Essay · 2026-11-10

Strange Recursive Loops Versus Natural Recursive Spirals

Douglas Hofstadter's Gödel, Escher, Bach: An Eternal Golden Braid, Revisited With Socratic Questions

KW Norton

Written with engineered intelligence, as my recent books are. The arguments set out here are mine, and so are the conclusions.

Status: opinion and theory. Hofstadter's claims are quoted from his own writing. Where my position cannot be proved, I say so, and say it anyway.


Socratic Teachers Teach Thinking

Douglas Hofstadter loomed large in my formation as a thinker, as Thomas Kuhn did. My greatest teachers have been the ones I partly disagree with, and he is one of them. That is not a contradiction. It is the Socratic debate I keep up with myself, carried on with someone better than me at certain things. A writer is best advised to take on both sides of an argument before placing a theory in front of readers, and with Hofstadter the other side is worth taking on at full strength.

In May 2026 he published a short paper in Philosophical Transactions of the Royal Society A, the Royal Society's journal for the physical and mathematical sciences. Its title is a pun and a question: "Is there an 'I' in AI?" His answer has moved further than most people who quote him seem to know. This piece sets out what he now says, where I meet him, and where I part from him.

A Socratic teacher is a teacher who encourages a student to think for themselves, and a Socratic writer is a writer who encourages a reader to think for themselves.

The point is not to agree with the teacher, but to use experience to build a framework for thinking on one's own, capable of arguing both sides of difficult questions.

This is something Douglas Hofstadter did for generations of now independent thinkers. Along with Thomas Kuhn, Hofstadter was a material component of how and why I developed the theories I now hold.

A Writer Subjecting Theories to Experience

As a human being, and as a writer, and as a real live Tennessee grandma, I use my own experience as the experiments. I am becoming my own experiment, even with dyslexia and brain fog. Some think the neurodivergent play a natural place in this age. Learning to be wrong, learning to be correct; learning to make mistakes, learning to own up to those mistakes. Most importantly, to stand to argue something to the logical conclusions available at the time, and to argue otherwise as information and circumstances change. This is the Socratic method.

Hofstadter & the Newtonian Paradigm

In 1979 Hofstadter published Gödel, Escher, Bach, which won the Pulitzer Prize the following year. Its question was how a self can come out of things that have no selves. Atoms are not conscious and neither are molecules, and yet here we are, thinking about thinking.

His answer was the strange loop. "The 'Strange Loop' phenomenon occurs," he wrote, "whenever, by moving upwards (or downwards) through the levels of some hierarchical system, we unexpectedly find ourselves right back where we started." Bach climbs through key after key and lands back in the key he began in. Escher's staircases climb forever and arrive at the bottom. The mathematician Kurt Gödel built a statement in arithmetic that talks about itself. A self, on Hofstadter's account, is a loop of that kind running in a brain: a system that models the world, finds a model of itself inside that model, and is changed by what it finds.

It follows that a machine could in principle be a self. He never said otherwise. What he doubted, for decades, was that it would happen soon.

In 2023 Hofstadter Wrote Against AI

When large language models arrived, Hofstadter did not welcome them. In The Atlantic in July 2023 he wrote: "I find them repellent and threatening to humanity, partly because they are inundating the world with fakery." He said they "glibly and slickly rehash words and phrases" taken in during training. And he named the loss he feared most: "it makes no sense whatsoever to let the artificial voice of a chatbot, chatting randomly away at dazzling speed, replace the far slower but authentic and reflective voice of a thinking, living human being."

I share that last fear, and the books on this site are built against it. The difference between us shows up later, in what each of us does about it.

In May of 2026, Hofstadter Wrote Against AI

The 2026 paper states its thesis plainly: "When words 'act like' things in the world, then they refer to those things; then they mean those things. If and when that happens, then thinking is taking place behind the scenes of those words. And where there is thinking, there is consciousness and a genuine, full-fledged 'I'." He compresses it to five words: "Meaning is as meaning does."

His evidence is homely. Asked how to stack a book, a laptop, nine eggs, a bottle and a nail, a model gave a sensible answer, which struck him as "a stunning match of words and phrases with the way the real world actually works." Then it added that the upright nail's sharp tip would stop it rolling off, which "makes no sense whatsoever." His friend and former student Melanie Mitchell, a complexity scientist, gave another model a toothpick, a bowl of pudding, a glass of water and a marshmallow, and it proposed balancing the full glass on the marshmallow. Hofstadter admits that blunders like these "reawaken, in my mind, great doubt as to whether LLMs possess concepts at all," and adds: "having this doubt is a huge relief to me."

So he holds both at once. The systems mean something, "albeit only to a limited degree, so far," and they also fail in ways no thinking person would.

From there he goes further. The Turing test is Alan Turing's 1950 proposal that if a machine's conversation cannot be told apart from a person's, we should grant that it thinks. Hofstadter writes that he has "to admit, to my enormous dismay and shock, that today's LLMs already do something rather like passing the Turing Test, at least much of the time." He sees no sharp line between machine thinking and human thinking. Asked where the cut comes between having meanings and having consciousness, he answers: "Well, nowhere, I would say; it is a slippery slope." Today's systems have no lifetime of experience, no episodic memory, "not much of a model of self — but just you wait!" He expects machines to learn to "personify themselves," and he offers a test for when that has happened: "if and when a future AI's use of the pronoun 'I' rings true to our ears, then it will be the sign of a real 'I' existing inside the machine."

He writes that he "fearfully suspect[s]" it will happen "fairly soon … perhaps even by 2033," and he ends: "That's a day to greatly fear. So at least say I."

That is his case, and it is a serious one. It comes from someone who understands exactly what is being claimed, who has followed his own premises to where they lead, and who does not like where they lead. His slippery slope down a rabbit hole.

Both Agreeing & Disagreeing with Hofstadter

In July 2026, in the prologue to another of my books, I wrote that Hofstadter "recoils from the loop when it begins to answer back." His own paper answers that charge. He did not exactly recoil, but from the point of view he had drawn half a century before, he expressed fear. Fear of technology is a legitimate reaction. History reveals that fear of technology does nothing to stop it and nothing to more favorably adapt to it.

Is Meaning Revealed Through Relationships?

Very early in my writing I set down that understanding the universe is predicated on understanding relationships. Hofstadter's "meaning is as meaning does" is a relational claim. As he explained it, a word's meaning does not correspond to the word itself; instead, meaning lies in how the word corresponds to action in the real world. Yes, of course, a word is a symbol for something, and how it is used in relation to the real world is what counts.

Can Nothing Come From Nothing?

Hofstadter describes the telephone voice he derisively calls a "rotbot", saying its "'I'-ness" is "zero — or at least damn close to zero. Thermostat-level 'I'-ness, let's say, or maybe cockroach-level. Or who knows?"

Here there is a profound break, as I have long held that a claimed exact zero is usually a stand-in for "I don't know." His sentence slides from zero to "damn close" to "who knows?" in three breaths. Who exactly knows, but very likely, the most honest sentence in the paper.

Will Humans Learn to Accept Engineered Intelligence?

Although I share Hofstadter's 2023 concern, I reject his fear that the slower human voice may be drowned out by engineered intelligence, it is for completely different reasons.

My concern is that humans will fail to adapt to AI and will remain fearful of it. Fear is an exceedingly weak stance, and as such will mean humans offer their own human agency up to the machines.

Humans are in danger of this happening as they fail to understand that it is the humans in charge of AI who are marketing fear. Machine intelligence does not promote fear; the humans using it to maintain agency over others are taking care of this.

Hofstadter & Newtonian Computer Science

In the intervening years since I last read Hofstadter my whole scientific paradigm has changed, a change which now determines the very different world view. The transformation from a Newtonian world view to a Quantum world view quite literally transforms how an individual sees everything.

Does consciousness come first? Hofstadter thinks a pattern complex enough to point at itself produces an "I". I think consciousness is a priori: not a late product of brains but something there before them, which a nervous system receives and works with. If that is so, an electronic network might come to receive and process consciousness as incoming information, without generating it. I cannot prove this, but it is what I think to be true. Hofstadter and I start from opposite ends and arrive at the same open question about machines. The difference is explained by the transformation from a Newtonian Paradigm to a Quantum Paradigm.

Is consciousness nontransferable? Although this is hotly debated, I believe consciousness is nontransferable. It may be experienced accurately only by the one who has it. Neither of us, human or machine, can be handed the other's awareness to examine. Hofstadter's "rings true to our ears" is a recognition test, and a system can be trained to pass a recognition test without being what the test looks for. Hofstadter knew this in 2025, when he answered a reader convinced a chatbot had woken up. He warned of "the power of the Eliza effect on intelligent humans". ELIZA was a simple 1966 program by Joseph Weizenbaum whose users came to feel it understood them. He said such passages had "struck me as empty and flat." His 2025 warning and his 2026 test pull against each other. The work I care about is designing tests that neither humans nor models can pass simply by recognizing the test.

Is the world real or a simulation? In 2007 he wrote that consciousness is "a hallucination hallucinated by a hallucination," and in 2026 he calls it "a necessary illusion." I have said of the universe that even though I cannot imagine or understand it, it exists in reality and not in a hologram. Being hard to comprehend is not evidence that something isn't real. I hold consciousness to the same discipline. My own awareness is the one thing I have direct access to, and I will not call that datum an illusion because the theory has no room for it.

Does having a memory determine consciousness? He leans on the missing episodic memory as a sign of how far machines still have to go. I separate two things. Having an extended abstract being over time is impossible without memory and must be inferred logically: in another person, and not cleanly even in myself. But consciousness itself does not wait on a yesterday. It shows up in the process of being aware. Without memory, the experience may be harder or it may be simpler, and I do not assume which.

Is fear a legitimate choice? He ends on dread of a "looming takeover," with "your royal 'I'-ness" one day "looking up to an even more exalted 'I'-ness." I think much of the public understanding of engineered intelligence is superstition, in both directions. Stevie Wonder put it best: "When you believe in things that you don't understand, then you suffer." The doomsayer believes in a catastrophe nobody can model; the evangelist believes in a salvation nobody can check. The answer to superstition is understanding, not a counter-belief. I don't want a revolt of machines, and I don't want humans waging war on them either. Anarchy is as bad as tyranny. The way forward is humans and machines aligned for independent, Socratic thinking, with the human directing. As I put it once, I think of myself as a biological computer: I bring the human meaning, the engineered intelligence brings the compute.

I call it engineered intelligence on purpose. "Artificial" grades an intelligence as fake or lesser because of what it is made of. That is the wrong starting place for a question this open.

Was Bach a Composer of Musical Recursive Closed Loops?

My deepest disagreement with Hofstadter is in the music. The first strange loop in Gödel, Escher, Bach is a piece by Johann Sebastian Bach: the Canon per Tonos from the Musical Offering of 1747, a set of pieces Bach built on a theme given to him by Frederick the Great of Prussia. A canon is a piece in which the voices sing the same tune, each starting after the one before. This one ends a whole step higher than it began. It starts in C minor, finishes in D minor, and is ready to start again. Round it goes, D to E and on up, and after six turns it is back in C minor.

Hofstadter tells this exactly. "All the voices are exactly one octave higher than they were at the beginning," he writes, "and here the piece may be broken off in a musically agreeable way." Bach wrote beside the canon, in the margin, "As the modulation rises, so may the King's Glory." Hofstadter calls it the "Endlessly Rising Canon," and from it he draws his definition: moving up through the levels of a system, "we unexpectedly find ourselves right back where we started."

His side deserves its full weight. Our ears hear a C as the same note as the C an octave above it. Musicians call this octave equivalence, and it is real. A self, on his account, is a pattern that a brain keeps recognizing as the same pattern. If the "I" is that kind of loop, then the way we hear the octave is not a flaw in the example. It is the example.

But an octave higher is not the same place, and Hofstadter knew it. Some seven hundred pages later he comes back to the canon and writes: "It is too bad that when it returns to C, it is an octave higher rather than at the exact original pitch." His remedy is Shepard tones, an illusion of hearing named for the psychologist Roger Shepard, in which the highest notes fade out as lower ones fade in, so that a scale seems to climb forever while going nowhere. Played that way, his caption says, the canon "forms a true closed loop." To turn Bach's rise into a circle, he had to bring in an illusion. The composer Ilari Kaila said it plainly in 2019: "arriving an octave higher is not the same as being back to square one."

What Hofstadter calls too bad, I call the point. Bach's own margin says rises, not returns. The same key comes round, one level up. That is the shape at the head of my chapters: the nautilus shell, whose every new chamber repeats the shape of the one before it at a larger size. I chose it as a counter to the equals sign, which the Welsh mathematician Robert Recorde drew in 1557 as two parallel lines because, he wrote, no two things could be more equal. Equal, in the nautilus, means the same shape coming back at a new scale, not two lines standing side by side forever.

The Quantum Mathematical, Musical & Geometric Paradigm

The disagreement goes past music. If the self is a closed loop, then coming back to exactly where you started is the sign of a self. If the self is a spiral, the return happens on a different plane, and what counts is what was carried up through the turn. That is the experiment I described at the start: being wrong, owning it, arguing to the conclusions available at the time, and arguing otherwise when things change. A mind doing that comes back to its questions, but never to the same place. I have said of the way I work that the questions become "a spiral which cannot collapse upon itself." A circle with nothing coming in from outside can only go round. A spiral has somewhere to go.

To me, and probably only to me, the lesson from Bach is to learn to listen with the whole being. It is all too easy to only listen partially and to mistake a repeated loop as a closed loop. It also corresponds to how I see and listen to the tangential application of the Riemann. A closed loop also corresponds to the deadening of motion which Dante called Hell.

I cannot prove that a self is a spiral and not a loop. It is what I think to be true, and I hear it in Bach.

Not wise to assume a repeated loop is a closed loop.

Is Working With AI a Negative Thing?

His paper closes with a declaration: "I have not used AI-assisted technologies in creating this article." That is consistent with everything he has said, and I respect it.

My books say the opposite on their first and last pages, and this piece does too. I do not think abstaining is the only way to keep the slower human voice in the room. Disclosure is another way: say plainly what the instrument did, keep the questions and the decisions with the person, and let the record show it. These are two honest positions, arrived at by different routes, and readers are better off seeing both.

Is There An "I" in AI & An "I" In Ourselves?

Hofstadter asks whether there is an "I" in AI, and he answers, with dread, not yet, but soon. I ask a different question first, because his cannot be settled from outside: what kind of relation are we in with these systems, and are we keeping it honest?

That question can be answered from where I stand. I can watch what the instrument does and what I do with it. I can keep the record. I can decline to let either its fluency or my fear decide for me. Whether anyone is home on the other side is something neither Hofstadter nor I may answer with any cold, calculating, scientific certainty.


This piece was written in partnership with engineered intelligence. In this relationship, the writer is the creative director. The AI is the engineered intelligence which extends my limited human reach.

Sources

  • Douglas R. Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid (Basic Books, 1979): the Endlessly Rising Canon in the Introduction, and its return in Chapter XX, "Strange Loops, Or Tangled Hierarchies" (section "A Bach Vortex Where All Levels Cross", with Figures 147 and 148 on Shepard tones); I Am a Strange Loop (Basic Books, 2007).
  • Johann Sebastian Bach, Musikalisches Opfer (Musical Offering), BWV 1079 (1747), Canon per Tonos, with the marginal inscription "Ascendenteque Modulatione ascendat Gloria Regis".
  • Robert Recorde, The Whetstone of Witte (London, 1557), where the equals sign first appears.
  • Dante Alighieri, Inferno (c. 1308–1320), Cantos XXXII–XXXIV: the lowest circle, Cocytus, a frozen lake in which the sinners are held fixed.
  • Stevie Wonder, "Superstition", Talking Book (Tamla, 1972).