← Digital, E-Commerce & AI

AI · Data licensing

AI training and data licensing agreements

A dataset’s value depends on the actual right to use it. The agreement must distinguish between ownership of the medium, content rights, database rights, personal data and confidential information.

dataset provenance TDM and copyright model and output
Subject matterIdentified data, version, format and documentation
RightsTraining, fine-tuning, evaluation, reproduction and distribution
Critical evidenceThe provenance chain and source restrictions
01

What is actually licensed

The agreement identifies the sources, version, fields, metadata, labels and documentation. A link to a continuously changing bucket is not sufficiently precise subject matter without versioning rules.

Use for training, fine-tuning, testing, evaluation, RAG and subsequent improvement is distinguished. The right to access data does not automatically imply the right to reproduce, extract and use the content for all these purposes.

02

Copyright, databases and text and data mining

The dataset may contain protected works and may benefit from protection of its structure or the sui generis database right. Law No. 8/1996 and the European framework must be analysed for each source.

Directive (EU) 2019/790 permits certain reproductions and extractions for text and data mining from lawfully accessed content, but for general uses rights holders may appropriately reserve their rights, including through machine-readable means. The agreement must not assume that all public content is freely available for training.

03

Personal data and trade secrets

If the dataset contains personal data, the parties must establish their roles, legal basis, information, minimisation, retention, individuals’ rights and transfers. Pseudonymisation does not automatically take data outside GDPR.

  • The origin of the data and data subjects’ expectations.
  • Sensitive data, children’s data and information about convictions.
  • The possibility of reidentification and model extraction tests.
  • Requests for access, rectification, erasure and objection.
  • The supplier’s or third parties’ trade secrets.
  • Technical measures for access, storage and training environments.
04

Quality, model, output and termination

The agreement defines quality standards, representativeness, labelling accuracy and documentation of limitations. An absolute warranty that data contains no errors is rarely realistic; a reporting and correction mechanism is more useful.

Rights over the trained model, checkpoints, embeddings, evaluations and outputs are established. Termination provisions clarify deletion of the dataset, retention of the model and situations where data or a third party’s rights are withdrawn.

05

How we work together

  1. 01
    Inventory and architecture

    We clarify technology, actors, data flows, interface and the intended commercial outcome.

  2. 02
    Legal classification

    We establish roles, applicable regimes, risks and information requiring completion.

  3. 03
    Drafting or audit

    We prepare the licensing agreement and dataset documentation and align the document with the product, technical processes and available evidence.

  4. 04
    Implementation and review

    We deliver the final version, priority actions and reference points to monitor as products or legislation change.

QUESTIONS

Frequently asked questions

Can public internet data be freely used for AI?

Not automatically. Public access does not eliminate copyright, database rights, TDM reservations, GDPR, contracts or secrets.

Does pseudonymisation eliminate GDPR?

No. Pseudonymised data generally remains personal data if the person can be reidentified using reasonably available additional information.

Does the data licence grant rights over the model?

Only if the document establishes this. The dataset, training process, model and output are distinct assets and must be addressed separately.

Need an agreement for data and AI training?

Send your documents for a legal assessment and a solution tailored to your commercial objective.