Forum: Author Hangout

Sadly Asimov's three laws weren't as exciting as Skynet.

> Since the use for training involves only one copy, the loss to the holder is miniscule. The loss to society is potentially huge, and the burden on the companies to compensate copyright holders individually would be huge.

I don't think your logic on loss follows. Loss would be calculated based on how it's used by the person who violated copyright, not based on how many copies were taken. If somebody copies a book and sells it on amazon, they took one copy from the author but profited on the thousand copies they sold; did the author lose the one copy, or the thousand potential sales that went to the infringing item?

And saying that the burden to repay copyright holders would be huge is irrelevant, isn't it? You shouldn't, in a just system, be exonerated of a crime just because the magnitude of your crime exceeds your capacity to pay for the crime. Companies are fined in excess of their holdings and go bankrupt, and there is a legal process to identify who has senior claims on the money.

It's not my logic. It's the guy who committed suicide who's giving his beliefs. I guess he was some bigshot in the development of AI and then started warning people of the dangers of AI.

Edit: Wait, you're not even the person I was quoting. Yes, I agree it's not your logic, but I don't believe I implied that it was when I quoted an entirely different person.

If somebody copies a book and sells it on amazon, they took one copy from the author but profited on the thousand copies they sold; did the author lose the one copy, or the thousand potential sales that went to the infringing item?

That would be 1001 copies, not one, and the number of copies combined with the profiting commercially from the the infringement would be elements that would weigh in favor of a finding of copyright violation and awarding damages.

And saying that the burden to repay copyright holders would be huge is irrelevant, isn't it? You shouldn't, in a just system, be exonerated of a crime just because the magnitude of your crime exceeds your capacity to pay for the crime.

Yeah, I wasn't clear here. I was thinking more about legislation -- schemes that might be set up which would weigh the loss to the public and the burden on those developing what could be a revolutionary technology for society against a small, remote, and mostly hypothetical loss to an individual copyright holder.

Finally, copyright infringement isn't a crime; it's a civil matter. Including an author's work in an LLM's training materials doesn't "steal" the author's copyright. They still have it. At most, it is the loss of one sale of that work. Which, of course, isn't what authors are concerned about -- they are worried about AI putting them out of work. As are many, many others. We don't know yet if those worries are justified, but I personally lean toward the view that these systems are going to be VERY disruptive.

Contrast that with pop music,

What you describe applies to all music, not just pop. And a large part of it is the fact that there is a compulsory licensing scheme for music in the US that is managed by a government sponsored entity*.

And what's meant by compulsory is that an artist can't refuse to grant or revoke a license if they don't like what their music is going to be used for or who is using it.

For example Bruce Springsteen's "Born in the U.S.A." is seen by a lot of people as a celebration of patriotism by a lot of people and frequently gets used by conservative politicians at rallies.

This is actually the opposite of what Bruce Springsteen intended when he wrote it and he has threatened to sue conservative politicians using it.

However as long as those politicians have paid into the compulsory licensing scheme, he doesn't have any legitimate basis to sue.

The compulsory licensing is also what keeps parody artists like "Weird Al" Yankovic from getting sued into oblivion.

You might not be so happy if such a scheme got applied to stories.

*I believe that the US is not the only country with compulsory licensing for music.

If that were true, then there wouldn't be any point in their stealing that work in the first place.

It's true, and the motivation wasn't to allow the LLM to analyze that one novel, but rather the millions in the datasets. And, of course, licensing agreements either already made or in the works that may drastically reduce the relevance of these points.

But the number of times that work is used by the AI's output could be zillions.

As I understand it, the "use" by an LLM of any particular work does not include accessing a copy of the work (which isn't "kept" in the model after training).

I think it is more accurate to say that, after studying the structure of millions of novels, plus huge amounts of other information, current LLMs can produce chunks of text that read very much like "stories" produced by authors. However, per Sturgeon's law, 90% of everything is crap, and the LLMs were trained on lots of crap, so... GIGO.

That said, while my few early experiments with getting LLMs to "write" turned me off to that effort very quickly -- I like my writing better! -- I don't generally find what they produce to be garbage. For example:

I've used them to write macros to take story planning info from an Excel spreadsheet and put it in outline form in a Word document, something I would not have known how to do myself.

I like writing about my stories and having the AI respond with strengths and weaknesses. So, I might put down my thoughts about events to occur in one section of the novel and then get feedback. I benefit just from getting my ideas down, and knowing I can get the AI to organize my stream-of-consciousness rambling frees me up from letting a focus on organization hinder creativity. Plus, sometimes the feedback from the AI sparks something I hadn't thought of, though generally it's not "That's it!" but "That isn't it, but it makes me think of this other idea I really like."

As Ethan Mollick says, LLMs are weird. What they are good at and what they aren't isn't unpredictable, a "jagged frontier." However, research into their use in organizations is already showing that they make professionals better -- at least the bottom 80%, though it also seems that they may be of value to the top 20%, just in different ways. And the only way to discover their benefits is to play with them. In fact, one CEO of an international corporation said it was his non-English speaking employees that produced the most significant use cases for AI, mostly because they adopted the tools early to help polish their written English.

I remember reading somewhere that the impact of technological developments between 1900 and 1920 was greater than that of rest of the century. I have a feeling the pace of change from 2020 - 2040 could parallel that of the early 1900s. I hope I live to see it.

a small, remote, and mostly hypothetical loss to an individual copyright holder.

Well, it's not "mostly hypothetical" - a large number of copyrighted works were used for commercial purposes without permission, ranging from published writing to videos. And "small" seems arguable as well; if the data was used in training, then I think there's a pretty reasonable argument that it's contributing in every single query made to the LLM, and now that's some amount of loss that the original artist is suffering. At over a billion queries per day, that's gonna add up at even tiny fractions of a penny. And "individual copyright holder" sounds like a deliberately myopic view of the situation, because while nailing down the exact number of people who's work was used without permission would be challenging, the number is going to be larger than 1, by many orders of magnitude.

Edit: and to be clear, I think the argument about LLMs being a revolutionary technology for society is another one that's worth interrogating a little bit - it's easy to say "wow this tool is incredible" when somebody else is footing the bill for it. As soon as the various ai companies start charging actual rate for their LLMs, I'll be interested to see how many companies decide that maybe the value prop doesn't make as much sense anymore.

That's not responsive to my argument, though. It costs a lot of money to keep ChatGPT up and running, both in terms of compute infrastructure and energy costs. Right now all that money is being subsidized by venture capital. If you have to pay a dollar every time you submit a question to chatgpt, are you still going there to get an answer instead of stackoverflow? If your question requires four rounds of prompt engineering after your initial query to get what you're looking for and you get charged for four pieces of garbage before you get an answer you can use, is the tool still cost effective? When you pay real money and the tool gives you what is functionally a confident answer that there are two Rs in "STRAWBERRY", are you still feeling like society is undergoing a revolutionary change, or do you think that you used to be able to get garbage for free on the internet and didn't have to pay for the privilege?

I'll grant you that the works were certainly transformed, but in this case it's more like they were transformed from "a collection of individual works" into "an oracle that can reproduce copyrighted texts word-by-word", which is maybe not what the concept of "fair use" is really looking for when it comes to transformation. It's absolutely possible to e.g. ask chatgpt to give you the first word of Falkner's "As I Lay Dying", and the second word, and the third word, and with enough prompt engineering you can just get it to spit out chunks of text from the book. Note that "As I Lay Dying" doesn't enter the public domain until next year, so that seems like a bit of a problem.

And yes, I agree that the mechanics of paying licensing fees to everybody who publishes works on the internet would be ridiculously expensive. It probably means that slurping up everything on the internet and using it in a commercial enterprise without permission is a bad idea! Maybe companies shouldn't do it.

> I can reproduce large chunks of the script to 'Monty Python and the Holy Grail', but I am not guilty of infringing its copyright.

If you offer a service in which you take money from people and on-demand produce copies of copyrighted works, you are absolutely infringing on copyright.

You're getting hung up on the idea that the text is "memorized", which a) is meaningless to a computer and b) is completely irrelevant to the issue. I can't memorize a book, then type it out and sell my typed copies for money, even though I memorized it; it's not my book. The memorization isn't the issue, the reproduction for money is.