My question is somewhat similar to this one, but I am interested in a case where there is no explicit allowance for models that are unable to reconstruct their training data.
To be concrete let's focus on training data licensed with creative commons v4 (BY, BY-NC and BY-ND.)
Suppose I create a machine-learned model by training it on tens of thousands of materials made available by these CC licenses. I then use the model and its outputs for commercial purposes.
- Have I made use of the training material in the BY sense?
- Have I made commercial use of the material in the BY-NC sense?
- Are the model or its outputs derivative works in the BY-ND sense?
It is not possible to recreate any of the training materials from the model or its outputs. A typical person familiar with the training materials would not be reminded of any specific one by the model or its outputs.
EDIT: As a rider that only occurred to me afterwards, would the model and its outputs fall under fair use?