How to use AI in writing?

One thing I was thinking about in relation to the Shy Girl controversy was Corey Doctorow. Many (but not all) of his works are freely available. Most of them have been released under Creative Commons licences that let you download, share, and often remix them. And yet he still makes a living predominantly as a writer, because you can support him financially by buying his books from his publisher. The free downloads actually help drive paid sales, build audience, and get his work into libraries and schools. 

So, what does a publisher P have to fear from Person Ps₁ taking book B and sharing it on qBitTorrent? With the proper business model, it could actually drive sales of the printed books. Unfortunately, one big difference with Corey Doctorow is that he clearly owns the copyright on all his books anyway. Ps₁ is sharing B because it cannot be copyrighted. That means that it really doesn't matter whether Person Ps₂ prompt-engineered it with minimal effort. There is no reason for P to pay Ps₂ any royalties for B because it was public domain in the first place. 

So, how does Ps₂ make any money honestly from prompt-engineering such works? One would have to get paid directly to prompt-engineer the product, which P itself could pay Ps₂ to do. However, given the state of the art, I do believe that such works are likely to be slop, so P would be a slophouse. How would you prompt-engineer a work of any quality? I suspect that so much creativity would have to be put into the prompt that the result would become a copyrightable work. You'd just have to make sure that you kept a clear audit trail of all your work. It would be your proof of sufficient human intellectual effort and authorship in the final work.

Nevertheless, one thing that clearly presses people's buttons today is AI detectors. The Shy Girl controversy started in earnest when Pangram reported that the novel was 78% AI-generated. Even though AI detector tests are as inadmissible in court as lie detector tests. I strongly suspect that no publishing house is going to touch a book that scores over 50% on an AI detector test. Even if the result is a false positive. Because the legal costs of proving human authorship under those circumstances would probably outweigh the perceived benefits of publishing the book. In fact, I strongly suspect that the book would have to score 100% human-generated to be considered for publication at all.

To that end, I ran my last blog post through Pangram to find out that it was 100% human-generated. Well, that's nice to know, although I'm not sure what to make of it. Does it mean that I don't exhibit the kind of bad style that AI does? Or does it mean that my own style is so bad that no AI would have generated it? I very strongly suspect that it is a combination of the two, because both humans and AIs exhibit bad style, just in different ways. To make matters worse, Pangram does not even provide any reason for its percentage score, so you have no way of knowing how much of each reason applies. But then again, it is typical of neural networks that they lack transparency.

How did Pangram fare with the converse input? I copy-pasted the last thing that Grok said to me that didn't involve a quotation from me, which would have tainted the input. Pangram said that it was 100% AI-generated, so on the surface, this thing seems to be really accurate. But with no transparency into its detection processes, there will always be reasonable doubt about it, so it is still inadmissible in court. Nevertheless, the story seems to me to have a clear moral. If you want to publish your work, don't get AI to generate the text of it. I think it can still be a good assistant behind the scenes, especially for research, but write the text yourself to protect your intellectual property.