I hope you had a great Labor Day Weekend. I spent mine doing one of my favorite things: reading legal briefs about fair use. Specifically, I read the parties’ briefs in the gigantic consolidated OpenAI multi-district litigation (MDL) case. My takeaway? The plaintiffs are living in a fantasy world.
Every legal brief rests on a case theory, a story that brings together facts and law to paint a picture of the world that persuades the judge of your position. The problem for the plaintiffs is that, like a hallucinating AI model, their case theory paints a picture that flatters their own biases and sounds believable to them, but bears very little relationship to reality.
Perhaps the wildest hallucination in the plaintiffs’ briefs is the idea that AI models can be easily manipulated into providing users with copies of works from their training data. This is perhaps the most oft-repeated and ill-understood claim in the whole AI debate. Unfortunately, the plaintiffs’ heavily-redacted briefs tell us literally nothing about the basis for those claims.
OpenAI’s briefs, on the other hand, are not so cagey. Their experts found that “in a sample of 95 million outputs from the 20 million ChatGPT conversation logs that OpenAI produced in this case [there were] only 14 instances containing verbatim or near-verbatim regurgitation—i.e., an alleged regurgitation rate of 0.00007%. Of these 14 instances, all but two were 32 words or fewer (the two outliers were 67 and 122 words), and none represented more than 0.017% of a [plaintiff’s work].”
What kinds of things happen 0.00007% of the time? Well, that’s the average person’s odds of being struck by lightning in a single year. It’s also the odds of a coin-toss coming up heads 20 times in a row, or of catching two fly balls in the same baseball game. In other words, it’s a freak occurrence. Copyright law does not impose liability on extraordinarily useful general purpose tools based on freak occurrences.
Compare these stats to the Book Search tool in Google Books, which the Second Circuit ruled to be a fair use. 100% of searches in Google Books revealed snippets of 25-45 words. Showing these snippets was a core function of the tool – it helped readers better determine whether they’d found a book that would be relevant to their search. Because Book Search “augments public knowledge” without revealing enough of any copyrighted work to provide consumers a substitute for purchasing the work, the use was fair.
In the plaintiffs’ imaginary world, AI models are trained primarily on news articles and novels, and they operate by rote copying of the expressive contents of those works. They imagine the core uses for AI models are to generate news updates and book summaries, to jump over paywalls and regurgitate full-text books, and to generate unauthorized sequels and new books in the style of famous authors like David Baldacci. It’s a strikingly narcissistic fantasy.
As the briefs for OpenAI and Microsoft show, the purpose of exposing the model to a wide variety of training data is to allow the model to develop generalizations: facts, ideas, rules of grammar, and other patterns that emerge from the data collectively. These generalizations and abstractions, not any particular work or portion of a work, are the building blocks for the model’s new outputs. Like a metaphorical humanities major, AI models engage with a wide array of texts in order to become well-rounded, well-spoken thinkers, not so they can regurgitate memorized snippets at inopportune moments.
As to the primary uses of AI models, anyone who is paying attention can tell you that writing novels and summarizing the news are not even in the top 30 most common or commercially promising applications of AI. Software programming, business automation, fraud detection, and drug discovery are the kinds of real world use cases that promise to improve people’s lives, and offer AI labs returns on their investments. AI models were never intended to replace novels or newspapers; it’s not what they’re designed to do, and it’s not how people use them.
Time will tell whether Judge Stein can be convinced to live in the plaintiffs’ fantasy world. But it’s hard to see how anyone who is even halfway paying attention would be duped by such a simple, one-sided story.