Copyright and AI
Three separate questions: what may be trained on, what the output is worth, and what to watch when third-party works are the input.
The three questions
| Question | Short answer |
|---|---|
| May you train on it? | Under the text and data mining exception, with lawful access |
| Is the output protected? | Generally not, without human creative shaping |
| May a third-party work go into the prompt? | It is a reproduction and needs a basis |
Text and data mining
| Case | Basis | Reservation possible |
|---|---|---|
| Scientific research | Art. 3 DSM Directive, national implementations | no |
| Other purposes | Art. 4 DSM Directive | yes, machine-readable |
For companies the second is the normal case. The reservation must be expressed machine-readably, for instance via robots.txt or metadata. Whether and how far particular forms of expression are effective is the subject of ongoing litigation.
Protectability of the output
A work requires a personal intellectual creation. For purely machine-generated content that is absent. It follows:
- A generated image or text is generally in the public domain. Anyone may use it.
- Exclusivity is achieved by contract, secrecy or database protection, not by copyright.
- Substantial own editing can create protection, but only for your own contribution.
- The provider's terms of use may still restrict use. Those are contractual, not copyright.
When the output comes too close to a work
Regardless of the output's own protectability, it can infringe a third-party work where it takes over that work's creative features. This is not theoretical: models occasionally reproduce training content nearly verbatim or image-identically.
- For text: sample checks for verbatim reproduction of longer passages.
- For images: checks for recognisable imitation of a living artist's style, increasingly assessed as problematic.
- For trade marks and logos: do not generate third-party signs.
- For music and voice: separate rights that differ from the image position.
Third-party works as input
A text in a prompt is reproduced and transmitted to a service provider. For protected works that needs a basis:
| Case | Basis |
|---|---|
| Your own work | unproblematic |
| Licensed work | within the licence, which must cover transmission to third parties |
| Quotation | narrow, with a quotation purpose and attribution |
| Third-party work for analysis | text and data mining, with lawful access |
| Confidential third-party work | additionally trade secret protection |
What to document
- For your own training: provenance of the data, check of the reservation, time of retrieval.
- For generated content intended for publication: prompt, model, date, labelling applied.
- For use of third-party works: licence or exception basis, with a reference.