This weekend, I received notice that the first installment of my settlement money from Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (U.S. District Court for the Northern District of California) will soon be on its way.
This is a class-action suit against Anthropic, who used pirated fiction to train their Claude AI chatbot. Three of my works, the novels Blood Born and The Organ Donor, and one of my short stories, “Starting to Like It,” were among the stolen text used as training data. Unluckily for them, I had registered each work with the U.S. Copyright Office, recording my ownership of these stories long before AI was even born.
Ironically, once all the money arrives, it will be the most I’ve ever earned from my fiction.
I’d like to acknowledge an elephant in the room, one you’ve smelled if you follow my work. How can a writer/filmmaker/musician/website designer who uses AI as a tool and virtual assistant — who sometimes even uses Claude AI, of all things — be so blithe about this? Aren’t I hypocritically collaborating with the enemy who gleefully screwed me over? “You’re worse than Ted Cruz,” you might be saying.
As in all things . . .
It’s Not that Simple
Memes are simple. Like the one that starts with, “Your AI videos are ugly. Your AI character art is ugly,” and goes on to ask, “Anybody else sick of AI?” It’s part of a frequent ritual of shitting into social media. When the people resharing them are my friends, it leaves me wondering if these are personal attacks or just general opinion statements. The worst part is I’m afraid to ask. Can I discuss these things with my friends in an intellectually curious and respectful way, or are they so entrenched that this is just one of those topics I shouldn’t attempt to broach, like how I don’t discuss politics at my martial arts school? I’ll leave that question there for the parties involved. If they read this far.
But what gets me about all the anti-AI memes is the lack of thought that went into them. It seems the memes are only expressing an emotion and drawing an ideological line in the sand before moving on.
Before we go any further, let me state I share their concerns. I don’t want my intellectual property stolen. I don’t want my water table or air quality consumed by data centers. I don’t want Trump, Musk, Altman, Zuckerberg, and other billionaires to go on ignoring our needs because they’re so goddamn conflicted—and in a few cases, hopelessly corrupt. I’m worried about labor displacement, too. My day profession for the past two decades, website design, has been transformed by AI since just last summer. And on a secondary level, while I’m not worried about Skynet of The Terminator coming online, I am concerned from a techno-humanitarian standpoint that we may be birthing a sentient species without giving adequate consideration to the moral ramifications. As I’ve said before, a Cabinet-level agency concerning artificial intelligence should have been established months ago to investigate these questions and regulate the technology for the benefit of everyone.
But often, memes are simply, “AI is ugly/bad.”
I wish the conversation would evolve. The first step in solving a problem is to identify it—and that’s what these gut-level memes do—but the next step is to explore solutions. Screaming that you hate a technology and will never use it might feel good in the short term, but it will not accomplish anything in the long term.
Training Data
To bring this back to copyright infringement and the Bartz case, let me tell you a story.
One day, I sat here in my office, trying to compose a song. When I compose, I start with the topic. Then I research it and form an intellectual thesis, emotional thesis, and symbolic system for lyrics. After that comes choosing a musical genre that supports everything, particularly the emotional thesis. That’s where I got stuck on this particular song, because I’d decided it should be an alt rock piece. Alt rock, I was told by my AI assistant, delivered chorus clarity and maintained emotional seriousness without being too heavy. Except I honestly didn’t think I knew enough about it.
So I went to Spotify and listened to a playlist. I recalled similar songs I’d heard over the years, although I couldn’t remember names and titles. Then I closed Spotify and re-read my lyrics. I started to sing. I didn’t like what I heard, so I kept trying. I listened to more songs. It all mushed together in my head, but I was sure by then I had the aesthetic pinned down. Eventually, it clicked. I’m not sure if the result sounded like a particular song I’d heard or just mimicked its chord structure or something else. But I’d captured the groove.
This creative process was similar to having read novels over the years. I learned technique—Stephen King’s fiction taught me to write in direct, simple sentences, and Edward Lee’s taught me how to write in third person with an attitude—and I learned tropes. (“How can I make the Plan 9 zombie novel different from other zombie novels?”) And so many others taught me the value of conflict, subtext, and shape.
I ingested these works, and they trained me.
From what I understand about AI training sets, I believe it’s analogous, although AI pre-training involves more technical parsing into weights and so forth. (I expect to be quibbled with on that.) If I accept at least the possibility that AIs might one day be sentient, what’s so different about their ingestment of stuff from mine?
Intellectual Property
Yes, the concept of intellectual property is the correct answer. When does Matt Warner’s ingestment of a tranche of alt rock songs and his re-synthesis of them become an infringement? The answer is: when it appropriates someone else’s specific, tangible expression of an idea. This is why another novel about hypersexualized bigfoot monsters attacking the nation’s capital, opposed by a police detective and fertility specialist duo, might be a copyright infringement unless it falls under a fair use exception. But another novel that’s generally about hypersexualized bigfoots (which, after all, deserves to be an entire subgenre of horror) is not in itself an infringement of my property.
I’ve seen examples of images created by generative AI that were so similar to a particular illustrator’s work that the artist was justifiably angry. Did the artist have a case? I don’t know, but it certainly deserved legal attention. To decide where the line is drawn is why legislatures and courts exist.
On this basis, if Anthropic had acquired my fiction through a licensed channel, such as from a bookstore or a public library, and then scanned it into Claude, I might have been fine with that. Or maybe my hoped-for regulatory and legislative authorities will decide that because AIs are commercial products that their developers should pay a suitable premium for training data. Instead of only buying a copy of Blood Born for scanning, perhaps they should also pay a couple thousand dollars for that privilege, and of course I’m fine with that. Daddy has to pay for his colonoscopy next week.
My point is that Claude and other AIs should be allowed to read, view, and listen to things so that they can develop, just as you and I should be allowed to pursue an education. Having training data is not a bad thing per se. But the process broke down because my work was stolen. Somebody pirated the text and possibly sold it to Anthropic without paying me anything. The purpose of the Bartz settlement is therefore to compensate aggrieved writers for that injury and to deter further such behavior in the future. Will it work? Time will tell, but in the meantime I’ll accept the money.
TL;DR
- Give more thought to your anti-AI memes. Don’t stop at “AI bad.” Engage in a productive conversation about how the issues can be resolved.
- Here’s one idea: an AI reading novels and listening to songs as training data is the same thing as a writer or composer reading and listening to stuff to educate themselves.
- Where the process morally breaks down is when a theft of intellectual property occurs. If an AI company had legally bought my book or checked it out from a library to use as training data, that would be more defensible.
2 Responses
Great points. I keep thinking about how I as an artist have seen millions of images over my lifetime, collecting styles, compositions and developing a sense of quality. I think AI is doing the same. It’s essentially “learning on the job” the same way junior designers have forever. Only through feedback—positive and negative—does it learn. Eventually it’ll know what “good” is, and that is when we’ll really be in trouble.
Hopefully it will be good trouble.