If I learn to write news articles by reading the NYT, I’m not then able to duplicate myself infinitely and fill every role at every news publisher. I am but one human, producing one human’s output and consuming one human’s space within whatever pursuit I undertake. Critically, I still leave room for others to do the same thing.
Eventually I also die and then someone else can come along to replace me. I’m finite, AI is not. It doesn’t get used up or retire.
If you consider that there’s a fixed amount of societal capacity for whatever undertaking is in question (news journalism, art generation, etc.) then as a human, I take up only a certain amount and only for a certain amount of time. I will never arbitrarily duplicate to produce the work of 10, 100, 1000, etc. humans. I will also shuffle on after about 50 years and someone else, having potentially learnt from me, can now gainfully exist within the world in my stead.
The capacity for infinite commoditisation that AI brings is necessarily a critical distinction to humans when it comes to considering them performing equivalent functions. They must be treated differently.
This is a luddite argument that can equally apply to any automation. A robotic arm can be trained to do the same thing a human line worker does, but the robotic arm can be copied infinitely and work 24/7 leaving zero room for other humans to do the same thing. Should we ban robotic arms?
Nowhere did the poster say that is “sufficient” reason to ban ai. They were clarifying how software is different from humans and only that. You need to go up a couple of comments and combine this explanation with the other part of copyright infringement concerns to see why the “whole” thing is concerning for the news industry.
And that is empty statement.
If humans are contextually special here, a "passing the torch" argument seems unconvincing, to me.
What is the purpose of laws at all if humans aren't special?
> If something takes in information, learns from it, and spits out information as a result of its learning, then why does the implementation matter at all?
Yeah, so who cares if the implementation is human and if that implementation breaks?
I really don't want to troll you, I believe it is worth it to point out the absurdity of the "humans aren't special" argument this way.
Humans are not machines and machines don't have human rights.
And humans using AI as a tool changes this fact?
Yes. It's an object or writing. Not a person. You're arguing for giving personhood to software right now and that's crazy.
I'm not sure anyone in the thread is actually arguing that. I think what they are saying is that we should look at what behaviour is considered acceptable for humans in order to help us decide what behaviour is acceptable for the tools humans use.
(I heavily edited this comment after realizing I could make the point in far fewer words. Sorry.)
It's not. That was the point of my replies. That it's time to assert software personhood is crazy.
> If we decide that a certain kind of usage of software shouldn't be considered acceptable, then we could sue the user who used it, or the developers who created it, or something.
We can already do that. That's what this article is about. The people are being sued. That's what all of my replies are about. I don't understand why you are replying to my comment with a re-summary of my comments as if it's a rebuttal to them.
It doesn't. And all of my comments are about that. Like I just said in the previous reply. You're replying to my comment where I also just said this. Please stop replying to me saying that I'm saying that.
The article is about people (well, companies) being sued. Not about software being sued. Software can't be sued.
Whether or not there are additional laws written about what's acceptable behavior for software (whatever that means? It's assuming software can make decisions) is irrelevant. You can't sue software. People are being sued because the plaintiffs think that people broke people laws and are liable for damages. Software can't break laws and can't be held liable.
I'm having to reword this over and over because you keep replying to me. I think you might be replying to me repeatedly just to have the last word.
What would be your financial responsibility to keep AI running?
No, arguing that a human using an artificial brain instead of their own biological brain for learning and derivative creation is an implementation detail of little relevance, and dismissing personhood related arguments like "software can't marry" has nothing to do with arguing for the personhood of software. Explain how one leads to the other, because I'm not seeing it, and that's not at all what I was attempting to communicate.
There is a possibility that you're not saying that, but it's the only interpretation of your comment I could come up with. Because your comment consists entirely of comparisons of software to human brains about whether or not something should be considered legal, and this only makes sense if the software itself can be held liable.
Respectfully, I don't know how you're interpreting it that way. Until we demonstrate that the current generation of AI is genuinely intelligent, instead of clever algorithms, a piece of software is no more or less liable than an individual firearm is after it's been fired at someone. My observation is that your argument appears to be that there is something special about humans learning and creating derivative works from that learning over humans using a tool that does the learning and create derivative works.
> There is a possibility that you're not saying that, but it's the only interpretation of your comment I could come up with.
That's fair. I just don't get it.
> Because your comment consists entirely of comparisons of software to human brains about whether or not something should be considered legal, and this only makes sense if the software itself can be held liable.
Human brains and software are both tools. The question I'm invoking is what is it about a person doing the learning and the derivative creation that's different from a person (since, as you say, software itself has no personhood) using an artificial brain to learn and perform derivative creation.
I think the disconnect here may be that I'm operating from the assumption that of course there are human beings liable for the software, but your interpretation of what I'm saying is that software in a vacuum should effectively have personhood applied to it. These are two different things. I'm referring to both humans/brains and software as the interchangeable variable in the question of why the choice of tool means applying entirely different legal principles.
Sorry if I wasn't clear or still am not being clear here. I wanted to make sure I was being understood correctly, but if all we can do from here is agree to disagree, that's fine, and I'd offer to just shake hands.
There is one other way I have figured out to read your comment. Which is that it doesn't matter how software or a brain functions since it's only the action of the outcome that matters. But this is not really a relevant statement regardless of whether or not you agree with it, because the article is about a lawsuit and liability. A group of people, acting as a company, is suing other groups of people as companies. And software is not a person, and can't be held liable. So for that to be the case, the software would need to be made into a person, or equivalent. The fact that software and brains are or are not similar is irrelevant, because software is not a person and cannot be held liable.
If we were so eager to give personhood to corporations and now to software can we finally give it to other animals as well?
Legally, the implementation may not matter at all, but the scale does.
The precedent is absolutely clear in legislation, for just about any category of crime or civil tort you can think of.
Just one example: you get caught with a single joint (in a place where it isn't legal, of course) you're looking at a fine ... maybe. Most places have had exemptions for possession of a single joint.
You get caught with 225 tons of weed all processed and packaged in a warehouse, you're going to jail!
You need to justify why you believe that, in the case of LLMs and AIs, an exception should be made so that the scale is not considered.
I haven't seen any justification why the justice system should make an exemption for LLMs when it comes to scale.
Scale matters, and has matter in every modern 1st world jurisdiction going back hundreds of years.
You want to overturn that? Provide an argument other than "If it doesn't matter when a single article is used to learn, it shouldn't matter when billions of words are used in the learning."
If you're going to draw a sharp distinction between reading/words/people (software isn't people, etc.) and software I would argue that the legal and copyright considerations should be stronger for software, not weaker, precisely because of your argument.
it's not. Why isn't photoshop being sued as copyright infringement? People can reproduce infringing works in it easily enough.
How is pressing buttons in photoshop different to pressing buttons in chatGPT?
It's not! That's the entire point. The liability is with the people, not the software. The software isn't being sued! You can't even do that, because the software isn't people. Software isn't humans. The people developing ChatGPT incorporated other people's copyrighted works into their product, and they're the ones being sued.
https://www.adobe.com/legal/dmca.html
Just because you produce an infringing work doesn't mean it will be discovered and then a legal case made.
Would NYT be suing Adobe (the people who made the software), if one of their users utilized Photoshop to produce an image that violates copyright? I don’t think so.
My comparison addresses exactly that, as Adobe is the company that made the software (in this specific example, Photoshop).
You cannot retroactively switch up your argument to something entirely different and then claim that my comparison was incorrect and bad.
I don't owe you any further replies, but I wanted to make this clear to anyone else skimming the thread.
In all three of your transformation examples, you likely would’ve paid the NYT either indirectly via ads or directly via a subscription (contentious ad-blocking aside). And Mazda, GM, and VW surely generated far more value from the car designer in revenue than the designer got in salary and knowledge. Broadly, humans pay something (tuition, apprenticeship, etc) to learn knowledge and skills. Even schools and libraries are not free, though they may feel so at time of use. _Some_ free learning is beginning to emerge on the internet, but much of it is as upsell to paid learning.
OpenAI on the other hand has been purely extractive in its learning relationships. A single human contravening normal learning costs would go unnoticed but also would not scale their own downstream impact. But an organized large group doing so would certainly run into issues (10k interlopers couldn’t just sneak into a single college course, for instance). And I think a large group of aligned/associated people is more akin to AI than a single person is, given the differences in scalability.
It’s buried a bit by the lede but IMO the more fundamental question is whether AI models can be trained off of publicly available, copyrighted content without compensating the owner of said copyright. This is regardless of how the eventual work is presented and how “different” it is.
It’s where this falls on the “fair use” spectrum. This is uncharted territory for any sort of fair use doctrine as written.
1) You cannot prove that this does not happen in some form. The fact that manual electrical stimulation of the cortex triggers random memories is sort of the proof that you are wrong here; if there wasn't some kind of "copy, transformed somehow into a 3D space we call the brain" in there, then that localized stimulation wouldn't trigger a particular memory at all.
2) And yet, this is not true of AI's either. Knowledge encoded into either real neurons or a simulation of them may be similar both in how it's not a "literal copy" but a kind of copy nevertheless.
If we are only concerned with "literal" or "verbatim" copies, then there is no case here.
And that's seems to be the problem. Apparently ChatGPT was spitting out entire copies of NYT articles without attribution and falsely attributing things to NYT it had made up.
If I read a lot of New York times and then go and write a political thriller, that's fine.
If I read the New York times and then go and report the same news, but in my own voice, that's debatable.
If I take the New York times article and publish it verbatim, that's a problem.
All of these cases are ultimately on a spectrum - and the answer will go beyond AI I think. Tesla's lead designer previously worked at Mazda, GM and Volkswagen - so does Tesla owe these companies money because their designer learned car design there? This might sound silly, but it's a fair argument when it comes to AI - how much distance does there need to be between the learned from material and the output?