Or similarly, if a company trains a model and then bins the training data. How would you verify any artists work was in there?
Your first question is based on a technicality (human can learn too, should we then …?), but the idea of right is not based on physical properties or similarities, but on realistic desired outcomes of its application. (Of course we see flaws of laws all the time, but that is tangential.)
company trains a model and then bins the training data. How would you verify any artists work was in there?
Like any shady company it will drag a set of risks with it forever. It may fly under the radar for some time, but will be snitched on after few internal disagreements. This well-known dynamic disincentivizes big companies from choosing going this route, and small companies are not a big deal.
As others have already mentioned here before:
- a "style" isn't copyright-able
- Looking at how others have done art, imitating them to learn, and then combining what we learned with everything else we've seen in our life is how humans produce art. And it's the same way that stablediffusion produces art: train on data, reproduce based on what it's learned, combine everything its seen with a prompt to create something new.
This is different to most copyright cases where all that matters is the end result, not the process.