> While this likely has no legal weight

I wouldn't be quite so sure about that. The AI industry has entirely relied on 'move fast and break things' and 'old fart judges who don't understand the tech' as their legal strategy.

The idea that AI training is fair use isn't so obvious, and quite frankly is entirely ridiculous in a world where AI companies pay for the data. If it's not fair use to take reddit's data, it's not fair use to take mine either.

On a technological level the difference to prior ML is straightforward: A classical classifier system is simply incapable of emitting any copyrighted work it was trained on. The very architecture of the system guarantees it to produce new information derived from the training data rather than the training data itself.

LLMs and similar generative AI do not have that safeguard. To be practically useful they have to be capable of emitting facts from training data, but have no architectural mechanism to separate facts from expressions. For them to be capable of emitting facts they must also be capable of emitting expressions, and thus, copyright violation.

Add in how GenAI tends to directly compete with the market of the works used as training data in ways that prior "fair use" systems did not and things become sketchy quickly.

Every major AI company knows this, as they have rushed to implement copyright filtering systems once people started pointing out instances of copyrighted expressions being reproduced by AI systems. (There are technical reasons why this isn't a very good solution to curtail copyright infringement by AI)

Observe how all the major copyright victories amount to judges dismissing cases on grounds of "Well you don't have an example specific to your work" rather than addressing whether such uses are acceptable as a collective whole.

> but have no architectural mechanism to separate facts from expressions

Sure they do. Every time a bot searches, reads your site and formulates an answer it does not replicate your expression. First of all, it compares across 20.. 100 sources. Second, it only reports what is related to the user query. And third - it uses its own expression. It's more like asking a friend who read those articles and getting an answer.

LLMs ability to separate facts from expression is quite well developed, maybe their strongest skill. They can translate, paraphrase, summarize, or reword forever.

'old fart judges who don't understand the tech'

If this intended to refer to Judge Alsup, it is extremely wrong.

> The idea that AI training is fair use isn't so obvious

> Observe how all the major copyright victories amount to judges dismissing cases on grounds of "Well you don't have an example specific to your work" rather than addressing whether such uses are acceptable as a collective whole.

Well, all a judge can/should do is to apply current law to the case before them. In the case of generative AI then it seems that it's mostly going to be copyright and "right of publicity" (reproducing someone else's likeness/voice) that apply.

Copyright infringment is all about having published something based on someone else's work - AFAIK it doesn't have anything to say about someone/something having the potential to infringe (e.g. training an AI) if they haven't actually done it. It has to be about the generated artifact.

Of course copyright law wasn't designed with generative AI in mind, and maybe now that it is here we need new laws to protect creative content. For example, should OpenAI be able to copy Studio Ghibli's "trademark" style without requiring permission?

> The very architecture of the system guarantees it to produce new information derived from the training data rather than the training data itself

A “classical” classifier can regurgitate its training data as well. It’s just that Reddit never seemed to care about people training e.g. sentiment classifiers on their data before.

In fact a “decoder” is simply autoregressive token classification.