In the latest SHuSH newsletter, Ken Whyte fiddles around a bit with some of the current AI large language models and tries to decide how much he and other publishers should be worried about it:
The literary world, and authors in particular, have been freaking out about artificial intelligence since ChatGPT burst on the scene sixteen months ago. Hands have been wrung and class-action lawsuits filed, none of them off to auspicious starts.
The principal concern, according to the Authors Guild, is that AI technologies have been “built using vast amounts of copyrighted works without the permission of or compensation to authors and creators,” and that they have the potential to “cheaply and easily produce works that compete with — and displace — human-authored books, journalism, and other works”.
Some of my own work was among the tens of thousands of volumes in the Books3 data set used without permission to train the large language models that generate artificial intelligence. I didn’t know whether to be flattered or disturbed. In fact, I’ve not been able to make up my mind about anything AI. I’ve been playing around with ChatGPT, DALL-E, and other models to see how they might be useful to our business. I’ve found them interesting, impressive in some respects, underwhelming in others.
Unable to generate a newsletter out of my indecision, I called up my friend Thad McIlroy — author, publishing consultant, and all-around smart guy — to get his perspective. Thad has been tightly focused on artificial intelligence for the last couple of years. In fact, he’s probably the world’s leading authority on AI as it pertains to book publishing. As expected, he had a lot of interesting things to say. Here are some of the highlights, loosely categorized.
THE TOOLS
I described to Thad my efforts to use AI to edit copy, proofread, typeset, design covers, do research, write promotional copy, marketing briefs, and grant applications, etc. Some of it has been a waste of time. Here’s what I got when I asked DALL-E for a cartoon on the future of book publishing:
In fairness, I didn’t give the machine enough prompts to produce anything decent. Like everything else, you get out of AI what you put into it. Prompts are crucial.
For the most part, I’ve found the tools to be useful, whether for coughing up information or generating ideas or suggesting language, although everything I tried required a good deal of human intervention to bring it up to scratch.
I had hoped, at minimum, that AI would be able to proofread copy. Proofreading is a fairly technical activity, based on rules of grammar, punctuation, spelling, etc. AI is supposed to be good at following rules. Yet it is far from competent as a proofreader. It misses a lot. The more nuanced the copy, the more it struggles.
