News for hackers. Stuff that matters.
Re:OpenAI's justification for why training data is fair use,... by thomastjeffery on Thursday October 05, 2023 @01:37PM (#37781492)
I'm all about getting rid of copyright.

This is...bullshit.

> patterns inherent in human-generated media

That's what you hope your model is, not what your model literally is!

> The aim of this process

That's a lovely thought. Now let's get back to talking about the process itself.

I'm not interested in treating LLM tech like a black box. That's intentional stupidity.

OpenAI's entire argument hangs on not understanding how its tech works! I can't see any difference between that and fraud.

Re:OpenAI's justification for why training data is fair use,... by devmor on Thursday October 05, 2023 @01:40PM (#37781532)
>OpenAI's entire argument hangs on not understanding how its tech works! I can't see any difference between that and fraud.

This is how I see it as well.

If you can't explain, in detail, how your mechanism is not violating copyright then you have no defense.

Re:OpenAI's justification for why training data is fair use,... by Tadpole9181 on Thursday October 05, 2023 @01:48PM (#37781662)
That's not how law works in the US. You are presumed innocent and there must be a preponderance of evidence of guilt.

If they can't prove the AI model contains or reproduces exact copies of their copyrighted work, there's no case to be had.

Re:OpenAI's justification for why training data is fair use,... by thomastjeffery on Thursday October 05, 2023 @02:07PM (#37781908)
But we can. There are countless documented instances.

OpenAI describes these as "limitations". A more accurate word would be "features".

Re:OpenAI's justification for why training data is fair use,... by skeaker on Thursday October 05, 2023 @02:53PM (#37782558)
Is this true? As far as I am aware there are zero pixel for pixel recreations by an image generating AI of any pre-exisiting images and no written work has ever been word-for-word reproduced in full. The closest I've seen are images highly similar to the Mona Lisa which is already copied virtually everywhere without AI and brief passages from commonly copied written works that veer into hallucinations after a paragraph or two.
Re:OpenAI's justification for why training data is fair use,... by thomastjeffery on Thursday October 05, 2023 @05:51PM (#37784744)
Why does it need to be exactly the same? You have shifted the goalposts to absurdity.
Re:OpenAI's justification for why training data is fair use,... by skeaker on Friday October 06, 2023 @12:39PM (#37792995)
Because the parent comment says "But we can" in response to the claim that AI can't generate "exact copies." Generating exact copies would be a lot more legally notable than outputs that are only similar or partially copied. As far as I know exact copies haven't happened and was legitimately asking for cases where it has.
Re:OpenAI's justification for why training data is fair use,... by devmor on Thursday October 05, 2023 @01:51PM (#37781714)
No one said otherwise. There is already a case against them, their defense is that they don't know how it works, so it can't be true. This was clearly laid out in the comment I replied to, and the comment that reply was made to in turn.
Re:OpenAI's justification for why training data is fair use,... by blibble on Thursday October 05, 2023 @02:44PM (#37782403)
> That's not how law works in the US. You are presumed innocent and there must be a preponderance of evidence of guilt.

true for criminal cases but this is a civil?

so it's balance of probabilities, not beyond reasonable doubt

Re:OpenAI's justification for why training data is fair use,... by dcow on Thursday October 05, 2023 @03:28PM (#37783130)
civil cases is preponderance of evidence.

criminal is beyond a reasonable doubt.

in any light, yes civil cases are about whether you’re more likely than not to be liable