Glossary
Text and data mining
Text and data mining (TDM) means the automated analysis of digital works to obtain information, above all about patterns, trends and connections. In German copyright law, § 44b of the Copyright Act (UrhG) sets the conditions for copies. The exception is no blanket permission for everything one can do with a text.
The exception permits copies (legal term: reproductions) of works that are lawfully accessible, that is, available by a route the law permits. Public availability on its own says nothing reliable about this. Anyone who holds the rights to a work and does not want it used for this can set a rights reservation. Under the wording, such a reservation for works available online takes effect only if it is machine-readable. For research there is a separate rule in § 60d UrhG.
An example: A program reads thousands of news articles and counts which words appear together. To read them, it makes copies of the articles. The result, a statistic about word pairs, no longer contains any article. The question is whether the copies were permitted beforehand.
Not to be confused with training a model: TDM describes the analysis of the texts; training is a step that turns them into parameters. Whether the statutory exception also covers training is open and is currently being decided in court.
Where you’ll come across it: In terms of use and reservations on websites, in statutes, and in reports on lawsuits against AI providers.
Introduced in “Training Data: Copyright and Privacy”. The origin of the texts from which TDM copies are made is described in “Text Becomes Many Practice Problems”. For the training data itself, see there.