There is simple way to fix this.
Ban private large models trained on public data, require them to be public weights.
If a company wants to train large private model, they can do it with their own data.
Ban private large models trained on public data, require them to be public weights.
If a company wants to train large private model, they can do it with their own data.
But it seems unfair that a company can own such a model. It's not their work, it's a codified expression of basically our entire cultural output as a species. We should all own it.
Maybe, just maybe better to ask artist, who should own it?
We can fix it. Lets create "for-ai-training" tag. With default value "false", so, when tag is absent you can't use text and images for ai training.
The issue here is models generating copy written work verbatim.
People claiming that training on copyrighted work is a violation of copyright (its not) have no legal legs to stand on. They are purposelessly muddying concepts though to make it seem like it is. However any competent judge is going to see right through this.
To draw a parallel in software, we have MIT licenses that allow for-profit, private use of source code in the public. The copyleft license might be more aligned to what you are envisioning?
You can't ban it for open source either.
Real world policy is rarely black and white, all or nothing. An inability to prevent every instance of something is not a reason to forgo any action
Just because country X has slavery, doesnt mean the US can't ban it. There may be some consequences, and disadvantages to doing so, but that is a trade off, not a show stopper.
Same thing with open source. There is lots of illegal and bootlegged content that is shared on the internet, and cant be eradicated. That doesnt mean it can all be freely bought at Walmart.
They would be better off if you did it the other way around - ban large public models.
And I am in favor of more stringent licenses than that: you must pay me to use my data.