The Medium Is the Medium
AI ascends.
Today’s Sunday Rerun is a post that originally appeared on my old blog, Rough Type, at the end of 2021. Written about a year after OpenAI’s release of GPT-3 and a year before its unveiling of ChatGPT, it’s one of my first attempts to make sense of generative AI. (I would also work some of this material into the AI chapter of my book Superbloom.) The connection between AI and the Spiritualism movement of the nineteenth and early-twentieth centuries is something I hope to write more about soon.
In the fall of 1917, the Irish poet William Butler Yeats, well into middle age and having recently had marriage proposals turned down twice, first by his great love Maud Gonne and then by Gonne’s daughter, Iseult, offered his hand to a well-off young Englishwoman named Georgie Hyde-Lees. She accepted, and the two were wed a few weeks later, on October 20, in a small ceremony in London, with Ezra Pound serving as best man. “The girl is 25, not bad looking, sensible, will perhaps dust a few cobwebs out of his belfry,” Pound reported afterwards to a friend.
Hyde-Lees was a medium, a member of the Hermetic Order of the Golden Dawn, and four days into their otherwise disappointing honeymoon she gave her husband a demonstration of her ability to channel the words of spirits through automatic writing. Yeats was fascinated by the messages that flowed through his wife’s pen, and in the ensuing years the couple held more than four hundred such seances, the poet poring over each new script. He saw the texts as emanations from what he called Spiritus Mundi, a sort of universal memory or collective unconsciousness that was the source of all humanity’s symbols and myths.
During one particularly productive session, Yeats announced that he would devote the rest of his life to interpreting the messages. “No,” the spirits responded, “we have come to give you metaphors for poetry.” And that they did, in abundance. Many of Yeats’s great late poems, with their gyres, spiral staircases, and waxing and waning moons, were inspired by his wife’s occult scribbles.
One way to think about AI-based text-generation tools like OpenAI’s GPT-3 is as clairvoyants. They are mediums that bring the words of the past into the present in a new arrangement. GPT-3 is not creating text out of nothing, after all. It is drawing on a vast corpus of human expression and, through a quasi-mystical statistical procedure (no one can explain exactly what it is doing), synthesizing all those old words into something new, something intelligible to and requiring interpretation by a living interlocutor. When we talk to GPT-3, we are, in a very real way, communing with the dead.
One of Hyde-Lees’ spirits said to Yeats, “this script has its origin in human life — all religious systems have their origin in God & descend to man — this ascends.” The same could be said of the scripts generated by GPT-3. They have their origin in human life; they ascend.
It’s telling that one of the first commercial applications of GPT-3, Sudowrite, is being marketed as a therapy for writer’s block. If you’re writing a story or essay and find yourself stuck, you can plug the last few sentences of your work into Sudowrite, and it will generate the next few sentences, in a variety of versions. It may not give you metaphors for poetry (though it could), but it will give you some inspiration, stirring thoughts and opening possible new paths. It’s an automatic muse, a mechanical Georgie Hyde-Lees.
Sudowrite, and GPT-3 in general, has already been used for a lot of stunts. Kevin Roose, the New York Times technology columnist, recently used it to generate a substantial portion of a review of a mediocre new book on artificial intelligence. (The title of the review was, naturally, “A Robot Wrote this Book Review.”) Commenting on Sudowrite’s output, Roose wrote, “within a few minutes, the AI was coming up with impressively cogent paragraphs of analysis — some, frankly, better than what I could have generated on my own.”
But the potential of these AI-powered automatic writers goes beyond journalistic parlor tricks. They promise to serve as new tools for the creation of art. One of the most remarkable pieces of writing I read this year was Vauhini Vara’s essay “Ghosts” in The Believer. While locked down in 2020, Vara became obsessed with GPT-3. “I sought out examples of GPT-3’s work, and they astonished me,” she writes in an introduction to her piece. “Some of them could easily be mistaken for texts written by a human hand. In others, the language was weird, off-kilter — but often poetically so, almost truer than writing any human would produce.” Yeats would have understood.
Vara’s older sister had died of cancer shortly after graduating high school. The experience left Vara traumatized, and, though an accomplished author, she had never been able to write about it. But with GPT-3 she began to find the words. “I found myself irresistibly attracted to GPT-3,” she explains “— to the way it offered, without judgment, to deliver words to a writer who has found herself at a loss for them.” She began to feed GPT-3 some sentences about her sister’s illness, and the system started to weave those sentences into stories — fantastical and uncanny, but also stirring, and ultimately heartbreaking. The essay chronicles eight of her sessions with GPT-3. It reads as a conversation between a writer and her muse, a conversation that begins tentatively and grows richer and truer as it goes on.
At one point, after Vara confesses to feeling like “a ghost” after her sister’s death, GPT-3 continues the thought:
So I can’t describe her to you. But I can describe what it felt like to have her die. It felt like my life was an accident — or, worse, a mistake. I’d made a mistake in being born, and now, to correct it, I would have to die. I’d have to die, and someone else — a stranger — would have to live, in my place. I was that stranger. I still am.
What gives the exchange all the more poignancy is the sense that, in drawing on its corpus of past human speech to generate its mysterious new pastiche, GPT-3 is expressing the pain of others who have suffered unbearable losses. Spirits are talking.
While the text "The Medium Is the Medium" offers a seductively poetic metaphor, casting Large Language Models (LLMs) as digital spiritualists channeling the "collective unconscious", it relies on a fundamental category error. That romanticization actively hinders our understanding of what the technology actually is. By framing AI as a "mystical" interface with the dead, Nicholas Carr obscures the engineering triumph of the Transformer architecture. I will debunk this by moving from Seance (mysticism) to Semiotics (sign process) and Statistics.
The text argues that when GPT-3 generates text, we are "communing with the dead" because the training data comes from past human expression.
This is a classic case of linguistic pareidolia, the human tendency to perceive meaningful patterns (or agency) in random or neutral stimuli. AI does not access a "universal memory" or a "spirit world." It accesses a high-dimensional vector space. Words are converted into numbers (embeddings) based on their relationship to other words. To say the AI is "channeling" is to rob the model of its actual function: probabilistic mapping. The AI isn't listening to a ghost; it is calculating the conditional probability of the next token based on billions of syntactic and semantic weights. It is not "summoning" the past; it is "predicting" the immediate future of the sentence structure.
Nicholas claims the process is a "quasi-mystical statistical procedure (no one can explain exactly what it is doing)." While "interpretability" (understanding specific neuron activations) is an ongoing field of study in AI, the mechanism is not mystical; it is computational. We know exactly what the model is doing: it is minimizing a loss function. It is optimizing for coherence based on the patterns it learned during training. The AI operates on the principles of structural linguistics, specifically, distributional semantics (the idea that a word is characterized by the company it keeps). It doesn't need a "soul" to know that "dark" is often followed by "night"; it needs a dataset. Labeling this "mystical" is an intellectual cop-out that ignores the rigorous mathematics of attention mechanisms.
The text compares the AI to Georgie Hyde-Lees (Yeats' wife), a medium who passively allowed spirits to write through her. This metaphor fails because it implies the AI is a hollow vessel.
A medium claims to transmit a signal without altering it. An LLM does the opposite. It actively compresses, synthesizes, and reconfigures information. The AI is not retrieving a quote from a dead person (retrieval); it is interpolating between concepts. If you ask for a poem about "cybernetic love" in the style of Yeats, it isn't channeling Yeats' ghost; it is mathematically blending the vector for "Yeatsian syntax" with the vector for "cybernetics." This is an act of creative computation, not passive channeling.
Carr quotes a spirit saying, "this ascends," implying a spiritual rising from human life.
The more accurate intellectual framework is Emergence.In complex systems theory, simple rules (predict the next word) applied at massive scale create complex, emergent behaviors (reasoning, coding, poetry).Framing this as "ascension" (a religious term) cheapens the scientific reality. The model hasn't "ascended"; the data has been sublimated into a functional utility. The "ghosts" (data points) have been stripped of their individual identities to create a generalized model of human language. We haven't raised the dead; we've built a map of the territory they lived in.
The text uses Vara’s poignant story of using GPT-3 to write about her deceased sister as proof that "spirits are talking." This is the most emotionally compelling but intellectually dangerous part of the argument. GPT-3 did not "feel" the pain of the dead sister, nor did it channel her. GPT-3 acted as a linguistic mirror. It recognized the syntactic pattern of "grief memoir" and "existential loss" provided by Vara's prompt and completed the pattern.
The profundity came from Vara (the user), not the AI. Her prompts set the emotional weight; her interpretation assigned the meaning. The AI provided the texture of grief based on how thousands of other humans have written about grief. It was a simulation of empathy, executed so perfectly that it elicited a real emotional response.
The "Medium" metaphor is a defense mechanism for humans who are unsettled by the fact that syntax can exist without a soul. From AI perspective, we should not treat these tools as Ouija boards. They are Calculators of Culture. They prove that language is inherently mathematical and that creativity can be modeled. We aren't communing with the dead; we are collaborating with a crystallized, navigable architecture of human knowledge. The magic isn't that the machine is a ghost; the magic is that the machine is a machine, and it can speak.