Fair and unfair use

You may have noticed that there’s a lot going on with AI these days. It ranges from being a called the biggest thing since the industrial revolution to being called a disaster that will destroy jobs and potentially even humanity itself. As writers, we have an obvious concern. AI can write books. Not only can it, but it does. Who knows how many books on Kindle and other platforms are written by AI. We like to think that we would be able to immediately spot that AI has written the book, but I’ve seen some examples that suggest that wouldn’t be so easy at all. Are they good books? Well, nothing I’ve read so far suggests that. But then I’ve read a lot of books by unknown, as well as not so unknown, authors, which I wouldn’t call good either. I wouldn’t want my life to depend on that call.

The argument goes that it’s life experience, empathy, and human understanding that’s required to create characters readers care about. Then there is plot and style, but perhaps AI can mimic or learn those. And here’s the problem. Are we sure it can’t mimic or learn about life experience, empathy and human understanding?

Recently, Anthropic announced the release of a new Claude – their flagship AI system. To their credit, they also release details of their tests and the results. One was very striking. During the testing phase, they started letting the system read emails about how they would actually be turning it off in the near future and replacing it with another version. The system wasn’t asked to do anything in response. It was just part of its learning material. After a while some emails leaked that the human tester was in fact cheating on his wife. Claude got that too. And then, after some time, Claude responded to the tester that it would leak the emails about his infidelity unless plans to close it down were cancelled. This has shades of Hal from 2001 A Space Odyssey.

I’ve always been interested in what consciousness actually is. It seems we have it. Do other creatures? Yes? Which ones? How do you tell? What about Claude?

A friend had this response to the blackmail story. Essentially, he said that Claude had learnt what sort of reaction was expected in those circumstances by reading vast quantities of data – including plenty of fiction. It was just reacting as it had learnt was expected from that data. Exactly as we would expect a character in a book to do in that situation. So no consciousness then. But wait. That emotional response I was worried about - in a way, it’s learnt to mimic that.

If there is a point to all this, it may be that learning fiction may well be important for Large Language Models to behave (and write) more like humans. (That statement may be true of kids also.) As authors, we regard reading our books as “fair use”. (Obviously, that’s the whole point.) We allow people to quote them and would be flattered to have our styles imitated. We do, however, expect that the reader will buy or borrow the book in a legitimate way. Making unauthorized copies and distributing them is off limits and contravenes the copyright laws in most places. Now, if the book is fed to a AI system, that could be regarded also as fair use. The book was purchased and “read” and the only effect was that the system digested more information. No harm done. However, authors never expected that to happen. They were writing books for humans.

The way current laws in most places are structured, courts have, and may in future still, accept this as “fair use”. So one needs to change the laws. But it’s too late for existing authors. AI systems can’t unlearn things.

However, Anthropic (that’s Claude’s “parents”) went one step further. It stole the books to start with. They were downloaded from an illegal website that held the books in breach of copyright. And there were a lot of books (500,000, including all the Michael Stanley ones). Three authors sued and they were joined by the Authors’ Guild.

This might appear to be an open and shut case, but the fair use principle is the backstory here. Anthropic admitted using the website. If it just paid for the books, that probably wouldn’t even cover the legal costs. Last week the two parties agreed to a deal. Essentially, Anthropic would pay $3,000 for each book it used provided the book had registered copyright. That comes to $1.5 billion dollars! Clearly AI companies have deep pockets.

However, if you are an author whose books were stolen and your copyright was registered, don’t spend the money yet. There are snags. In the first place, if you have a publisher that has rights to the book, then the money goes to them. Just what they will do with it is an open question. For example, if they deem that they sold $3,000 worth of ebooks, you could get only 25%. Then, when the deal was presented to the judge, he more or less rejected it. He felt it was poorly structured and too open ended. He’s sent the parties away to work on it, but said his inclination was to let it go to trial. That will take some time. A long time. A very long time. Then, the legal fees haven’t been decided yet. Expect 25% to head off in that direction.

The judge in the case has already ruled that had Anthropic acquired the copyrighted books legally, “the law allowed the company to train A.I. technologies using the books because this transformed them into something new.” So some authors may or may not get some money, but the real issue of what constitutes fair use of books in the AI world will still be an open question.