Hot take
Time for another #hotTake - #github #copilot does not infringe GPL licenses. In any #copyleft #license you have permission to learn from that code, and you can use that learning elsewhere without having to respect the license.
#AI is just learning from these. Of course, if you force it by crafting an attack prompt, it is likely to give you a literal copy. But that should be hard, almost impossible in regular usage. And yet, we should wonder if copying a few lines infringes GPL.
Hot take
@deavid could the model the AI generated from the training data not be considered a derived work? And thus the output of the AI a derived work?
I also don't see an explicit grant in the GPL v2 or v3 for ' learning', what clause gives you this impression?
I genuinely think there will be some interesting test cases over AI and copyright. It won't be a short or pretty road.
Hot take
@intrbiz for me, it's the end user responsibility to use it fair. If the results are copying a lot and causes a legal issue, I don't think that GitHub would be claimed as responsible. It most probably would be claimed as end user fault.
Hot take
@intrbiz if it's derived work, then I cannot work anywhere because I learned from open source code, therefore all the code I produce is derived work and must be GPL, this unsuitable for confidential/private use inside a company.
#AI doesn't mash things together. It learns the relationship and builds from there. Unless we're talking about dumb AI, something at GPT-3 level should be considered original work (also with the exception of bad tunning with overfitting).
Hot take
Interesting point. Which could well hinge on the definition of 'use'.
This seems to usually be tested in court by how similar the outputs are to other existing works, often siding with the rights holders of the existing works.
I don't think that people will accept the view that a machine could be as 'lossy' as a human.
Hot take
@intrbiz I agree. The problem is that most people aren't going to take a machine as lossy or creative. This reminds me of the movie "bicentennial man"
Hot take
Yes, plus also sentiments echoed in "Humans".
Hot take
@deavid
Should the developer of the algorithm be considered to have a right to the resultant work?
Should the rights holders of the training data set be considered to have a right to the resultant work?
Or should the 'artist' who combined them and generated the work?
I don't think there is an easy answer here.
But I do think it's too simplistic to say that the licences of the works used to train the AI model can be ignored, which seems to be Github's current stance.