open source

In depth

The missing source code

The difficulty specific to AI is that “source code” has no obvious equivalent. For a piece of software, publishing the source is enough to allow study and reconstruction. For a model, one would need at the very least the parameters, the training code, a description of the data and the procedure followed: without those elements, nobody can reproduce the object, so nobody can really study it. It is this impossibility of rebuilding that sets the AI case apart, and that makes the word so contested.

Three families of licence

The licences actually encountered fall into three families. A few models are released under a recognised free licence, permissive, with no restriction on use. Many are released under a bespoke licence, which permits broadly but forbids certain domains or sets a threshold beyond which a commercial agreement becomes necessary. Others again are reachable only through a service, with no release of the parameters. These three regimes do not carry the same legal consequences, and confusing them exposes you to discovering late that an intended use was not permitted.

The habit to acquire

The practical consequence comes down to a habit to acquire: read the licence before repeating the word the publisher used. The useful questions are few. Is commercial use permitted with no threshold? Is the intended domain among the exclusions? Is a notice of origin mandatory in the final product? Are the outputs of the model subject to restrictions? Those four answers determine what an organisation can do, far more reliably than the label placed on the announcement page.