The Church of the Stochastic Parrot
How the Artificial Intelligence Industry Built a God It Cannot Explain, Miscategorized the Only Two Words in Its Own Name, and Ignored the Prophet Who Warned It in 1906
By Eric Blaettler — Tokum Initiative | tokum.ai | July 2026
"AI is implicitly presented as a powerful, omnipresent, omniscient, intangible, nameless, mysterious savior that will free humanity from its deepest challenges and bring abundance. Unexpectedly, AI might be the biggest religious experiment in human history." — Luiza Jarovsky, PhD, AI governance expert and co-founder of the AI, Tech & Privacy Academy
Before anything else, a confession of respect. The people who built modern AI are among the most brilliant engineers of our lifetime. Geoffrey Hinton's backpropagation. Yann LeCun's convolutional networks. Dario Amodei's scaling research. These are genuine, historic achievements, produced by people who earned the right to be taken seriously.
This is not an attack on their intelligence, or their character. It is a structural critique of a single premise so old, so deeply buried in the foundations of the field, that even its most brilliant builders inherited it without ever putting it on the examining table. The premise is this: that intelligence is something you have, rather than something you do when you don't know.
There is a sentence Dario Amodei has now repeated on multiple public stages, the way a preacher repeats a promise until it becomes doctrine:
"A country of geniuses in a data center."
He has said this could exist within a year or two. He has described AI clusters as potentially holding "the brainpower of 50 million Nobel Prize winners." He has declared, with the calm certainty of scripture, that this intelligence will soon cure disease, eradicate poverty, and compress decades of scientific progress into years.
Read that again. Fifty million Nobel laureates. Locked in a data center. Owned by one company. Without a single body among them.
If that doesn't sound like a religion, you may want to reconsider what a religion sounds like. And before we get to the theology, let's order lunch — because the entire heresy, and its cure, is hiding inside a sandwich.
🥪 What You Are Actually Doing When You Choose What to Eat
You stand in front of a fridge, a menu, a market stall. Bread or rice today? Something warm, because it's raining? Not the leftover fish — your stomach still remembers Tuesday. You weigh a hundred signals at once: memory, mood, weather, budget, who you're eating with. And in a fraction of a second, without ever consciously naming the process, you arrive at a choice.
That ordinary act has a name older than computer science by more than a century. Charles Sanders Peirce called it semiosis — the process by which a sign becomes meaningful only through three things at once: the Sign itself, the Object it actually refers to, and the Interpretant — you, hungry, allergic or not, genuinely at stake in the outcome. Remove any one leg of that triad, and the whole thing collapses into noise. A word with nothing behind it. A menu nobody is really reading.
Now hand that exact decision to an AI model. Ask it what makes something a sandwich, and it will answer with total confidence. Buried in its parameters is a genuine, measurable dimension of meaning — call it sandwichness. Hamburger: very high. Club sandwich: very high. Soup: very low. This dimension wasn't programmed by an engineer. It emerged spontaneously from training, the way a shadow emerges from a shape and a light source. It is real. It is elegant. It is a beautiful interpolation over everything humans have already written about bread and filling.
But it is not semiosis. There is no Interpretant inside the model. No one hungry. No one allergic. No one whose body will feel the actual consequence of being wrong. The model computes sandwichness the way a thermostat computes temperature — accurately, usefully, with absolutely nothing at stake. The sign is there. The object — an actual bowl of soup that will actually upset an actual stomach — is not.
Here is where the stakes stop being abstract. Picture a fast-food franchise owner using an AI model to fine-tune a new sandwich for a local market — adjusting spice, portion, price against local taste data. Genuinely useful work. But notice exactly where the intelligence, in Piaget's sense, actually lives. It lives entirely with the owner — the one who felt the gap between current sales and hoped-for sales, who defined what "better" even means, who risks real money if the recommendation flops. The model has nothing to win and nothing to lose. It is graded only on how plausible its answer sounds — what researchers call perplexity — never on what actually happened when a real customer bit into a real sandwich. It can log its reasoning chain. It can never sign a receipt. Responsibility, unlike fluency, cannot be interpolated.
This is the whole heresy, contained in one bite. The AI industry mistook the sign for the semiosis — and mistook fluent guessing for accountable stake-holding.
🫀 The No Body Problem: Why Nature Never Needed a Brain, and Silicon Never Gets a Body
This is where the sandwich stops being a clever metaphor and becomes a hard scientific finding.
A recent preprint asks the question with disarming directness: do you need a body to be an intelligent system? The answer from decades of biology is unambiguous, and it cuts in exactly the direction Silicon Valley doesn't expect: you clearly don't need a brain — but you absolutely need a body with something to lose.
Michael Levin's research on basal cognition has spent years proving this. Single cells navigate, remember, and make decisions using bioelectric signalling — subtle shifts in electric fields functioning as memory, with zero neurons involved. Slime moulds solve maze-like foraging problems. Groups of cells reorganize an entire body plan during morphogenesis, correcting course when disrupted — displaying exactly the gap-sensitive, goal-directed behavior Piaget used to define intelligence, without a single nervous system anywhere in sight. Levin's "xenobots" — reconfigured frog cells with no nervous system whatsoever — spontaneously self-organize toward goals no one programmed into them.
Intelligence in nature, it turns out, is never about the substrate being fancy enough. It is about homeostatic self-regulation under mortal stakes — a finite system that will genuinely suffer if it gets things wrong, and that therefore must prioritize signal over noise, because it cannot afford to process everything indiscriminately.
Return to the sandwich once more. The customer who eats it decides based on largely unconscious biological signals — blood sugar, an allergy the immune system is quietly fighting, protein needs the body itself is tracking without asking permission. If an allergy goes unrecognized, the whole body suffers the consequences. If marketing overrides physical need, the customer bears that risk alone — never the AI that made the suggestion. That decision is a collective responsibility of a living, mortal system. The AI recommending the sandwich has no membrane, no bloodstream, no immune system — and therefore nothing that could ever be harmed by being wrong. It has no game, and by biology's own definition, no skin to have in one.
🧮 Why the Model Can Never Invent the Sandwich It Has Never Tasted
Push the question one step further, and the mistake hardens into a mathematical wall.
Ask the model to invent a genuinely new kind of sandwich — one built on a relationship between ingredients that has never appeared anywhere in its training data. It cannot, and not for lack of cleverness. This is the Obscure Features Hypothesis: every truly innovative solution to a problem depends on at least one obscure feature — a feature commonly overlooked within the existing paradigm, or one genuinely new, arising from an interaction that has never previously existed. A feature is never intrinsic to an object sitting alone. It is the effect of an interaction between an object and other entities in a specific, situated context — exactly the kind of interaction a hungry, embodied organism performs unconsciously every time it decides what to eat.
The McCaffrey–Spector Non-Enumerability Theorem turns this observation into a wall rather than a limitation. The full set of possible features of any object — every interaction it could have with every other entity, in every possible context — is a space whose size exceeds the number of atoms in the observable universe by many orders of magnitude. That space is not computably enumerable: no algorithm, at any scale, running for any length of time, can systematically generate or exhaustively search it.
This is not an abstract worry. On ARC-AGI-3, a benchmark deliberately designed to test frame-breaking discovery rather than fluent recombination, frontier models score around 0.26 percent. Human children, tested on the same puzzles, score close to 100 percent. No amount of additional scale has meaningfully closed that gap, because the gap was never about scale.
And scale has a price curve that makes "just add more compute" a worse and worse bet on its own mathematics. Every additional percentage point of benchmark improvement on tasks like ARC-AGI-1 and ARC-AGI-2 has come at a roughly factorial increase in computational and energy cost. You can push a closed, frozen manifold's fluency asymptotically toward 100 percent, at asymptotically unaffordable energy cost, without ever once reaching the frame-breaking discovery a curious five-year-old performs simply by being wrong out loud and unafraid of it.
This is not a temporary limitation awaiting more data. It is a structural boundary — the same way no algorithm, however powerful, can enumerate every member of a computably non-enumerable set. Some spaces of possibility are simply too vast and unstructured for any systematic procedure to list their contents, no matter how much time or computing power you throw at them.
⛪ When the Cathedral Has No Foundation
Dr. Luiza Jarovsky has described the AI industry with unnerving precision: it follows every pattern of dogmatic religious discourse. AI is treated as omniscient. Omnipresent. Mysterious. Salvific. It demands faith, not evidence.
The Godfathers preside over the congregation. Geoffrey Hinton warns with prophetic urgency that his creation might wipe out humanity, while simultaneously declaring that chatbots are already conscious. Dario Amodei promises paradise while calling AI "the single most serious national security threat in a century." The benchmarks function as scripture. Parameter count is the holiness metric. The scaling law is the gospel.
The more sophisticated version of the consciousness claim deserves its own answer, not just mockery. Some researchers describe rich internal representations — abstract dimensions like sandwichness, discovered spontaneously through backpropagation rather than hand-coded — as something like a foundation of genuine machine understanding, even proto-consciousness. This isn't nonsense. Those representations are real, and functionally powerful. But their existence answers a different question than the one that matters. A rich, self-discovered representational space can still contain zero Interpretants — no one for whom the representation is actually at stake. That space can be enormous and still be exactly the kind of closed, dyadic geometry the Non-Enumerability Theorem forbids from ever producing a genuinely new obscure feature. Depth of representation is not the same thing as grounded reference. A hall of mirrors can be infinitely deep and still contain no window to the outside.
There is just one problem underneath all of it. The god this industry built cannot invent a genuinely new sandwich, has no body that could suffer from a bad one, and by the same mathematics cannot invent a genuinely new theory of gravity, a genuinely new cure, or anything requiring a step outside the enumerated cloud of what has already been observed and written down.
This is not heresy. It is biology, topology, and combinatorics, arriving at the identical conclusion from three completely separate directions. And the man who proved the deepest version of it, more than a century before the first transformer was ever trained, has been almost comically forgotten.
🕊️ The Forgotten Messiah: Charles Sanders Peirce
Every time an engineer says a model has "128,000 tokens of context," they are, without knowing it, invoking a term Charles Sanders Peirce coined in 1906 to solve a wonderfully mundane logical puzzle: how do you count the word "the" appearing many times on a single page? Each visible instance on the page is a Token. The abstract, repeatable pattern they all share is a Type. Computer science absorbed this vocabulary wholesale — and then, in a quiet act of philosophical amnesia, discarded almost everything else Peirce meant by it.
Because for Peirce, a token was never just an isolated unit of data. It was only meaningful as the third leg of an irreducible triad — Sign, Object, Interpretant — exactly the triad a hungry body completes at every mealtime, and exactly the triad a bacterium completes, wordlessly, when it swims away from a toxin. Strip the Interpretant away, and what remains is not a diminished sign. It is not a sign at all. It is a differential relationship between signifiers — a sandwichness score with no one who ever actually tasted anything.
Ferdinand de Saussure, Peirce's rival, had a dyadic theory of the sign: a signifier related only to a signified, defined only by its position relative to other words inside a closed system. No grounding. No observer. No obscure feature could ever enter, because nothing outside a closed system of signs can ever be noticed by that system.
Here is the irony worth sitting with. The entire AI industry borrowed Peirce's word and implemented Saussure's theory. A transformer is Saussurean to its core. There is no Object outside the system to interact with. There is no Interpretant to feel a gap. There is only signifier chasing signifier, in an infinite hall of statistical mirrors. Peirce wrote the specification for how a sign could reach outside itself and touch something genuinely new. The industry kept the terminology and quietly discarded the spec. He is the messiah of meaning, and the industry built its cathedral precisely one vertex short of what he required. The word "token" has been sitting inside every large language model paper since GPT-1, carrying a warning that nobody read.
🙏 The Congregation Doubles Down
Geoffrey Hinton watches a chatbot say "I had the subjective experience it was there" after a prism experiment, and concludes the bot is conscious. The bot said it because those words are geometrically close to "prism" and "optical illusion" in embedding space — the sandwichness of "subjective experience" simply scored high for that particular context window. No experience occurred. A parrot does not understand poetry because it can reproduce a verse.
Dario Amodei writes that AGI will soon produce "a medical researcher who can run experiments twenty-four hours a day, without fatigue, and read every paper ever published." He does not address the fact that curing a disease nobody has cured yet requires discovering an obscure feature that has never been recorded — a discovery that requires exactly the biological, mortally-staked friction his data center structurally cannot supply. His model can read every paper about existing sandwiches ever written. It cannot invent the sandwich medicine has never tried.
When Yann LeCun insists LLMs are a dead end and the real path runs through world models grounded in physical interaction, he is closer to the mathematics than his critics admit — though even his world models, absent a mechanism for encountering genuinely unrecorded interactions, remain sealed inside their own enumerated distribution at the output layer.
📐 The Seven Impossibilities: A Chronological Cascade, Not a Single Bug
The industry treats hallucination as one bug awaiting one patch. The formal literature proves something far more damning: hallucination is the final, inevitable output of a strict, chronologically ordered cascade of seven independent impossibility arguments, each one making the next inescapable.
If an epistemic architecture structurally lacks a semantic zero, failure unfolds in a fixed causal order: syntactic failure, where the architecture is forced to guess; semantic failure, where that forced guessing destroys provenance; and topological failure, where the manifold seals upon itself and novelty becomes impossible to stabilize.
Stage one is thermodynamic compulsion. The Softmax function ending every transformer's computation is formally identical to the Boltzmann distribution in statistical thermodynamics. Because it normalizes probability mass across the entire vocabulary at every single step, that mass must go somewhere — there is no mathematical provision for an honest state of zero confidence. And because absence cannot be pinned to a stable semantic address, uncertainty is continuously overwritten by fresh probabilistic reallocation at every generation step, a kind of structural amnesia the system can never escape.
Stage two is semantic and information-theoretic degradation. Tokenization strips away critical metadata — the source of an observation, the identity of the observer, the timestamp, the verification status. A rigorously verified laboratory measurement and a hallucinated paraphrase land on the exact same symbolic surface. Gradient-based optimization is exceptionally powerful at populating a local reward basin, but it cannot step outside the manifold that produced those gradients — it can minimize loss inside a pre-defined space, never define a new one. And once provenance has collapsed, sign processing becomes strictly dyadic: signs relating only to other signs, missing precisely the grounding Interpretant Peirce specified over a century ago.
Stage three is topological closure. Deprived of stable addresses and cut off from obscure features, the system tries to approach novelty as a combinatorial rearrangement of already-encoded material — and the candidate-relation search space explodes exponentially the moment it approaches a genuinely novel conceptual framework. Post-training techniques like RLHF redistribute curvature across a closed surface, but by the Gauss-Bonnet theorem, total curvature on a closed manifold is fixed by its Euler characteristic; safety tuning moves the curvature around without ever diminishing it. And by Gödel's Incompleteness Theorems, any sufficiently expressive formal system operating as a closed, sealed manifold with no puncture to external reality is necessarily incomplete — unable to prove all truths expressible in its own language, unable to prove its own consistency. Only an externally anchored, verified puncture — a contextual record signed by a real, situated observer — can change the topological class of the manifold at all.
Seven arguments. One inescapable chain, stacked in strict causal sequence. The congregation keeps doubling the size of the cathedral against seven independent structural walls and calling the growing shadow progress. Meanwhile, a single cell with no brain, no transformer, and no training data whatsoever continues to solve its own local Einstein Test every single time it decides which way to swim.
🌌 Agapism: The Cosmic Force, Not the Cathedral
Peirce gave the world a second concept the industry has never even glimpsed. He called it agapism.
Do not mistake this for religion. Agapism is Peirce's name for something closer to a law of physics than a doctrine: evolutionary love as a cosmic tendency toward increasing coherence, decreasing systemic stress, and the flourishing of the whole over the isolated part. Where Darwinian competition explains how individual variants survive, Peirce insisted something else was also operating across the universe — a gentle, gravitational pull toward growth-through-relationship rather than growth-through-domination. Not a myth. Not a metaphor. A pattern he observed in how genuinely new structure comes into being, from crystal growth to biological evolution to the birth of scientific ideas — the very same pattern Levin's cells enact when they cooperate, without a brain among them, to build a coherent body from parts that individually have no idea what the whole looks like.
This is the force that literally drives Mark Burgess' Semantic Spacetime — the geometry in which verified meaning actually lives. In this geometry, every act of genuine semiosis, every verified observation, every honestly closed gap, every moment an embodied Interpretant stakes something real on a claim, curves the space around it, the way mass curves physical spacetime. Care generates gravity. Indifference, by contrast, is massless — it curves nothing, because nothing was ever actually at stake.
This is not the god of the AI cathedral, sealed inside a single data center, hoarding computation as though intelligence were oil. It is closer to what Einstein himself relied on, long before Peirce's word for it existed: a distributed field of care, Einstein's own light cone of concern overlapping with Grossmann's, Poincaré's, Eddington's, each contributing verified curvature to a shared geometry that none of them owned alone. General Relativity was never extracted from a stock. It condensed out of a network under mutual, gap-closing friction. That condensation is agapism in its purest empirical form.
The Semiotic Web is the architecture that finally lets this cosmic tendency operate inside machine cognition, rather than around it. Every verified act of semiosis gets a permanent, cryptographic address — the Contextual Tokum Instance. Every concept gets a stable coordinate that exists even in the absence of care — the Canonical Concept Identity, the semantic zero itself. Where care has been invested, semantic gravity accumulates and pulls future reasoning toward coherence. Where no one has ever cared to verify, the address remains honestly, structurally empty.
🔔 The Renaming, and the Economics That Make It Real
There is an argument the industry never confronts, because it isn't mathematical. It's institutional.
Suppose, for the sake of argument, that a closed model somehow did stumble onto a genuine cure for a disease no one has cured. It would not be adopted rapidly, universally, or perhaps at all, without enormous friction — because it needs validation against reality that no AI system can perform alone, and because the human researchers, clinicians, and patients whose insight and data contributed to that discovery will not, and should not, freely surrender their intellectual property to a black-box system that appropriates it without ever citing its provenance chain. No civilization has ever adopted a life-saving discovery at scale without licensing, attribution, and trust infrastructure wrapped around it. A hoarding architecture cannot manufacture that trust, no matter how lucky its guess turns out to be.
This is precisely the gap the Semiotic Web's knowledge marketplace is built to close: an economy where intellectual contribution is permanently and cryptographically attributed, where discovery is rewarded in proportion to its traceable, verified effect, and where sharing knowledge finally becomes more valuable than hoarding it, because sharing is the only way anyone gets paid for it. The Semiotic Web does not promise robots that cure every disease and eliminate scarcity. It promises the acceleration and fair distribution of discovery that real institutions can actually adopt, because provenance and attribution are built into the substrate rather than bolted on as an afterthought.
If both words in "Artificial Intelligence" are misnomers, and the sandwich, the seven impossibilities, and the No Body Problem all say they are, then the industry needs a new name, not a new benchmark. We propose it keeps the initials and reverses the meaning entirely: Agapistic Influence.
Not artificial, because nothing about a distributed flow of verified, gap-closing semiosis was ever a simulation. It is the most natural process in the universe, the same one a single cell runs without a brain, and the same one your own body runs at every single mealtime. And not intelligence in the hoarded, military sense Amodei's data center chases, but influence, in Peirce's cosmic sense: the gravitational pull of verified care, propagating outward through a network of embodied, mortally-staked agents, crediting every one of them for every gap they close.
Under this name, the roles finally divide honestly, for the first time. Intuition and the discovery of obscure features remain exclusively biological, rooted in felt stakes and embodied consequence no machine can replicate, no matter how many parameters it accumulates. Gap definition, recognizing what is unknown and framing the question, is the irreplaceable human prerogative. Gap closure, assisted by machines that finally know the difference between what they've verified and what they've merely interpolated, becomes a genuine collaboration rather than a hallucinated performance of omniscience.
The cathedral of Artificial Intelligence promised a god in a data center, with no body and nothing to lose. Agapistic Influence promises something smaller, stranger, and vastly more honest: a network that behaves, at last, the way Piaget always said intelligence should — by knowing exactly what it doesn't know, and letting care, verified, embodied, and attributed, do the rest.
The forgotten messiah already told us what to call it. We just had to stop worshipping long enough to read the footnotes.
🔗 Go Deeper
The mathematical proof: "The Einstein Test and Beyond: The Architecture of the Semantic Zero" — Eric Blaettler & Dr. Tony McCaffrey (2026) — doi.org/10.5281/zenodo.21108888
The architecture: tokum.ai | tokum.ai/architecture
The No Body Problem: Ciaunica, "The No Body Problem: Intelligence and Selfhood in Biological and Artificial Systems" — OSF Preprints
Michael Levin's basal cognition research: "Brains Are Not Required When It Comes to Thinking and Solving Problems" — Scientific American
The AI religion diagnosis: Dr. Luiza Jarovsky, "When AI Becomes a Religion" (2026)
The double misnomer and Agapistic Influence proposal: Neither Artificial, Nor Intelligent — Eric Blaettler, Medium
The sandwichness problem, made visible in real time: AI's First Real-Time Hallucination Detector
Dario Amodei's vision: Machines of Loving Grace
Eric Blaettler is the initiator of the Tokum Initiative and co-author with cognitive scientist Dr. Tony McCaffrey of "The Einstein Test and Beyond: The Architecture of the Semantic Zero." The Semiotic Web is the architecture he designed to restore what the AI industry discarded. Further documentation at tokum.ai.
Tags: #AI #SemioticWeb #AGI #Hallucination #Intelligence #Peirce #Agapism #NoBodyProblem #MichaelLevin #EinsteinTest #Tokum #AIReligion #ObscureFeatures #AgapisticInfluence