacement data, narrower use or documented permission covering the intended activity. An indemnity cannot supply the missing right.

What does the upstream model licence permit?

Check model licensing and EU supply. Examine redistribution, fine-tuning, provider documentation and the limits of open-source exemptions.

The agreement should distinguish rights in customer inputs, supplier technology and outputs. US copyright requires human authorship; the UK retains a separate rule for qualifying computer-generated works. Assign the rights the supplier actually holds, and avoid promising exclusive protection for every generated answer. (17 U.S.C. section 102(a); Thaler v. Perlmutter, No. 23-5233, slip opinion, pp. 3, 9–10 (D.C. Cir., 18 March 2025); Copyright, Designs and Patents Act 1988, sections 9(3) and 178.)

Specify what happens to source files, indexes, embeddings, caches, logs and trained components when access ends or a rights claim arrives. Before agreeing to deletion or removal terms, the vendor should demonstrate that it can perform the promised process. Deleting a source file does not establish that its information has disappeared from trained weights. Where required removal cannot be achieved, suspend the affected use and assess replacement or retraining.

Negotiate claim notification, defence control, evidence preservation, replacement rights, refunds and transition assistance. Continued use after termination needs a surviving permission or another lawful basis. Preserve evidence required for a dispute under controlled access, and distinguish that retention from continued product use.

Where the service qualifies under the EU Data Act, statutory switching duties also apply. Check the service-specific exceptions and export scope; proprietary model weights do not automatically become transferable. (Regulation (EU) 2023/2854, Articles 2(8), 23–25, 30(6) and 31.)