I've recently noticed this opinion seems unpopular, at least on Lemmy.
There is nothing wrong with downloading public data and doing statistical analysis on it, which is pretty much what these ML models do. They are not redistributing other peoples' works (well, sometimes they do, unintentionally, and safeguards to prevent this are usually built-in). The training data is generally much, much larger than the model sizes, so it is generally not possible for the models to reconstruct random specific works. They are not creating derivative works, in the legal sense, because they do not copy and modify the original works; they generate "new" content based on probabilities.
My opinion on the subject is pretty much in agreement with this document from the EFF: https://www.eff.org/document/eff-two-pager-ai
I understand the hate for companies using data you would reasonably expect would be private. I understand hate for purposely over-fitting the model on data to reproduce people's "likeness." I understand the hate for AI generated shit (because it is shit). I really don't understand where all this hate for using public data for building a "statistical" model to "learn" general patterns is coming from.
I can also understand the anxiety people may feel, if they believe all the AI hype, that it will eliminate jobs. I don't think AI is going to be able to directly replace people any time soon. It will probably improve productivity (with stuff like background-removers, better autocomplete, etc), which might eliminate some jobs, but that's really just a problem with capitalism, and productivity increases are generally considered good.
There are multiple facets here that all kinda get mashed together when people discuss this topic and the publicly available/public domain difference kinda gets at that.
An AI model could be seen as an efficient but lossy compression scheme, especially when it comes to images... And a compressed jpeg of an image is still seen as a copy so why would an AI model trained on reproducing it be different?
Are you suggesting that the model itself is a compressed version of its training data? I think it requires some stretches of how training works to accept that.
It depends on how much you compress the jpeg. If it gets compressed down to 4 pixels, it cannot be seen as infringement. Technically, the word cloud is lossy compression too: it has all of the information of the text, but none of the structure. I think it depends largely on how well you can reconstruct the original from the data. A word cloud, for instance, cannot be used to reconstruct the original. Nor can a compressed jpeg, ofc; that’s the definition of lossy. But most of the information is still there, so a casual observer can quickly glean the gist of the image. There is a line somewhere between finding the average color of a work (compression down to one pixel) and jpeg compression levels.
Is the line where the main idea of the work becomes obscured? Surely not, since a summary hardly infringes on the copyright of a book. I don’t know where this line should be drawn (personally, I feel very Stallman-esque about copyright: IP is not a coherent concept), but if we want to put rules on these things, we need to well-define them, which requires venturing into the domain of information theory (what percentage of the entropy in the original is part of the redistributed work, for example), but I don’t know how realistic that is in the context of law.