Text and Data Mining for Literature Reviews in the EU: DSM Directive Articles 3 and 4 and Germany's UrhG Rules Explained

Text and data mining (TDM) in the EU rests on two exceptions in the Digital Single Market (DSM) Directive [1]. DSM Directive Article 3 covers research organisations and cultural heritage institutions mining works for scientific research; DSM Directive Article 4 covers everyone else, subject to a rights-holder opt-out.

The Directive defines TDM as any automated analytical technique aimed at analysing text and data in digital form to generate information, including patterns, trends and correlations (Art. 2(2)) [1]. Article 4 lets anyone copy and extract lawfully accessible works for mining, keeping copies as long as necessary, unless rights holders have expressly reserved that use in an appropriate manner, such as machine-readable means for content made publicly available online. Article 3 has no opt-out: rights holders may only take measures to protect the security and integrity of their networks and databases, and a contract cannot override the research exception (Art. 7(1)) [1].

Germany transposed both articles through the Act adapting copyright law to the Digital Single Market of 31 May 2021, which introduced § 44b UrhG and rewrote § 60d UrhG in the German Copyright Act with effect from 7 June 2021 [4].

§ 44b defines TDM as the automated analysis of one or more digital or digitised works to obtain information, in particular patterns, trends and correlations, and permits reproductions of lawfully accessible works for it. Copies must be deleted when no longer needed, and the use is allowed only if the rights holder has not reserved it; for works available online, a reservation is effective only if it is machine-readable [2].

§ 60d is the research rule. It covers universities and other research organisations that pursue no commercial purpose, reinvest all profits in research or act in the public interest under a state-recognised mandate. It also covers libraries, museums and archives, and individual researchers with no commercial purpose. An organisation cooperating with a private company that has decisive influence over it and preferential access to its results is excluded [3]. As with Article 7(1), contract terms cannot restrict the § 60d use [7].

Germany has the first appellate ruling on TDM for AI training data. In Kneschke v LAION, the Hamburg Regional Court held in September 2024 that the non-profit LAION’s creation of a training dataset was covered by § 60d. On 10 December 2025 the Hanseatic Higher Regional Court dismissed the photographer’s appeal, relying mainly on § 44b: the reservation in the website’s terms was not machine-readable, so it did not take effect. It also upheld § 60d as a further ground and allowed an appeal to the Federal Court of Justice, so the ruling is not yet final [5].

AI developers face a separate duty: since 2 August 2025, providers of general-purpose AI models must have a policy to identify and comply with Article 4(3) reservations [6]. For a review team the point is narrower: the research exception is real, but whether it applies depends on who you are and what you do with the copies.

Reading papers by hand is not TDM, but several review tasks can be: bulk-downloading full texts for automated analysis, running text-mining or machine-learning screening tools over publisher content, and building a corpus for citation or term analysis.

• Check who makes the copies. § 60d covers research organisations and non-commercial researchers. A commercial provider supporting a university team may not qualify itself, so agree in writing who runs the mining and on whose infrastructure.
• Check lawful access. Both exceptions apply only to lawfully accessible works, so your institution’s licence or open-access status comes first.
• Look for reservations where § 44b applies: publisher terms, robots.txt and other machine-readable signals.
• Keep corpora under control. Delete copies you no longer need. Under § 60d only research organisations and heritage institutions may keep them, securely, for verification; individual researchers should delete them.
• Document it. Record in the methods which automated tools were used, on which sources, and under which exception, alongside your search strings.

Most literature review support involves no TDM at all: searching, screening and synthesis are done by people. Where a systematic review uses machine-learning screening on titles and abstracts exported from a database you licence, the usual question is your licence terms, not the TDM exceptions. Plan the rest of the literature review steps as normal, and see our note on text mining in data extraction.