intellectual property and AI

In depth

On the way in

On the way in, the question is the right to place a document into the context. A contract covered by a confidentiality clause, a document subject to professional secrecy, a text under a restrictive licence are not given to a third-party service to read without checking. The point to examine is not only whether the material leaves the perimeter but retention and reuse: professional offerings generally exclude the use of content for training, consumer offerings rarely do, and it is this difference that makes shadow AI legally dangerous.

On the way out

On the way out, the dominant principle in Europe and in France is that a protected creation presupposes a characterised human intervention: a text produced entirely by a model from a brief instruction opens no copyright, which means that nobody can prevent its reuse. The more substantial the human contribution, choice, structure, rewriting, selection, the more plausible protection becomes again. This gradation has an immediate practical effect: whatever constitutes an asset for an organisation must keep a trace of its human work.

Unintentional infringement

A distinct risk deserves to be known: unintentional infringement. A model can reproduce an extract very close to a piece of content in its corpus, particularly on rare wordings, on code or on heavily repeated texts. It does not show on reading, and responsibility falls on whoever publishes. The reasonable precautions are known: check the passages that seem too well turned, be wary of proper nouns and of quotations, and look at whether the provider offers a contractual guarantee, several now offering one for professional use.