Key takeaway
A confidential dataset and a public training-content summary are different disclosure layers. Resolve their interaction before promising that every source detail will remain private.
Establish whether this is the provider’s obligation
A potential recipient wants business manuals for a model and offers to keep every reference to the source confidential. Before treating that promise as settled, establish which model and entity are involved, whether the recipient is a provider of a general-purpose AI model within the relevant EU scope and what it actually intends to train. An internal retrieval application, a general-purpose model and a downstream modification are not interchangeable descriptions.
Article 53(1)(d) of the AI Act requires providers within its scope to make a sufficiently detailed training-content summary public using the AI Office’s template. The Commission’s template FAQ explains the scope and describes the public summary’s purpose. This article addresses that specific EU disclosure question; it does not determine provider status, replace the template or decide whether any training activity has lawful source rights.
Separate the summary from a raw-record release
The Commission describes three public-summary sections: general model/content information, data sources and relevant processing aspects. The source categories include public datasets, private datasets, online scraping, user data and synthetic data, with different detail requirements. The FAQ also explains the balance with confidential business information and trade secrets. Neither a blanket promise of total secrecy nor an assumption that every training record must be published accurately captures that explanation.
For a proposed business dataset, ask the provider to identify the applicable template fields and a proposed source description. Do not infer the required detail for a private dataset from the rules for scraped domains or user interactions. A lawyer and the provider’s responsible team need the actual model facts and current template, together with the license’s confidentiality language, before settling what may or must be disclosed.
An illustrative three-layer disclosure map
In this hypothetical proposal, fictional company Alder offers 100 company-authored equipment manuals to fictional model provider Vale. Rights in those manuals are assumed only for this exercise; customer attachments and individual names are excluded from the described candidate. The draft confidentiality clause prohibits any external description of the source, while Vale says it expects to publish a training summary. No actual dataset, provider determination or agreement is established.
The completed map records the contradiction without inventing the legally required wording. Alder’s commercial sponsor asks for a reviewed public description and corresponding contract exception, while keeping raw content and restricted details on their own access path. The count of 100 is an invented inventory fact, not a template threshold or evidence that this package is sufficiently detailed.
| Disclosure layer | Proposed treatment in this example | Decision before agreement |
|---|---|---|
| Raw manuals and authorized access | Controlled delivery only if separately approved | Define exact material, purpose and users |
| Public training-summary description | Required scope/detail unresolved | Obtain actual template mapping and reviewed wording |
| Customer identities and attachments | Excluded from this candidate | Verify exclusions; do not expose them to resolve the summary question |
| Contract: no external source description | Conflicts with provider’s stated summary plan | Resolve wording and authority; do not sign the contradiction |
Keep the copyright policy beside the transparency question
Article 53(1)(c) separately addresses a copyright-compliance policy, including identifying and complying with relevant reservations of rights. Paragraph 2’s stated free-and-open-source exception concerns the documentation obligations in points (a) and (b); it does not list points (c) and (d). The Commission’s FAQ likewise explains that open-source general-purpose models within scope still need the public summary.
Publishing a source description therefore does not prove the provider had permission to use the material, and obtaining a license does not itself establish that the public-summary duty is fulfilled. Keep the grant, actual training use and public disclosure plan linked but separately reviewed. A title saying open source or licensed should not be used as a substitute for either assessment.
Resolve the actual statement, model and update responsibility
Alder’s finished preparation note names Vale’s model/version, the exact candidate source category, the draft public description and the confidentiality issue requiring agreement. It leaves provider status and mandatory detail with the authorized reviewers. The sponsor can now judge a concrete proposed statement instead of approving an unspecified future disclosure. No raw manual needs to be sent merely to write this initial scope note.
Also ask who checks later changes in model training and the corresponding summary. The Commission FAQ describes updating summaries when additional training changes their content. An agreement should identify the responsible people and review process for the actual situation, rather than quietly allowing any future public description. The result is a reviewable disclosure boundary, not a guarantee of secrecy, a public-summary compliance certificate or permission for an unapproved training use.
Use this review agenda with your legal and privacy advisers. Requirements depend on the records, jurisdictions and intended use.