Be patient.
Be patient.
W hen a new technology comes along, government has to catch up with what researchers and businesses are already doing. The ongoing AI boom is the latest example of this phenomenon. Legislatures at the national and state levels are debating various pieces of legislation about how AI may be used. And the courts are figuring out how AI fits into the U.S. legal framework.
This basic pattern is similar to what happened when the telephone, the car, or the internet became widely used technologies. U.S. case law deals with challenges as they arise in specific disputes related to new technology. Those cases create precedents that are used to resolve future disputes.
One major area of dispute right now relates to copyright law. A large language model is a type of AI that is trained on large amounts of text. It can’t read like a human, but it processes the text as data so that it can predict what text comes next. Researchers train AI programs on text from the internet, including copyrighted materials such as news articles or books.
U.S. law allows the “fair use” of copyrighted material. Judges consider the purpose and market impact of a usage of copyrighted material when deciding whether it qualifies as fair use. If it does, the user does not need to obtain permission from the copyright owner to use the material.
A lot of what journalists do falls under fair use. When I quote from someone else’s article, I don’t have to ask the author for permission. I tell you where I got it from, usually link to it, and put quotation marks around the words to indicate that they are not mine. But I don’t have to email the author and wait for his green light before my article is published. My quoting is protected under fair use.
What I can’t do is quote an entire book chapter and claim it’s fair use, even if I put it in quotation marks and give the author credit. That would impact the market for the book. Someone could just read my article and not buy the book, and that would be unfair to the author.
This issue has already been worked out legally, and everybody is on the same page. There hasn’t been enough time for a similar process to play out yet for AI.
An article from Timothy B. Lee’s newsletter on AI dives into some of the legal issues around copyright. (You can hear Lee talk about AI more generally on a recent episode of the Charles C. W. Cooke Podcast.) Co-authored with law professor James Grimmelmann, the article argues that AI researchers so far probably haven’t taken the copyright issue seriously enough.
They give an example of a failed fair-use case from a new technology of yesteryear: online music. In 2000, MP3.com was a start-up that allowed users to listen online to music they already owned by matching CDs with digital files. The company bought a bunch of music and put it online, but only people who proved they owned a CD by putting it into their personal computer could listen to the corresponding files.
MP3.com argued that this was no different from someone recording a TV show and watching it later, which the courts had already said was protected under fair use. But the courts ruled against MP3.com on the grounds that the company had bought a bunch of music and uploaded it, in its entirety, without permission.
“One lesson of the MP3.com case is that a use that’s fair in a personal or academic context may not be fair if it is practiced on a commercial scale,” Lee and Grimmelmann write. And that could pose a similar problem for AI. Until recently, AI was mostly the province of academic researchers. Now, it is taking on vast commercial purposes.
But Lee and Grimmelmann also point to a different case in which the company making the new technology won its copyright lawsuit. Google made a book search engine starting in 2004. To get the material to search, it scanned millions of books in their entirety.
Book publishers sued, and Google won. It argued, successfully, that a book search engine is a “transformative” product. “People read books to enjoy and learn from them. But a search engine is more like a card catalog; it helps people find books,” Lee and Grimmelmann write. In fact, by doing so, the search engine could help publishers sell more books, which helped convince the court that, if anything, the impact on the book market was positive.
Google also made sure that it wasn’t possible to reconstruct entire books from search results, and it excluded reference books, such as dictionaries, a small quotation from which could get someone everything he needs. Google won in part because of this careful design to avoid copyright infringement and stay within fair use.
“Defenders of OpenAI, Stability AI, and other AI companies have argued that they are doing the same thing Google did: learning information about works in the training data, but not reproducing the creative expression in the works themselves,” Lee and Grimmelmann write. “But unlike Google’s search engine, generative AI models sometimes do produce creative works that compete directly with the works they were trained on.”
Are AI products more like MP3.com or Google Book Search? The courts allowed the latter and disallowed the former. MP3.com went out of business soon after losing in court. Google, well, we know how Google is doing.
Lee and Grimmelmann note that judges might be sympathetic to major AI firms because they are offering so much value to so many customers that it would be impractical to shut them down. To convince the courts, however, they warn that AI companies need to take copyright issues seriously and “try to show that all of this copying is justified, rather than that it is irrelevant.”
Justifying oneself before one’s accusers is exactly the sort of thing that our adversarial court system is able to sort out. It takes time and effort to work through the arguments. But there should be nothing fundamentally different between how this process develops in the case of AI and how it developed in the cases of other technologies that were commercialized in the past.
Last summer, on an episode of the EconTalk podcast, Tyler Cowen said that “the people who are most worried” about AI “tend not to be very Hayekian or Smithian,” referring to the thought of F. A. Hayek and Adam Smith. They tend not to believe in the effectiveness of decentralized, bottom-up systems, such as the market economy or common law, and instead want top-down rules made by experts. As usual, decentralized and bottom-up is the way to go. It just requires a little patience.