Short answer
No. The DSM text-and-data-mining exceptions are input-stage exceptions. They do not authorise output-stage reproduction.
Read the operative words. Article 4(1) provides an exception "for reproductions and extractions of lawfully accessible works and other subject matter for the purposes of text and data mining". Article 3(1) is similarly framed, for reproductions and extractions made "in order to carry out, for the purposes of scientific research, text and data mining".
Article 2(2) defines the activity:
"'text and data mining' means any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations" —
The exception attaches to copies made in order to analyse. An output that reproduces protected expression and is delivered to a user is not an analytical copy — it is a fresh act engaging the reproduction right, and potentially the communication-to-the-public right.
So a developer relying on Article 4 to build a model cannot rely on the same provision to justify what the model emits.
Two separate stages, two separate acts
Input stage. Copying works into a training corpus, and the reproductions inherent in processing them, are acts of reproduction. The DSM exceptions can cover them, subject to lawful access and — for Article 4 — the absence of a reservation under Article 4(3).
Output stage. Where a model generates material reproducing a protected work in whole or in part, that engages separate rights:
"Member States shall provide for the exclusive right to authorise or prohibit direct or indirect, temporary or permanent reproduction by any means and in any form, in whole or in part: (a) for authors, of their works" —
Note the breadth: direct or indirect, temporary or permanent, by any means and in any form, in whole or in part. "Indirect" matters here, because an output is not copied directly from the source file.
Article 3 adds the communication and making-available rights, and Article 3(3) provides that those rights are not exhausted by any act of communication or making available.
Neither Article 3 nor Article 4 of the DSM Directive is drafted as an exception to output-stage acts. They list the specific rights they derogate from, for the specific purpose of mining.
When does an output reproduce a work?
Not every similarity is a reproduction. EU law protects expression, not ideas, facts or style.
Reproduction in part is possible with very little material. In Infopaq International (C-5/08) the Court addressed whether storing and printing an extract consisting of a small number of words could amount to reproduction in part, holding that this depends on whether the extract contains an element expressing the author's own intellectual creation.
Recognisability can matter. In Pelham and Others (C-476/17) the Court considered sampling of a phonogram and the concept of reproduction "in part", addressing when taking a short sound sample falls within the phonogram producer's right and when a sample modified so as to be unrecognisable to the ear does not. The Court returned to that framework in C-590/23 (2026).
Style and function are not protected. Output in the manner of an author, without reproducing their protected expression, is not reproduction.
The practical question for a developer is therefore not "did the model see this work?" but "does this output contain protected expression from it?". Memorisation and regurgitation of training examples is the risk case.
The three-step test constrains any reading
Even if one tried to stretch the TDM exceptions, Article 7(2) of the DSM Directive applies Article 5(5) of the InfoSoc Directive to them — the three-step test. Exceptions apply only in certain special cases which do not conflict with a normal exploitation of the work and do not unreasonably prejudice the legitimate interests of the rightholder.
An output that substitutes for the work in the market conflicts with its normal exploitation almost by definition. A reading of Article 4 that licensed substitutional outputs would fail the second step.
Article 7(1) is also instructive on what the legislature was protecting: it makes contractual provisions contrary to Articles 3, 5 and 6 unenforceable — but not Article 4, which is instead subject to the Article 4(3) reservation. These are carefully calibrated input-stage rules.
What the AI Act does and does not add
Article 53(1)(c) of the AI Act requires a GPAI provider to have a policy to comply with Union copyright law and to identify and comply with Article 4(3) reservations. Article 53(1)(d) requires a public training-content summary.
Neither authorises output reproduction. Recital 108 is explicit that the Regulation "does not affect the enforcement of copyright rules as provided for under Union law", and that the AI Office monitors compliance without conducting a work-by-work copyright assessment of training data.
So the AI Act regulates the provider's processes. It leaves the substantive copyright question about outputs exactly where it was.
Who is liable for an output?
The Directives do not answer this directly, and national law fills the gap. The realistic candidates are:
- The user, who prompted the generation and who reproduces, publishes or otherwise exploits the output.
- The provider, on the basis that the model as deployed makes infringing reproductions available, or on accessory-liability principles under national law.
- Both, depending on the facts and the national framework.
The Court's communication-to-the-public case law on intermediary and platform liability is not a direct analogy but is the nearest developed body of reasoning.
Contractual allocation between provider and user — indemnities, output warranties, acceptable-use restrictions — is now common. As with dataset licences, it allocates risk between the parties and does not bind rightholders.
What mitigates output risk
- Reduce memorisation. Deduplication, regularisation and training choices that reduce verbatim retention of training examples.
- Output filtering. Detection of substantial verbatim overlap with known protected material before delivery.
- Retrieval discipline. Where a system retrieves and surfaces source text, that is straightforwardly a reproduction and needs its own basis — a licence, an applicable exception such as quotation under Article 5(3)(d) of the InfoSoc Directive where its conditions are met, or the rightholder's permission.
- Attribution and linking do not by themselves make a reproduction lawful, although they may be conditions of a quotation exception.
- Licensing for high-risk content categories, particularly news, music, images and code.
- User terms addressing responsibility for use of outputs.
Operational checklist
- Treat input and output as separate legal questions with separate bases.
- Do not cite Article 3 or Article 4 as authority for output reproduction.
- Assess outputs by reference to protected expression, not to whether the work was in the corpus.
- Test for memorisation and verbatim overlap, and record the testing.
- Give retrieval and quotation features their own legal basis.
- Apply the three-step test when assessing any exception-based argument.
- Keep AI Act compliance separate from copyright authorisation.
- Allocate output risk contractually, and do not treat that as permission.
Conclusion
The TDM exceptions permit copying in order to analyse. They say nothing about what the analysis may afterwards produce, and they are drafted as derogations from specific rights for a specific analytical purpose.
An output that contains protected expression is a fresh reproduction — direct or indirect, in whole or in part — and where it is delivered to users it may also be a communication to the public. It needs its own basis: a licence, a distinct exception on its own terms, or the absence of protected expression in the first place.
Developers who conflate the two stages are building on a permission that was never granted.
Sources
This article provides general legal information, not advice for a specific matter. Laws, facts and procedural deadlines may change the result. Follow the linked primary sources and obtain local advice before acting.