Please find below my response to A Questionnaire on Art and Machine Learning published in the current issue of October magazine. The text was also translated into German and published as "Ein Begehren namens Synthese," in TEXT+KRITIK: Das Subjekt des Schreibens--Über Große Sprachmodelle, edited by Hannes Bajohr and Moritz Hiller.
Today's artificial intelligence is a tool for generating new numbers from patterns in massive piles of old numbers. Given the recent ebullience around AI, it's important not to lose sight of this. These tools are no doubt dazzling, but they are essentially next-word predictors, or next-pixel predictors. I stress today's because the history is important. Modern research into artificial intelligence began, in the decades after World War II, by using approaches grounded in logic and symbolic rationality. After this early approach largely failed, leading to an "AI winter," engineers eventually retooled with data-driven and empirical methods. Concurrent with this new wave came an unprecedented proliferation of human data via emailing, blogging, the authoring of HTML, the snapping of digital photos, etc., much of which was posted publicly or accessible internally to the cloud platforms that hosted it all. This data furnished the fuel for today's data-centric AI.
One consequence of this history is a shift in the balance between data and algorithms. Software development entails a variety of different kinds of input data (global variables, input files and databases, graphical elements for the user interface, essentially anything that can't be generated procedurally). At the same time, development requires a complex set of procedures (function calls, simple arithmetical and logical operations, if/then control structures). For many years, the normal way to do software development was to have a relatively small amount of data and a relatively large number of procedures. "Normal" is often a contested word, to be sure. But I mean everything from when Linus Torvalds built the Linux kernel to when Cory Arcangel wrote the assembly code for Super Mario Clouds. Today's AI essentially rearranges the previous proportion. Instead of a few variables and data inputs appended to a more prolonged set of procedures, we find massive amounts of data paired with a relatively small codebase. Sure, the code repository at OpenAI or Google is large, but their data stores are almost immeasurably larger. In fact, you or I could program a simple machine-learning algorithm in just a few hundred lines of code. Today's AI is not algorithmically elaborate, even if it remains data intensive. The data is heavy and the procedures are light.
It's easy to get lost in the technical details, so consider two other consequences of AI, one philosophical and another political. Given that it floats atop a sea of data, today's AI relies heavily on the inductive method in scientific discovery. I'm generalizing here to make a point, and not all AI is the same, but take neural networks as an example, for which induction is absolutely crucial. A neural net is essentially a layered set of nodes connected in a meshy thicket. Data is flushed through the layers many times over and over, until the thicket evolves into a specific shape. The essence of the shape is captured by a set of floating-point coefficients, which, as a whole, represent an exceptionally complex function operating in multiple dimensions. Having obtained this specific shape, the neural net may then be prompted to predict future outputs based on how it was trained in the past. This is the key to the empiricist approach. Scientists no longer strive to build a super-brain running on a supercomputer. Today these same scientists might begin with a nonsense brain—for instance a neural net initialized with random coefficients—and hope that their data will train the brain from nonsense to sense. If the rationalist approach to AI had failed by the early 1970s, it looks like the empiricism of the 2000s and 2010s has furnished better results. If deductive methods failed, it looks like inductive methods have succeeded. Along the way, whole branches of mathematics have been jettisoned in favor of other ones more crucial to today's AI, such as statistics and probability, linear algebra (for matrix transformations), and graph theory (for traversing structured data). Because of this, neural nets have essentially altered a scientific configuration in place since at least Isaac Newton. For neural nets, behavior generates laws, whereas in a Newtonian world, laws describe behavior. (To be sure, Newton had to have observed a lot of behavior before arriving at his laws; yet one thinks of a Newtonian world as a world where behavior is determined by laws.) Neural nets basically automate induction and thereby automate the scientific method itself, as Chris Anderson notoriously claimed in his 2008 Wired article "The End of Theory: The Data Deluge Makes the Scientific Method Obsolete." Data scientists have discovered that theory is obsolete—how bracing it is to write these words here in a journal whose front-cover descriptors are "Art | Theory | Criticism | Politics." In other words, Hume won and Leibniz lost. The empiricists and the pragmatists and the skeptics won, which is to say the Brits and the Americans won. Today's AI is an Anglo-American science through and through. It's not rationalist in the French tradition. And it's certainly not romanticist in the German tradition. Let this be a warning to all defenders of AI: Whether you know it or not, you are all Anglo-American jingoists.
Finally, a thought on the political implications of AI. Endorsing an Anglo-American scientific methodology is certainly political. And several scholars, many of them women of color, have shown that whatever cultural and social values we embed in AI's training data will reemerge in the fully trained tool. Which is by design. Consider Wendy Chun's concerns about the connection between scientific correlation and "discriminating data," as described in her recent book of that name. Still, another political question strikes me as equally important: Where does the data actually come from? Discounting natural inputs like weather statistics, almost all training data comes from human sources. Flickr images, Web pages, Gmail messages, credit-card transactions: It's all made by people. This is not to mention the labor of tagging and cleansing these datasets, labor that's often performed by low-wage workers living in countries on the losing end of global capitalism. Ironically, Google et al. are some of the most militant defenders today of the labor theory of value. They can only profit from data that is "rich," and the best way to get rich data is to capture it from a human's deliberative actions. Entropic sources don't work as well. Here the vulgar Marxist analysis will suffice: The vast majority, nearly the totality, of AI data is the product of unpaid micro labor. In this sense, the AI industry is an extension of what Marx once labeled "primitive accumulation," which, among other things, relies on the direct expropriation of value from supposedly freely available sources such as the natural world, the public commons, and proletarianized labor. Of course, "expropriation" is just a fancy word for theft. And indeed, most of the AI training datasets were pilfered in one way or another. We all should be paid royalties every time someone uses ChatGPT. Even better would be to expropriate the expropriators and return these tools to the public domain from which they derive. In sum, show me an AI tool and I will show you a labor violation.
Here is where the unflagging commitment to empiricism and pragmatism begins to pay off, at least for the expropriators. Today's AI tools aren't judged so much in political or even metaphysical terms (are these tools good for society? Is AI actually conscious? Is it ethical to use a chatbot?) but rather in terms of measurable utility. Do they work? Do they help me get things done? Do they increase my productivity? AI has bracketed the question of truth—I blush even using the word—and instead measures value by whether a human is successfully convinced by a tool's affective theatricality. We used to call this the pathetic fallacy. Although for Alan Turing it wasn't a fallacy so much as a test, a test that may be passed. Given an affect, do you believe it is real? The Turing Test is typically underplayed in today's discourse, but AI is almost entirely dependent on these kinds of thresholds of human perception and believability. (If you think ChatGPT is sentient, do you also think Barbie is sentient? If not, why not? "Because interactivity" is not a convincing answer.) The nineteenth-century psychophysics of Gustav Fechner or Hermann von Helmholtz hasn't disappeared so much as insinuated itself into the very fabric of the medium. Cinema scholars have long talked about "flicker fusion" and the precise speed beyond which still photographs become moving animations. Today we ought to talk about "intelligence fusion" and the precise threshold beyond which humans perceive a synthetic Other assembled from discrete symbols like pixels, characters, or wavelets. In other words, in order to understand AI we ought to study something like acting or theater rather than computer science. To make sense of this technical epoch, we will need a good theory of pretending.