If we ask a lawyer they will probably cite precedence such as Authors Guild vs Google (https://towardsdatascience.com/the-most-important-supreme-co...) and thus this entire hypothetical interpretation of the legal framework is already not how the legal institutions see it.
If the courts are to set new precedent I think it's important to consider all the downstream ramifications, and I think it's a lot more complex and challenging than a lot of people here seem to think. There is a lot more to AI than just generative neural networks. A lot of "boring" technology we all take for granted can be caught up in it.
> If we ask a lawyer they will probably cite precedence […] interpretation of the legal framework is already not how the legal institutions see it.
Well, now this is an entirely different discussion, and FYI "precedence" is only a thing in half of the world's legal systems. Specifically, the common law (English) half. The other half, civil law (French) based systems, have no concept of "precedence"; verdicts from other courts have absolutely no law-like meaning. For every case and every judge, they are supposed to find the correct, applicable meaning of the laws as written by the legislative.
Personally speaking, I find the "precedence" approach taken by English / Common Law incredibly silly — and actively harmful, it intermixes two branches of power (legislative & judicative) that should be 100% separated. Judges' interpretative rulings should not have (almost) the same effect as the legislative passing a law.
(Google "Common Law vs Civil Law" for more info.)
Anyway the original argument was that you could apply some license-like terms onto spam mails, and for that — no, you very much can't. The situation for AI is, to my knowledge, still very muddy at this point.
> I find the "precedence" approach taken by English / Common Law incredibly silly — and actively harmful, [...] Judges' interpretative rulings should not have (almost) the same effect as the legislative passing a law.
Precedents are visible. They're the outcome of prior cases. If the people or the legislature doesn't like the rulings they can look at the judges' reasoning and fix the law, invalidating the old precedents at the same time.
> Anyway the original argument was that you could apply some license-like terms onto spam mails, and for that — no, you very much can't.
By the act of giving you the email they're implicitly giving you permission to do email things with it - read it, forward it, store it, etc. But you don't own the copyright and can't create and publish derivative works, etc.
We never questioned the anti-spam use because it's obvious. You aren't storing data for the purpose of recreating the spam, you're storing details about what spam looks like for the purposes of recognizing more of it.
> The situation for AI is, to my knowledge, still very muddy at this point.
The question is if the NN weights in an AGI are materially different than the NN weights in a spam filter.
> If the people or the legislature doesn't like the rulings they can look at the judges' reasoning and fix the law,
That's exactly the point. In common law, when a precedent is cited in a later case, it is (like a law) largely protected from "reasoning about". You need to involve the legislature to change it. In civil law, other cases are of course also invoked as references - but not law-like, they're just shortcuts in transferring prior reasoning, and fully open to challenge. Unlike laws.
(But this is really off-topic here anyway.)
> By the act of giving you the email they're implicitly giving you permission to do email things with it - read it, forward it, store it, etc. But you don't own the copyright and can't create and publish derivative works, etc.
None of these things come about from something written in the e-mail. They are that way because it is an e-mail. If you want to tack on other semantics, that's an entirely different thing.
> The question is if the NN weights in an AGI are materially different than the NN weights in a spam filter.
No, that's completely besides the point. The question is whether NN weights trained on data that you received, in this case without any agreement, are materially different from NN weights trained on data that you crawled and that had "you may retrieve and use this data under terms XYZ" restrictions attached. It legally very much matters whether the data got thrown at you or whether you went looking for it on your own accord.
I very much disagree that you can't specify license terms on spam. Lawyers certainly seem to think you can, as they always have a huge legal blob at the bottom of their emails about what you can or can't do with their email.
Lawyers are trained to lie. Unless they're specifically prohibited from lying (such as to a judge), lawyers will lie if it gains them an advantage. Do you think every attorney who claims at a press conference, "The facts will vindicate my client!" is telling the truth?
Even when lawyers are prohibited from lying, they are trained to and expected to mislead.
The "confidentiality" blocks in e-mails are completely unenforceable, unless there is a separate contract which it is included under.
"Automatic e-mail footers are not just annoying. They are legally useless"
Looking at / searching around this with a wider lens, in some cases the footers seem to serve a function in clarifying the intent of the mail (e.g. "this mail does not establish an attorney-client relationship") when the remainder of the mail may be unclear. But that's not a license or contract, that's a clarification of intent.
And with this I'll take my leave from this discussion as it no longer feels fruitful. But Thanks for the interesting thought exercise!
If the courts are to set new precedent I think it's important to consider all the downstream ramifications, and I think it's a lot more complex and challenging than a lot of people here seem to think. There is a lot more to AI than just generative neural networks. A lot of "boring" technology we all take for granted can be caught up in it.