08 · Aesthetics & Phenomenology

This chapter sits in Part I, the Court of Philosophy — the eighth and final hearing before the benches switch, where philosophy asks the machine what the critics and the artists notice that the builders do not.

What do the critics and the artists see that the builders miss?

In October 2018, Christie's auctioned Portrait of Edmond de Belamy, a blurred, unfinished-looking image of a stout man in a dark coat, estimated at $7,000–10,000. It sold for $432,500. The work had been produced by a generative adversarial network trained by Obvious, a three-person Paris collective, on roughly 15,000 portraits scraped from WikiArt — and, as reporters soon discovered, on code adapted almost unmodified from a public GitHub repository written by Robbie Barrat, a teenager Obvious never publicly credited until pressed (Hyperallergic, "Christie's Sells 'AI-Generated' Art for $432,500 As Controversy Swirls Over Creators' Use of Copied Code"). Where a painter signs a canvas, Obvious printed, in the lower-right corner, the mathematical objective their network had trained against: min G max D 𝔼x[log(D(x))] + 𝔼z[log(1−D(G(z)))] — the generator and discriminator's loss function, standing in for a hand. The painting had, quite literally, signed itself with its own machinery instead of an author.

Every question this chapter has to ask is already sitting in that signature. Whose work is it — Obvious's, for choosing the dataset and the frame and the price? Barrat's, for writing the network that actually learned the weights? The eighteenth- and nineteenth-century portraitists in WikiArt, whose brushwork the model had statistically absorbed and recombined? Or no one's, because a loss function has no hand at all? The builders of computing have their own, older answer to a related question — not who made a thing beautiful, but why beauty should matter to engineering at all — and it is worth hearing their case before the critics take the stand.

G. H. Hardy opened A Mathematician's Apology (1940) with a claim that sounds like vanity and is actually an epistemic demand: "Beauty is the first test: there is no permanent place in the world for ugly mathematics" (Hardy, A Mathematician's Apology, 1940). Computing inherited the demand and gave it teeth. Edsger Dijkstra's 1968 letter to Communications of the ACM, "Go To Statement Considered Harmful," argued that unrestrained jumps in a program's control flow make it impossible for a reader to hold a correctness argument in mind — not merely ugly, but unverifiable, because a mind that cannot trace a program's structure cannot check its proof. Donald Knuth, who had begun the project in 1962, published the first volume of The Art of Computer Programming that same year and has spent more than six decades still writing it, volume by volume, because for Knuth an algorithm's elegance and its correctness were never two separate virtues to trade against each other. In this tradition, beauty is not decoration applied after the logic works. It is the condition under which a human being — not just a compiler — can tell that the logic works at all.

Hubert Dreyfus thought the builders were solving the wrong problem entirely. In a 1965 RAND paper commissioned to evaluate artificial intelligence research and titled, pointedly, "Alchemy and Artificial Intelligence," he argued that treating cognition as rule-governed symbol manipulation missed what Heidegger and Merleau-Ponty had already shown about ordinary human competence: skillful coping in a world — knowing how to finish a sentence, cross a room, read a face — is holistic, embodied, and pre-theoretical, not decomposable into the explicit rules a program could follow (RAND, Alchemy and Artificial Intelligence, P-3244, 1965). The expanded argument, What Computers Can't Do (1972), was met with outrage from the AI community it targeted and quietly vindicated by a discipline that had not yet been founded. Rodney Brooks's subsumption robots, described in "Elephants Don't Play Chess" (1990), abandoned symbolic world-models entirely for layered, reactive behaviors coupled directly to sensors and motors — Brooks's own slogan, "the world is its own best model," is close enough to Dreyfus's phenomenology that embodied robotics reads, in hindsight, like Dreyfus's critique built rather than merely argued. Large language models complicate the picture from the opposite direction: fluent, contextually apt conversation produced by systems with no body, no world to cope in, and nothing at stake in any outcome — an achievement Dreyfus's argument, on its face, should not have permitted. Whether that is a refutation or a change of subject remains open: fluent text is not obviously the same accomplishment as skillful worldly coping, and defenders of the critique — including Dreyfus's own later essay "Why Heideggerian AI Failed" (2007) — would say the goalposts moved rather than the argument broke.

Heidegger's own later work supplies the vocabulary for what happened to Barrat's code once it entered Obvious's pipeline. "The Question Concerning Technology" (1954, developed from a 1949 Bremen lecture) argues that modern technology's essence is a mode of revealing he calls Gestell, enframing: the world shows up not as something encountered on its own terms but as standing-reserve (Bestand) — resource, on tap, awaiting further ordering (Wikipedia, "Gestell"). Barrat's repository, released in the spirit of open-source sharing, was absorbed by Obvious as exactly this: not an authored work to be encountered and credited, but raw material on tap for a saleable print. This is also the moment this chapter owes its reader the fact usually left out of Heidegger's aesthetics: he joined the Nazi Party on 1 May 1933, ten days after being elected Rector of Freiburg, delivered a rectoral address endorsing the regime, and never fully accounted for either afterward. A philosopher who diagnosed the ordering of the world into mere resource is not thereby cleared of having served, at the one moment his institutional power was greatest, a regime that ordered human beings into exactly that category.

Andy Clark and David Chalmers's "The Extended Mind" (1998) offers, unexpectedly, the frame that ties the signature back to the hammer. Their parity principle holds that a part of the world counts as part of a cognitive process if it plays the role a part of the brain would play were the process run entirely in the head — illustrated by Otto, their thought-experiment Alzheimer's patient, whose notebook functions as his memory precisely because he consults it as unreflectively as a person consults recollection (Clark & Chalmers, "The Extended Mind," Analysis 58(1), 1998). That transparency-in-use is close to Heidegger's earlier and more famous idea, from Being and Time (1927): a hammer, skillfully wielded, withdraws from attention and becomes ready-to-hand (zuhanden); only a broken hammer becomes present-at-hand (vorhanden), a mere object demanding inspection. Elegant code disappears into a reader's understanding the same way; Otto's notebook disappears into his remembering the same way; and a programmer pairing with a language model that completes half their keystrokes is asking, in practice, exactly Otto's question — where does the boundary of authorship sit once part of the work happens somewhere the hand never touched. Obvious signed their canvas with an equation because, whether they knew it or not, that was the one honest answer available: the hand had already withdrawn.

Connections

Chapter 04 supplies the deeper argument about symbols, minds, and the Chinese Room that Dreyfus's critique runs alongside. Chapter 05 returns to Barrat's code through Wittgenstein's rule-following problem — what a repository "says" once someone else runs it. Chapter 13 continues the embodiment question through the grounding problem in learning machines.

Open questions

If elegance is a precondition for verifiability rather than decoration, does an unreadable but provably correct proof — the kind machines increasingly produce — count as beautiful in Hardy's sense, or has that criterion quietly stopped applying? Was Dreyfus's critique refuted by fluent language models, or did fluency simply stop being the test that mattered? And once authorship can be distributed across a dataset, a borrowed repository, a curatorial choice, and a signature that is itself an equation, is "who made this" still a question with one right answer?

Go deeper

• Donald Knuth, The Art of Computer Programming, Addison-Wesley, begun 1962, first volume published 1968 — the canonical case that correctness and elegance were never two different virtues.
• Hubert L. Dreyfus, "Alchemy and Artificial Intelligence," RAND Corporation, P-3244, 1965, read against Andy Clark & David J. Chalmers, "The Extended Mind," Analysis 58(1), 1998 — the critique of disembodied cognition against the theory that lets a notebook, or a network, count as part of a mind.
• Stanford Encyclopedia of Philosophy, "Martin Heidegger" — the standing account of enframing, readiness-to-hand, and the biography neither can be read apart from.