This article was updated on September 13, 2026.
Kevin Kelly, on his SubStack, shared some compelling points in his Arguments in Favor of AI Fair Use, and I urge anyone who knee-jerks at the thought of AI companies “stealing” their training data by hoovering up the entire web to take a look. He gets a lot right. But I think it misses the mark.
His argument centers on the idea that LLMs transform their sources by reducing them to a mathematical expression, something we’ve never seen before, in the process creating something completely new, which, at least in the States, is exactly the kind of claim companies need to make to free them from litigious copyright lawsuits. Free use, as it exists currently, guards the ability of artists and authors to employ a copyrightable entity in the pursuit of adding something to the cultural conversation, the most famous example perhaps being using Mickey Mouse to create a statement Disney might have never intended. If LLMs are indeed creating something new, their creations are free use.
I find Kelly’s argument flawed in this understanding, though. First, Kelly mentions that the way the information is stored is completely new, something akin to a math formula, and is therefore free from copyright. And he’s right about how data is organized. For the first time, you can’t point to a spot in an LLM’s datastore and say, here is Mickey Mouse. Here is the Declaration of Independence. The data is abstracted and scattered and weighed, stored in an unforeseen amount of places with an unpredictable amount of influence on other data, and AI experts will readily admit that we’re not even really sure how that all works. But organizing data isn’t the issue. In fact doing, organizing data privately isn’t breaking any laws and never has been. It really doesn’t matter if your method of organizing data is brand new. It’s how you obtain the data, how you’re offering access to the data, and what you’re creating with the data, that represent the real issues the public has with AI.
Kelly breezes over the laws we have in place to protect how works are used electronically. Obtaining perfectly free data that isn’t bound by copyright isn’t an issue at all; there’s lots of that online. Stealing data that explicitly isn’t meant to be freely distributed is an easy fault to understand, and AI companies have done that lots, and that’s pretty black and white. But there’s a mind boggling amount of data on the web that’s licensed by the Creative Commons, much of it with attribution. This means the author doesn’t really care what you do with their work, so long as you give them credit for their work. Just because an LLM atomizes data to the point that it’s indistinguishable from other data doesn’t free the creators of LLMs from following the law to cite their sources. It doesn’t free them from using electronically distributed data in a way that the author never allowed. And it’s not on the public to shoulder that burden, it’s up to the companies offering access to the data to do the right thing. Kelly mentions that to provide the source of everything an AI says would represent some of the most complicated and resource-intensive calculations ever performed, and you know what? That sucks for AI companies, but it doesn’t change the laws the currently exist, and it doesn’t change the necessity for those laws to exist.
Kelly coins this term “Public Intelligence,” which is an altruistic notion that purports that it’s the benefit of everyone to have LLMs trained on data that everyone should have access to. And this is frankly fantastic, especially when it comes to information like the laws, like history, like well-established science. But no one is providing AI for free. Companies are not tantalizingly close to world shattering IPOs because they want to offer their services for free. Whole economies are not quaking and upending because everyone is going to benefit from LLMs; there is a strong disparity around who has access to LLMs and they’re certainly not being developed with everyone in mind. It’s hard for the realist in me, at the moment, to marry with the optimal view of AI technology. If everyone is to benefit, why are nations in an arms race to get to AGI? Why isn’t it a resolution at the UN to work together? Why has the government failed to regulate the industry? I wish every car manufacturer would collaborate on self-driving, but that’d be suicide for the industry. As it currently stands, it’s all rose colored glasses.
But the most visible gripe people have with AI is with what AI produces. As of right now, they’re not smart enough to create something new. They don’t have their own thoughts, they don’t come to new conclusions, they instead present a average or probability. If they solve a mathematical problem that no one solved before, it means they had all the resources to solve it, not that it was unsolvable until they came along. They connect the dots, and they certainly know more dots than you or I or any human might ever have. You could view them as strict academics, librarians, scholars who can organize a view of the world to build perspectives. But they are not artists creating new ideas because they are not capable of cognitive leaps that advance art and culture. Every thought they’ve ever had, someone else has had and written down. The way they organize data may be unique, but at the end of the day they are averaging the weight of words that someone else has already penned, looking for the most practical next thing to say, which is plainly not creating something new to push the cultural conversation forward. I don’t believe they can do that, but I guess that’s because I don’t believe AGI has arrived yet, regardless of what the shovel sellers in this gold rush are saying.
Let me give an example. One can’t prompt an LLM into creating jazz music in a world without Jazz. A human might be able to convince an LLM into creating jazz music if they know what the LLM needs to do to create jazz music, but without direction, left to its own devices, it would consistently create instrumental music that sounds like everything that came before jazz and not Jazz. This definitely brings the philosophical conundrum of originality into the mix. Is all art derivative? I would argue not, but I also concede that much of the art being created by humans is. Kelly claims that 99% of all thought is unoriginal, and 1% is all we can claim in a new idea. That’s a bullshit made up number meant to float his idea. But he’s not wrong that new ideas are built on top of old ones. That’s why copyright law protects artists from derivative work, overly formulaic reproductions of art that come too close to the original to be distinguishable. Because creating something new is a gift, and it’s not easy. Not all art tries to alter a conversation, much art simply wants to extend the conversation; how many still-lives of roast duck and pears are there in the world? The question is whether my AI generated Mickey Mouse adds something to the cultural conversation, or simply copies it. I think the latter.
I think I could return to this topic for a long time, and perhaps I will. I truly think Kelly has done a wonderful job outlining the best case for fair use, and I want him to be right because it means we’ve figured out the problems and have found solutions, but I think the hand-waving Kelly employs dismisses the biggest issues there are, and doesn’t really solve them.