EU AI Act Copyright Compliance Leaves Training Claims Open
Primary lensTrade policy
Sub-topicPolicy monitoring
Evidence base7 records used
Use casePolicy monitoring
EU Compliance Leaves the Underlying Copyright Question Open
EU AI Act compliance does not settle whether a model's training infringed copyright. The law requires a public account of training content that can help rightsholders examine potential claims, while leaving individual works outside the AI Office's copyright review. For a US general-purpose AI model provider, a compliant release can therefore expose sourcing decisions to closer scrutiny. That is the consequence of the division described in AI Act recitals 107 and 108.
The distinction now has enforcement behind it. The Commission's AI Act enforcement framework, updated August 24, confirms that its powers over general-purpose AI model obligations apply from August 2, 2026. Those obligations include respect for copyright and transparency about training. Providers of models placed on the market before August 2, 2025 retain the separate August 2, 2027 compliance deadline.
This is a concrete test for the international cooperation being discussed around the September 1-2 G20 Innovation Ministerial. In its August 31 announcement of the EU delegation's agenda, the Commission said Henna Virkkunen would call for interoperability to reduce regulatory fragmentation. The announcement describes a planned position, not an agreement changing copyright obligations.
A US provider's trade counsel should review the release file before EU market placement. It must show which model and version are being offered, which provider duties apply, and which copyright questions remain unresolved after those duties are met. A business that merely uses another company's model does not acquire all model-provider duties on that fact alone.
A US Training Decision Does Not Complete the EU Review
Article 53 of the current AI Act requires a policy to comply with EU copyright law, particularly identifying and respecting reservations of text-and-data-mining rights. Recital 106 explains that this policy obligation applies when a provider places a model on the EU market regardless of where the copyright-relevant training acts occurred.
A US fair-use assessment addresses another legal question. 17 USC 107 calls for consideration of the particular use, the work, the amount used and the market effect. A conclusion reached under that provision cannot substitute for the EU policy obligation. Equally, recital 106 does not itself decide that every training act outside Europe infringes EU copyright. Applicable law and infringement must still be assessed for the disputed conduct.
The AI Office's role has its own boundary. Recital 108 says it should monitor whether the policy and summary obligations have been fulfilled without assessing copyright compliance work by work. Existing copyright enforcement remains unaffected. Counsel cannot turn a regulatory compliance conclusion into assurance that a rightsholder has no claim.
The Public Summary Gives Rightsholders a Starting Point
The Commission's training-summary FAQ describes a mandatory template covering the model, its training sources and relevant processing. It is intended to help parties with legitimate interests exercise their rights. The summary must identify the models it covers and be publicly available by EU market placement, subject to the older-model transition.
The legal significance is the connection between disclosure and enforcement. A description of a dataset or collection can help a rightsholder decide what to investigate. It does not prove that a particular work was used, that the use required permission, or that infringement occurred. The claimant still needs the relevant facts and legal basis.
For the provider, the practical risk is an inconsistency between the published account and the sourcing records needed to answer a later challenge. Release counsel should reconcile the summary with records of acquisition, permissions and rights reservations before publication. That is a proposed review discipline, not an additional statutory filing.
One Training Record Serves Different Recipients
The AI Act distinguishes public disclosure from information supplied to authorities and downstream system providers. Treating them as interchangeable can produce either excessive disclosure or inadequate compliance. The following map identifies what each recipient receives and what that material cannot establish by itself.
Recipient
Required material
Limit of what it establishes
Public, including rightsholders
Training-content summary under Article 53(1)(d), using the AI Office template
A starting point for examining sources, not an inventory proving lawful use of every work
AI Office and competent authorities
Technical documentation under Article 53(1)(a), available on request
Evidence for supervision, not a work-by-work copyright clearance
Downstream AI-system providers
Integration information under Article 53(1)(b) and Annex XII
Information about capabilities and limitations, not a blanket warranty of training rights
Disclosure map for covered GPAI model providers as of September 2, 2026. Based on Article 53 and its confidentiality provisions and the Commission's training-summary template. Qualifying open-source exemptions affect the documentation rows, but preserve the public summary and copyright policy.
Trade-secret protection shapes how the obligations operate. It does not make the public summary optional. The template uses different levels of detail for different sources, while Article 53 protects confidential material through the relevant channels. Conversely, the summary requirement does not require a provider to publish its entire technical dossier or model weights.
The Commission's provider guidance also distinguishes voluntary use of a code of practice from the obligation to demonstrate compliance. Choosing an alternative route does not remove the mandatory summary. Qualifying open-source models retain the copyright-policy obligation even where certain documentation duties fall away.
Reopen the Release File When the Model or Evidence Changes
A release assessment should be tied to a named model version and its EU market-placement history. Keep the public summary, the supporting sourcing record and the legal assessment aligned when further training changes the data. The Commission's summary guidance calls for updates at six-month intervals, or sooner when added training data requires a materially significant change.
The policy question differs from the cross-border recognition of technical evidence examined in Traverse's analysis of the proposed EU-US cybersecurity mutual recognition agreement. Recognition can reduce duplicate assessment. Here, the unresolved question is what a compliant disclosure proves about the underlying rights.
A substantiated claim, newly identified rights reservation, changed training source or authoritative ruling can require counsel to revisit that conclusion. Revised binding requirements could change the regulatory part of the file. Until then, the EU release decision should state the remaining copyright exposure expressly, even when the provider has met its disclosure obligations.
From reading to review
Run the numbers on your lane.
The duty calculator runs the current stack for any HTS code and origin. A free account opens full tool output, AD/CVD detail, Chapter 98 processing, and available exports.