Intellectual Property Won't Stop AI
And it shouldn't!
(I’m Henry Snow, and you’re reading Another Way.)
The following critique of AI has become commonplace: by training on stolen copies of books, AI is plagiarizing my work. It’s a thief. A recent piece out in the Atlantic on pirated books used for training AI has prompted another round of complaints like this– my colleagues are, understandably, incensed to find their work in the database that Meta used for training their language models. But intellectual property-focused critique of AI training risks missing the nature of AI developers’ theft of knowledge. AI models like the one behind ChatGPT are indeed stealing from writers, but what they are stealing is more diffuse and more profound than our individual works. They are stealing our capacity to do that work.
AI is not another industrial revolution, but a comparison to the original one can still tell us something about what’s going on here. Three hundred years ago in Europe, skilled work was done by artisans. For centuries, blacksmiths, shipwrights, weavers, and other artisans developed their craft, taught it to their apprentices, and advanced it together with their communities.
Then came the mechanizers. Men like Josiah Wedgwood (in pottery), Richard Arkwright (in textiles), and Samuel Bentham (in shipbuilding) developed new processes and tools for making the same goods. Their systems relied on the knowledge artisans had carefully built up– Wedgwood and Bentham were only able to do their work because they were trained in it as apprentices themselves– but it broke the cycle of reciprocity that knowledge depended upon. Instead of sharing knowledge, they centralized and hoarded it. By mass producing pottery with machines built on potters’ knowledge, Wedgwood slashed the value of pottery and ruined the livelihoods of traditional potters. Many were forced out of the small, autonomous production world they had lived in and into factories where they did not control their time, their actions, and ultimately their lives. I’ve written about all of this elsewhere in this newsletter, and at length in my book, which you can look forward to next year– for our present purposes, what matters is what this did to workers and how that compares to AI.
Imagine a world in which artisan skills were protected by modern intellectual property laws– and instead of flouting them, mechanizers decided to pay workers. This would resolve little. Had Samuel Bentham paid all of the shipbuilders he learned from, or even every living shipbuilder, a lump sum for his heist, he still would be robbing their children and damaging their communities by rendering their labor less valuable. Josiah Wedgwood made pottery cheap, to the point that making a living as a traditional potter became increasingly impossible. Weavers and potters went from small producers working on their own schedule to replaceable human parts in a factory.
All of this would be true even if mechanizers had paid for the knowledge they stole. The lump sum payment would have been whatever weavers or their representatives, if they chose some, could negotiate for– it is impossible to imagine it would be a life-changing sum for everyone. The transition generation would be marginally better off than otherwise, but the mass dispossession and miserable poverty of the industrial transition still would have happened. If we wanted to address the mechanizers’ theft of knowledge, we would have needed mass social programs to help with the transition: profit-sharing to ensure weaving communities could prosper and transition into a new and valuable place in the new economy, and guaranteed working conditions for those in factories, a collective vision for how we want our world to be. AI proponents have no plans for any of this. They don’t even want to pay a lump sum.
What’s worse is that AI developers would not have to pay all or even most of us in order to ruin us. Samuel Bentham did not need to learn from every shipwright, because he was not stealing a collection of individual shipwrights’ knowledge– he was stealing a pattern of knowledge that they all shared, and that any sufficient number of them could represent. Likewise for AI. OpenAI might want books in particular domains to generate more plausible bogus answers in specific subjects, but the underlying connections about syntax and language that these models are modelling appear in any English text. An AI trained only on public domain works, or on a smaller number of texts that had been paid for, would still work. It would still be threatening to our livelihoods, our ability to teach, etc. (Tech executives think they need all the data they can get to improve these systems, but it is clear by this point that “more data” is not producing better results in AI.) If anyone with these skills treats them as a commodity, it threatens everyone who has these skills– even those of us who don’t choose to sell them.
This gets at what commodification itself is- the ability to sell something, especially as an individual, to an individual, within the frameworks we are used to, also makes it a particular kind of thing. Land ownership in America is a perfect example of this: much of the land this country is built on was bought, but also stolen. In fact, it was stolen by purchase. Using force, colonists and then state and federal governments “granted” individual indigenous people and communities the “ability” to sell land. This was a purchase and a theft: land had not, prior to this point, been a commodity for sale in this sense. The mere act of making it available for sale was theft itself, rather than an alternative to it.
Likewise, we must resist the notion that knowledge is a commodity for sale at all. When we use our skills and knowledge, we are borrowing from a shared resource built by everyone we ever learned from, and one that everyone who learns from us will need. This stockpile is not ours to sell. Instrumentally, I am happy to use copyright law to bury AI, if that works. But let’s not confuse the legal and moral cases here. Already, numerous media organizations and academic publishers have signed various deals with AI companies to make their data available for training. The Atlantic, for example, has a “strategic content partnership” with OpenAI. Many of the same people who complain about intellectual property in these piracy cases also complain about these deals. But such arrangements are exactly what an intellectual property-based critique of AI implies we should set up. If we want to fight AI and not sell our future to it, we cannot rely on individual property claims.
What’s at stake here is not just the livelihood of individual writers or artists, but the ability of our society to produce new knowledge and art. AI’s infinite firehose of plausible-sounding but wrong knowledge reduces the value of existing writing skills. In so doing it makes it difficult or impossible to get paid for writing, and in turn, increasingly unlikely for individuals to choose, and eventually to have the choice, to learn to write. In actual industrialization, changes in work processes can still produce long-term good. It is a positive development that more people can do less work, in many areas. The loss of skills, while lamentable, is sometimes acceptable– I don’t think we collectively need the ability to build an 18th-century man-of-war today, and the skills of wooden shipbuilding, while less than profitable, live on in various heritage institutions and individual artisans. Transitioning out of an artisan world with the kind of social program I outlined earlier could have ensured us the benefits of industrialization without many of the harms– though democratic principles and basic justice in my view mean that artisans should have been involved in that decision.
AI, however, threatens fundamental skills– the ability to write, the ability to make art, the ability to think critically– and does not replace them with anything at all. AI is not at all like industrialization’s technologies. A power loom produces more reliable cloth in greater quantities than a human. But AI is less reliable than a human being– unable to tell fact from fiction, they constantly “hallucinate,” and unable to follow instructions deterministically, they produce unpredictable results (this is the only reason they can be “generative”!). For technical reasons I will go into in another post, these problems are in my view inherent to the way these models work. We already have a machine that will think for you and follow exact instructions. It is called “a computer,” and you are using one right now.
Every economic activity is also a political activity, both because it changes the balance of political power, and because the act of doing and justifying new kinds of work produces new and impactful political ideas. The ideas coming out of the AI world are very bad for us. But that doesn’t mean every opposing idea is good. How we critique, and fight, AI matters for what kind of future we make. The problem isn’t that Meta or OpenAI are stealing our books, but that they’re stealing our craft, and whether they take it or pay for it, it’s still theft.