Hacker News Reader

Comment by tinco | Hacker News Reader

Why do you assume it's a legitimate thing to do? Nothing changed from a copyright perspective. You can copy a GPL file directly, or take a photo of it, then decode the JPG, OCR the bitmap and put the result in a file, or you can train a LLM to reproduce it. It's all the same, if a judge feels you've copied a substantial part of the file then it's an infringement.

The same way an LLM would generate that code without the license, another LLM could find that code and determine that it infringes. If anything, the emergence of LLM powered agents will empower the free software movement to find infringements.

The only way free software would "end", is if judges would not be able to enforce copyright, because for example the copyright infringement is obfuscated too much. For example if LLM powered agents are going to both generate and run the code without ever saving it to disk. We're not quite there yet.

View on HN

If legimate, then failure. The precondition implies the postcondition.

I don't personally think running copyright material through a magic box that you write lots of pretty marketing about removes copyright. Much like running it through grep to delete the copyright notices doesn't remove the copyright. Or deleting the licence file from a fork.

Seems totally obvious to me, but only because I don't buy that there's an artificial intelligence composing code from scratch that just happens to have the same expletives in comments that the GPL code did.

I, nor anyone else in this thread believes that the copyright is removed from the output of the LLMs, in fact we're arguing quite the opposite.

Your point is obvious to everyone. The question is why do you think anyone else believes differently?

I'm inferring from people using GitHub copilot that they're using co-pilot. Reflexivity.

So either they believe it isn't a copyright erasure tool, or they believe it is and don't care, or they haven't formed a belief in either direction.

That case analysis looks complete to me.

Ah, right so that's only true if you believe that all output from Github Copilot infringes on copyright. Most responses from LLM's don't include copyrighted information, or at least not substantial portions of such.

It's funny how this topic seems to make people's brains and reading comprehension just switch off.

I can train a human (such as myself) to reproduce entire song lyrics, verbatim.

We have entire professions (musicians, actors) who develop this ability.

Is that, in your view, 'illegitimate' too?

Why or why not?

It's legitimate because the law says what is and isn't legitimate.

LLMs/Copilot/etc. copying other people's code is not legitimate for precisely the same reason. But big and powerful beats small and week.

Also, those laws say that remembering a song is fine. And playing it in your living room too. But that you'll have to get the licensing right if you play that song in front of an audience, in a pub or on street. At least in most countries.

To translate that to LLMs: training an LLM on songs (or code) is fine. It's the output and what you do with that output that might be problematic.

> To translate that to LLMs: training an LLM on songs (or code) is fine. It's the output and what you do with that output that might be problematic.

Only if you assume that LLMs are treated like people. Does an LLM have a right for decent working conditions? Is it illegal to make an LLM work 24h a day?

LLMs are not humans, therefore it seems reasonable to not blindly translate everything that applies to humans.

The courts often handle new technologies by default simply by assigning responsibility to the people who operated it for the specific uses in question, such as reproduction.

What the courts do is one thing, what the law should say is another one.

One problem I see with LLMs is that it is killing the job of many artists by using the copyrighted work of those very artists. The more the artists work, e.g. by doing truly original content that an LLM cannot generate because it was not trained on that, the better the LLMs get at killing the job of the artists.

It is a very weird situation: the LLMs need the artists in order to finish killing the artists. When that is finally done, it's not clear at all that LLMs will be able to generate new truly original content or whether they will be stuck in the world that the artists created before dying.

It's very different from, say, a calculator that replaced the job that real person had. Because the calculator can do at least as well as the human was doing, and the calculator was not feeding from the output of those humans.

I think the most likely outcome is that AI will eventually produce art that most humans prefer to that created by humans, and humans may not even remember a time when art came from humans.

Are you old enough to remember when humans planned in advance which video you would watch next?

> I think the most likely outcome is that AI will eventually produce art that most humans prefer to that created by humans

Well if it prevents humans from producing art, of course this will happen. Doesn't mean it will be better art. I would argue that it will be worse by definition.

> Are you old enough to remember when humans planned in advance which video you would watch next?

Very much so. But now that time is back.

Most people I know do exactly that. The time when everyone consumed TV passively is over.

AFAIK, having humans recite copyrighted songs from memory and then performing it without permission is illegal in most jurisdictions.

However, copyright law seems to make a distinction between expression/implementation of an idea (which can be copyrighted), and the idea itself (which generally speaking, cannot be). This is why clean room implementation is a legally sanctioned protocol to copy functionality without running afoul of copyright law.

Whether the language model is copying the "expression" (or a particular implementation) of a computating function, or whether it is just somehow re-implementing it given user's instructions, is going to be an interesting debate, and I really doubt people who think it's one way or the other has actually understood the nuances.

Yeah, I agree that being confident about this is a good sign you don't understand it.

I'm unsure what you're getting at, but when people sing copyright songs or act copyright plays that do, in fact, get a license. So yes, if you train a human on someone's copyright content I do expect you to have a license to it.

I'm sure you've heard of covers? Well every cover that is published affords a royalty to (at least) the original authors of both the lyrics and composition. The artist may get some money, depending on how the work was licensed.

Read what you just wrote again.

Do you see a difference between training and reproduction?

Yes, someone still needed to be trained on the original copyright sheet music before they could train someone else. Ultimately the source of that knowledge was a copyright piece of material that was licensed appropriately.

Yes someone may teach another person based on their memory, but even if that person still performs that work they still are legally required to license the work.

Yes, if the only thing they were doing was training, sure. But it’s not. They’re training and then presenting the data and given the way LLMs are trained, there is no guarantee a transformation even takes place.

At the end of the day LLMs should be licensed under the current copyright system. Maybe OpenAI need to donate some money to a few politicians for that to change.

rather infamously, “it depends!”

https://variety.com/lists/song-copyright-infringement-cases-...

There's a youtuber that makes excellent videos about this topic. Here's a longer one, but the shorter ones covering particular lawsuits are great too:

https://www.youtube.com/watch?v=MAFUdIZnI5o

It doesn't really matter what my or anyone's view on it is, unless they're either a law maker or a judge.

People do not get sued for having abilities. They get sued for publishing copyrighted works. What tools they use to create or publish those works is not relevant.

The only interesting legal question in my opinion is wether the large language model itself can be considered a carrier of copyrighted information, like for example an encrypted DVD would be.

If you can come up with a prompt that itself does not contain substantial copyrighted information, and that prompt leads the model to come up with a substantial amount of copyrighted information, then in my opinion it's no different from an encrypted DVD, and as such would illegal to distribute without licensing. Also I believe it would be illegal to provide access to it over the web like OpenAI does.

IANAL, but I wouldn't hold stock in any company that publishes (access to) models that are easily goaded into producing substantial copyrighted information.

I think the encrypted DVD is the best analogy I've heard so far.

In fact, I once asked an actual lawyer about the information-theoretic questions raised by illegal/infringing content known to be on an encrypted storage device.

Her answer was "Yeahhhh, no, the Judge is probably not going to see it that way." :-)

If you reproduce those songs and charge a subscription fee for anyone to use them you will be going to court. If you use samples of a song (which I think is much closer to an LLM) you will pay per sample use.

Also, comparing a human to a service that runs on billions of dollars of infrastructure is silly.

No one is arguing that AI LM's can reproduce copyrighted material verbatim.

I have AI LMs running locally on my GPU. They cost less than a photocopier or CD burner to operate. Pennies.

Reply on news.ycombinator.com