Theory Fun Raph Koster GPT plugin

One day famed game designer Raph Koster awoke to discover that his classic book, A Theory of Fun, had been turned into a plug-in for ChatGPT. Which was somewhat strange, because Raph Koster had not authorized his book to become a named discourse in ChatGPT. In fact, he only found out about the plug-in because it showed up on Google/Google Alerts.

So his immediate reaction was an ambivalent, "Uh…"

I located the creator of the plug-in. Who as it turns out, was a game developer and educator in Asia. And his specific reaction was: "😱😱😱"

He told me the plug-in was created for a game design course he's teaching, and he was thinking about incorporating queries about games through ChatGPT for it. He chose Raph's book, he explained, because it was small enough to fit into ChatGPT. (He also incorporated writings by fellow game designers Marc LeBlanc and Nicole Lazzaro.)

In any case, he added, the plug-in was only meant as an experiment and wasn't intended to be published. (I am therefore not linking to it nor listing his name.)

While this specific instance seems to have been an innocent mistake, it's also an early warning sign of a future trend: Well-known authors' works ingested into ChatGPT and other LLMs without their permission, but still represented as their work and thoughts.

As for Raph, he has ambivalent thoughts about the experience, and what it means for the future of copyright, and his own work:

"I have mixed feelings," Raph Koster puts it to me in an DM exchange. "If a person read my book and then posted summaries, or was interviewed about it — I usually feel happy about that. There are many examples of just that — one came out on YouTube just a few days ago!" (See above.)

"So what's different if an AI database ingests it?"

"Not much, really. Except that I don't know if the AI database bought a copy, if you know what I mean. But it's possible OpenAI did! It's also possible it was just scraped from a pirate ebook site.

"This case goes a little further — training an AI specifically on one work then issuing it publicly feels like the creation of a derivative work. There are bounds on derivative works, in copyright. I don't think those bounds are very clear for a trained LLM, but selling it would probably exceed the bounds.

"Copyright law is gonna break — it was already bent beyond recognition by digital content in general. Which is something I've been talking about for years (you might remember my articles about CopyBot back in the day, or my long talk just about this issue." (Here's my own article on CopyBot, a program that was able to replicate user-generated commercial content in Second Life, and subsequently banned.)

"I am too familiar with remix culture, oral culture, or for that matter the history of culture, to think that cultural works getting swallowed up by other cultural works is evil," Raph goes on. "In fact, it's usually good and creates more culture. But the compensation mechanisms for cultural product are another story. How much cultural content is only enabled by the ability to profit from it? And what will we lose if creators cannot profit anymore?"

Those are many unanswered questions! Then again, they are currently being answered in US courts.

For myself as an author, I definitely do not want my copyrighted books ingested by LLMs without my consent, nor turned into a plug-in. I'm not even sure I would license them. LLMs are quite literally designed to produce mediocre average results, and I like to think my books are not that. (Let alone LLMs tendency to hallucinate, i.e. produce output attributed to me that I didn't even write.)

I'm also incredibly skeptical that the fair use doctrine has any validity in the context of LLMs. Fair use is intended to encourage editorial commentary and critique, which itself implies authorial intent. (I.E., by a human author.)

I'm overjoyed when my writing is cited at length in a review by a smart person who actually read my books, even if their take is critical.

When it's randomly, automatically hoovered up by an LLM? Not so much.

Posted in

One response to “Sign Of LLM Controversy to Come: Raph Koster’s Classic Game Design Book Converted Into a ChatGPT Plug-In Without His Consent”

  1. FlipperPA Avatar

    I’ve been pushing back on using terms like “hallucinate” for a while now. It is a category mistake. They do not hallucinate: they intentionally lie. Giving LLMs human attributes removes the ethical and moral responsibility from the programmer and model trainer. They only reason for the LLMs lying is because a program used a random-seed function to intentionally make up a response, to avoid having the LLM return that it has no answer. That’s because of the pressure to make them seem like omnipotent, all-knowing machines rather than glorified Markov chains. Here’s the talk I gave to DjangoCon US last year, if anyone is curious. https://www.youtube.com/watch?v=57zRoB4gx5M

    Like

Leave a comment