Hacker News Reader

Comment by ang_cire | Hacker News Reader

Something that is very disturbing in seeing people compare AI models to humans is the seeming implication by many of them that humans only have rights because a law says they do, and that there are no rights inherent to being a human (which AI models are not).

Arguing that because a human can do something, an AI model should be allowed to as well, shows a remarkable lack of value of humans.

Humans are more valuable than AI models, and should retain many inherent rights that AI models should not.

AI models are just pieces of software, they're not actual intelligence or sentience or life. And companies using them are just profit-driven companies, not benevolent child-rearers of a new lifeform.

This entire case is just a corporate power-grab seeking to profit from the work of others for free without legal repercussion, and it's sad seeing so many people jump to support the corporations.

View on HN

> Humans are more valuable than AI models, and should retain many inherent rights that AI models should not.

This argument feels disingenuous. There aren't people out there trying to get AI votership or property rights. Even the most ardent AI supports will agree that AIs shouldn't automatically enjoy the inherent rights we enjoy as human beings. This is a pure strawman.

The issue at hand is copyright laws, and IMO the conversation is a necessary one. With or without AI, a significant amount of people have been wondering whether or not the current set of legislation is effective in enforcing the spirit of the copyright law, which is to incentivize and protect original ideas. The fact that we have AI's that can learn and iterate at blazing speeds was just a catalyst that exposed some large holes in our legal system - the holes were always there.

The question now is where do we draw the line for AI? Even if everyone in the world agrees that "Humans [...] should retain many inherent rights that AI models should not", we still need to decide where that line should be drawn.

This is honestly the best response so far, and I agree that my statement was probably too ambiguous.

I am not claiming that anyone says AI should have all the rights of humans, I'm saying that people are arguing that:

-> the Right existing for humans (to organically learn by ingesting works of others)

-implies that-

-> learning must also then be allowed for non-human learning (i.e. AI models) as a Right

That argument does not make sense to me unless you start from the assumption that "learning" is what grants the Right (i.e. that everything has a right to learn, computers included), rather than being human, which is what I mean by "devaluing" people.

I agree the court case is necessary, but only as a vehicle to enshrine what should be an obvious truth:

Humans learn from experience whether they endeavor to or not, so they literally cannot ingest media (e.g. for the purposes of entertainment that the creators intend, per OpenAI's brief) without assimilating it into their mind.

That is not true for computers. AI models do not consume content for the purposes set out by the creators. They consume them for business purposes, and to allow them to do so, without license or permission, completely destroys copyright protection against corporations.

We do not need a new law; the current law already doesn't allow companies to use copyrighted works without permission. We just need the existing law enforced against AI model companies.

It's the other way around. Arguing from similarity to humans is a counter-argument to the idea that a process similar enough to a human practitioner[0] should be covered by IP laws as if it was regular piece of software - IP laws that are mostly archaic and already stretched to near-breaking point to cover computers in the first place. It's not about whether or not humans have inherent rights. It's about logical consistency.

This is perhaps a "failing"[1] typical to STEM/tech people: expecting laws to be consistent. Consistent in a way similar to mathematics, in that for any ruling you can trace the arguments back to underlying interpretations of laws, and trace those down to written rules, and trace those to some more fundamental rules, and eventually to some sort of intuitions about morality and fairness. And at least to me, arguments that generative network training ought to be treated like compiling and obfuscating a regular dataset of copyrighted works, is justified merely by "it's different because computers".

This is not to say that current models are learning and creating art and text the same way humans are, but rather that the process just seems to be close enough. And the point of mentioning logical consistency is this: I may agree with copyrighting the living hell out of LLMs for pragmatic reasons[2] - because it shakes the boat too much, has potential to destroy livelihoods of too many people in too short a time, while further centralizing power, etc. Those are all valid arguments. "Because it's a computer", to me, isn't. Not when the process and effects are already eerily similar to how humans work. Not when such argument would apply just as much in a hypothetical future where we develop sentient AIs[3].

[0] - In very limited scope, but also the very scope subject to legal issues.

[1] - I don't really consider it a failing. There's both beauty and efficiency in things making some kind of sense.

[2] - Intellectual property laws themselves are mostly pragmatic in this sense anyway.

[3] - Even in this future, we'll have to face pragmatic issues. I read an interesting take on this long ago, I think in one of Eliezer's essays: how do you handle democracy, fair resource allocation, basic ethics, in the presence of AI people that exist in silica? Such AIs will likely be able to reproduce much faster than humans - bringing new individuals into existence at the speed of factories pumping out GPUs. Equality and democracy are all fine until suddenly there's 10 trillion of AI people, and only 10 billion of human people, and everyone gets an equal vote. How do we deal with that?

> the process (of human vs AI learning) just seems to be close enough

It only seems that way because the people who understand it best also stand to profit from it, and are choosing to misrepresent it so that people who don't understand it falsely believe it is at all close.

That so many people on here are pushing the false narrative that they are at all similar is sad, since this is ostensibly a forum for people with technical knowledge. Nothing about machine learning is similar to human learning, by any measure of evidence. If I am wrong, show me that evidence.

"Data go in, something happen, metadata come out" is the closest you can approximate the 2, and too many supposedly tech-savvy people seem content to treat that as "close enough".

> "Because it's a computer", to me, isn't. Not when the process and effects are already eerily similar to how humans work. Not when such argument would apply just as much in a hypothetical future where we develop sentient AIs.

The argument is not "because it's a computer", it's "because it's not sentient", as I mentioned in my comment. Animals also have rights, because they are living.

ML models are none of those (living, intelligent, or sentient), and the popular conflation of ML models and "AI" with "AGI"/ the kind of AI that may or may not ever end up existing, is just a convenient excuse by companies looking to maximize profit and minimize licensing costs, not to regulate the plain old python that we're actually talking about now.

> that humans only have rights because a law says they do, and that there are no rights inherent to being a human

It may be disturbing, but that is the nature of reality. We humans indeed have no inherent rights, only those written in law. Sometimes the law says words like "these are universal and apply to all", but thinking those words really have that power because they say so is like thinking Harry Potter has to be a true story because none of the characters say "I'm fictional" when under the influence of Veritaserum.

> AI models are just pieces of software, they're not actual intelligence or sentience or life

Life is just a collection of chemical reactions.

Putting "intelligence or sentience or life" next to each other like that implies you think these are connected.

I don't think "life" is a requirement for either of the other two.

If the abilities demonstrated thus far by AI do not meet your standard for the word "intelligent", what does?

If intelligence were a prerequisite for sentience, would that be with a hard cut-off of "you must be IQ ≥ XYZ to be sentient"? Or do you suspect that sentience is a variable, with less intelligent systems also being less sentient, but also no cut-off? If there is a cut off, where? Why?

When I was a kid, we could happily assume that intelligent computers were pure sci-fi; now that models like GPT-3 have a complexity similar to the brain of a medium sized rodent — the scale may be important or may be a red herring, but we should take the questions seriously and try to come up with actual answers to each before it matters.

I'd rather treat a mindless automaton with unnecessary kindness than treat one that is able to suffer with cruelty; and on the off-chance that full-brain-uploading becomes a thing, I'd want to know for sure if an upload of me could or couldn't have qualia.

> This entire case is just a corporate power-grab seeking to profit from the work of others for free without legal repercussion

Well, some sure is. Some's given away for free, so it's not capable of being a power-grab.

> a corporate power-grab seeking to profit from the work of others for free without legal repercussion, and it's sad seeing so many people jump to support the corporations

lol who do you think posts on HN? a tech startup blog run by an incubator. half of us are those people, or aspire to be.

I'm a Bay Area startup alumni, so closer to the actual VC scene than most on here, and all the actual successful founders I know have a pretty strong belief in the good that their products will bring, even if they usually overstate the likely impacts quite a bit.

"Move fast and break stuff" != "...and run people over in the process". Are there greedy a-holes in every space? Of course. But if you think that's the default mindset at startups, I don't feel like you're going to have much success in the space.

Do I think OpenAI probably thinks that their work is less impactful to Steven King's sales than he is claiming? Absolutely.

Do I think they're lying to themselves about the actual negative impacts of their work because they think their product is really cool and has so many potential uses? Absolutely.

Do I think they secretly all just want to be the guy with the boot, crushing artists' livelihoods and pushing for regulatory capture, like you seem to be implying they (and others here) are? No.

A shoe has no rights. No right to loiter, enter an airport, or climb a mountain.

Nor does an AI.

A human has a right to use a shoe. Or an AI.

I see no reason one tool is inherently profit-driven while the other benevolent.

(An exercise left to the reader: whether leather or pages are the more renewable and humane resource.)

> A human has a right to use a shoe. Or an AI.

What right is that, exactly? I don't remember the "Right to use a tool".

You are trying to make "using a tool in a way that harms people" have greater legal or ethical weight, by falsely calling it a right, than the harm they're doing.

The argument was not "human can do it so should machine qed". The fact that humans so easily do something is a talking point in a large subthread, not the conclusion.

Yes, an AI model is just a digital machine that is presently hard for many people to distinguish from magic.

> This entire case is just a corporate power-grab seeking to profit from the work of others for free without legal repercussion, and it's sad seeing so many people jump to support the corporations.

This is where you're wrong. There's very clearly an active and relevant dialog surrounding the concept of machine generated content and how it should be interpreted by society (now that machines are getting really good at doing things that were difficult in previous decades). So at a surface level it's hardly decided.

Further your assertion that people who use copyrighted works to create a product that provides value to society are unfairly being enriched at the expense of copyright holders is naïve and uninformed by reality. People use impression of copyrighted material all the time in their day to day lives in everything they do. I read copyrighted textbooks to learn skills required to provide value to society. I read the news all the time and make decisions based off it. So to say that OpenAI can't feed a machine the text of a popular book (one which is has legally obtained a copy of) in order to calibrate weights used to help it emulate reality is not based in any current notion of fairness even as protected by copyright today. A naïve reading of copyright would argue you had to copy the work to do that and you don't have the right to that, but a sophisticated understanding would argue that that's fair use because it happens billions of times a day and it doesn't realistically compete with the author's market for the book, interoperating with a machine is protected under the DMCA, it's a damn computer anybody using one is copying around bytes in memory when they launch their ereader to view a copyrighted work fairly all the time that's just how digital things work etc. etc. Ideas can't be copyrighted.

What is arguable is whether there's a problem if an LLM reproduces parts or all of a copyrighted work. Some people think AI should launder copyright, some think the human generating the work should be held to all the normal standards of not plagiarizing or otherwise committing copyright infringement in what they decide to do with the LLM output, and some argue OpenAI should pay a royalty every time it does (or something to that effect).

Long story short, this isn't some random corporate power grab. As a fellow human, I very much believe society can navigate the introduction of LLMs without devolving into "every piece of training data must be specifically licensed to be used as such and don't forget royalties". I don't like the society I imagine if we take that stance (frankly I don't like copyright as it is deployed today) and I believe the value LLMs provide will do more positive than negative so it's irresponsible for society to unduly limit the proliferation of this technology in a bygone effort to protect old school content creators (well really publishing conglomerates and large media corporations).

This is a lot of text just to say, "I care more about ML systems getting better than I do about whether they do so ethically".

> people who illegally (ftfy) use copyrighted works to create a product

Yes, this is something that most people who aren't illegalists agree is wrong.

> that provides value to society

This is just wholesale coolaid.

AI is a tool. The tool doesn't provide value, only the use of a tool does, when that use is positive. The use can also be negative.

This case is not about the legal permissibility of using copyrighted works to train AI. That is a misrepresentation.

This case is about the legal permissibility of using copyrighted works, against the wishes of the works' creators, to train AI, which is clearly wrong.

"What do a few individual humans matter versus the advancement of humanity?" is an argument that has been invoked many times, by the evilest people.

Thats not my argument and if you really interpreted my comment as such we are probably at an impasse.

If I write a poem and you buy a copy and feed it to your LLM and make a billion dollars I don’t think I’m innately, ethically, or fairly entitled to a single dollar of that money. I think you’re within your rights to do that since you paid me fairly for the copy and my understanding is that you can use that copy as you see fit because you own it.

I fundamentally don’t think you understand the role that ideas and the derived creative expressions thereof play in society. For humanity to progress ideas must flow freely. Attempting to put roadblocks in front of the free exchange of ideas is the evil profit grabbing motive at work. The reality is that any one individual is not entitled to collective royalties from the commons just because they voiced an idea. In my opinion that’s whats poisonous about this whole thing.

> and you buy a copy

This is the step that OpenAI and other companies are missing. They didn't pay for the content they are using.

> For humanity to progress ideas must flow freely.

AI models are not and will never progress humanity. The ability to have a computer write, draw, or vocalize for you does not change you.

You are fundamentally misunderstanding machine learning. It's just fancy python code that approximates a derived pattern. Nothing less, nothing more.

> This is the step that OpenAI and other companies are missing. They didn't pay for the content they are using.

Where's the evidence that OpenAI pirated all the content they trained models on and didn't access it normally just like everyone else? It's quite the allegation that OpenAI torrented `all-media-archive.xz` to train their model.

> It's just fancy python code that approximates a derived pattern. Nothing less, nothing more.

You are sorely mistaken if you think that's all that people are throwing billions of dollars behind.

Did the steam engine progress humanity? Did the internet progress humanity?

Obviously LLMs themselves aren't autonomously progressing humanity. But what they do as a tool to democratize (hopefully we make sure it's democratized) cheap expression already is pushing us forward and creating new space for humans to imagine and express new content that we've never seen before. You'd have to be blind or ignorant to not see that. Just go join MidJourney's Discord and see for yourself.

And it's not just art. More and more people with less knowledge of high computer science are able to complete simple scripting level tasks with the help of LLMs. Even experienced professionals are accelerating their development flows and productivity by leveraging AI to skip a lot of the mundane exploratory/discovery level research that comes before diving into a task. I personally know people who use AI code tools in their workflows today.

> Where's the evidence that OpenAI pirated all the content they trained models on and didn't access it normally just like everyone else? It's quite the allegation that OpenAI torrented `all-media-archive.xz` to train their model.

This shows a fundamental lack of understanding of copyright law on your part. Getting a PDF of The Shining by googling/ web scraping, versus getting a PDF of it by torrenting, does not make obtaining the PDF any more legal. Publicly-accessible != legal to download.

Only the grant of permission by the copyright owner or a licensed distributor(when talking about digital copies) gives you the right to obtain a copy.

There is substantial evidence (which OpenAI freely admits to) that they downloaded copies of copyrighted works without permission by the copyright holders. That makes it copyright infringement.

> Did the steam engine progress humanity? Did the internet progress humanity?

Not in the least bit.

Caveman hit caveman with stick progressing to caveman hit caveman with nuke is not a progression of humanity. Trains and the internet did not make humans any better than before, they just consolidated power into certain groups.

Mistaking technological progress with social or certainly evolutionary progress is why people roll their eyes at techbros.

"Cavemen stop hitting each other" is progressing humanity, and that won't be a result of technology.

> You are sorely mistaken if you think that's all that people are throwing billions of dollars behind.

I know exactly what I'm talking about, and that is exactly what ML is. And if your only argument is, "surely people wouldn't sink millions of dollars into something unless it was revolutionary!" you've never dealt with corporate software licensing and development.

People pay Splunk billions of dollars to run regex on their logs that they could write themselves.

People pay billions of dollars for (F)OSS just to have proserv support contracts they never use.

Billions of dollars are wasted every day across the world for B2B cruft.

ML is cool and all, but it's just code. Code running on a computer. Code you can run on your home pc. It's not magic, or inscrutable, or greater than the sum of its parts. Data go in, data come out, same as every other program.

Can you point me to where OpenAI admitted they illegally acquired copies of works for training data? I’ve always said “if you acquire the training data legally then all is fine”. I never asserted that anybody definitely did acquire it legally. If OpenAI torrented some books then they should be made to pay for those.

Anyway at this point you’re being deliberately obtuse with your tirade about how a freely flowing exchange of ideas and culture (and tools that facilitate as much) does not progress humanity. No point in continuing to argue…

> Can you point me to where OpenAI admitted they illegally acquired copies of works for training data?

Yes. It's on the first page of the linked PDF:

> For this response, we draw on our experience in developing cutting-edge technical AI systems, including by the use of large, publicly available datasets that include copyrighted works.

They then go on to try to claim this is just how everyone does it, as though that makes it okay, including making copies of the copyrighted works:

> Modern AI systems require large amounts of data. For certain tasks, that data is derived from existing publicly accessible “corpora” (singular: “corpus”) of data that include copyrighted works. By analyzing large corpora (which necessarily involves first making copies of the data to be analyzed)...

> at this point you’re being deliberately obtuse

No, I just suspect we fundamentally disagree about the value and importance of many technologies.

Well I never argued that under the current rules OpenAI shouldn’t have to acquire an initial legal copy of copyrighted works, just like libraries do. So we agree.

If I were OpenAI I’d argue that a corpus built for the training of an AI model is transformative since it doesn't detract commercially from the original works. No one wanting wanting to train a model would otherwise be using that text to read the book instead.

You are misunderstanding the purpose of transformation as a Fair Use criteria; it's there to prevent using copyright claims to quash parody and critical (as in criticism) discussion of the copyrighted work, so for instance using copyright to attack a journalist who quotes shitty writing in a book review, or a comedian who alters a quote to make fun of a Disney character, etc.

Not only is training ml models not a form of critical expression that enjoys 1A protection, but using the book itself to train a model isn't transformative; you didn't alter the book, you used the book verbatim to train on. That the model doesn't maintain the book doesn't matter; copyright doesn't only cover output (ie a derived work like said comedy skit), it also covers input (how the original work is obtained), and in this case they were not obtained legally (which means with a license for said use).

> I’d argue that a corpus built for the training of an AI model is transformative since it doesn't detract commercially from the original works

The reason OpenAI isn't arguing this is because they know that would be DOA. You cannot just package copyrighted works together or in a different way and claim that is transformative; there is actually direct case law about that because torrenters tried to claim exactly that, and lost. In fact, transformation is explicitly ruled in several cases to not apply to the entire original work being present.

Reply on news.ycombinator.com