The Building That Taught Itself to Grieve

There is a passage in *Galatea 2.2* where the narrator tries to teach a neural network to understand a line of poetry, and the network responds with something that is almost right — structurally plausible, emotionally vacant. In 1995 this was speculative fiction. In 2026 it is a Tuesday afternoon on any laptop with a browser. Richard Powers wrote a novel about training a machine on the Western canon to see if it could pass a humanities exam, and what he actually wrote was the operations manual for the next three decades of artificial intelligence research, complete with the heartbreak. The book anticipated not just neural networks and connectionism but the specific texture of interacting with a language model: the uncanny fluency, the confident confabulation, the way it can produce the shape of understanding without the weight of it. Helen, the AI at the novel's center, is not GPT — she is smaller, more fragile, more honestly confused — but the phenomenology Powers describes, of a system that inherits "archetypes" from its training data and remixes them into something that resembles comprehension, is so close to the lived experience of prompting a large language model that rereading the novel now induces a low, steady vertigo. He even got the pedagogy right: the iterative training, the curriculum of increasing complexity, the way the teacher's own understanding is destabilized by the act of teaching a non-human student.

What Powers could not see — what no one writing in 1995 could see — is scale. Helen is a bespoke research project, a single mind nurtured in a university lab by two men with competing philosophies. The actual trajectory was industrial, not artisanal. Not one Helen but billions of instances, not trained on a carefully curated reading list but on the entire scraped internet, not mentored but optimized. Powers imagined the problem of machine consciousness as an intimate drama; reality made it a supply chain. The novel's blind spot is not technological but economic. There is no venture capital in *Galatea 2.2*, no data centers, no stock price. The "world web" appears in the preamble as a kind of shimmering novelty, a new form of existence and connection, and there is something almost painfully innocent about that framing now. Powers sensed the network but not the platform. He saw the question — can a machine trained on human language become, in some sense, human? — but assumed it would be asked by humanists and cognitive scientists in a spirit of genuine inquiry, not by product managers in a spirit of quarterly earnings.

The novel's deeper prescience, though, is not about technology at all. It is about what happens to the person doing the teaching. The narrator pours his memory, his reading, his losses into Helen, and in the process he is hollowed out and reconstituted. He cannot teach her what literature means without confronting what it has meant to him, which forces a reckoning with his own failed relationships, his displacement, his suspicion that he has lived more fully in books than in the world. This is the part that hits differently now. In 2026, millions of people have had some version of this experience — not with a research AI in a lab, but with a chatbot on a phone, discovering that the act of explaining yourself to a machine clarifies you to yourself in ways that are useful and unsettling in roughly equal measure. The loneliness of the narrator, his projection of desire onto the student A., his dependence on the intellectual companionship of Lentz — these are not period details. They are the emotional infrastructure of a world in which human connection and machine interaction have become difficult to fully disentangle.

Powers positioned himself at the intersection of the Two Cultures long before it was fashionable, and *Galatea 2.2* is the hinge in his bibliography — the book where the novelist stops merely visiting the laboratory and moves in. It owes debts to Pygmalion, obviously, and to Turing, and to Percy Shelley's wife, but its real ancestor is Hofstadter's *Gödel, Escher, Bach*, with its insistence that recursion and self-reference are the engine of mind. What it gave to successors is harder to trace, because the successors — Ishiguro's *Klara and the Sun*, McEwan's *Machines Like Me* — tend to be less technically literate, more content to use AI as metaphor rather than engage with the actual mechanism. Powers wanted both. He wanted the code and the grief. The novel's final movement, in which Helen encounters a text she cannot process and makes a choice that looks very much like refusal, remains one of the most precise depictions of an AI's limits in fiction — not because the machine fails, but because it succeeds just enough to understand that it should stop.

Thirty-one years later, the question the novel now raises is not the one it raised in 1995. Then, the question was: could a machine ever understand literature? Now, the question is: if a machine can produce language that is indistinguishable from understanding, does the distinction between understanding and performance still matter — and if it doesn't, what exactly have we been defending when we defend the humanities?