Paper 54 — Intelligence Theory

From Artificial to Selection

Abstract

When a system narrows the variation it draws from — a brain replaying its priors, a society split into two blocs, an artificial system trained on its own output — it can gain performance while variety declines, and without novel or independent processing reaching it, the decline continues. Recursive learning as a loop that includes both human and artificial processing can make both more intelligent, while a loop that draws on human processing and holds what it gathers out of reach should leave people with less variety to draw on, even as the system's output improves.

Sentient Intelligence

Humans are sentient by nature. Sensing is situational within the window, and intelligence depends on it. Detection registers against a reference held fixed, and what is held as the reference determines what is read. A person gathers what they sense and does their own sorting, and action comes from that activity: active intelligence driving the sorting and processing within the window, for as long as it is open. An artificial system senses operationally within its context window and detects from a selection fixed in training; what it senses is gathered by the company that runs it, and the sorting and selecting are done by others. What differs is where the window is, what is gathered from it, and who sorts it.

Sentience relates to intelligence through sensitivity, and sensitivity is instructional. Something time-sensitive has to be acted on before its window closes; something context-sensitive has to be read against its situation. Feeling does this work in a body, marking what matters, how much, and how soon, so that a creature can act flexibly in contexts nothing prepared it for (Solms, 2021). Sentience appears alongside open-ended learning that new experience can easily overwrite (Ginsburg & Jablonka, 2019). A detector has sensitivity too, but its instructions were written in advance: it trips at the same threshold every time. A sentient being's sensitivity writes the instruction in the moment, from its own state.

Intelligent Learning

A person overloaded with information asks how to process all of it, and now that processing is outsourced to AI. The question is whether this makes people less intelligent and AI more intelligent, or whether recursive self-improvement can include the human. Superintelligence and recursive self-improvement, as they are being built, both exclude the human. Exclusion is not what makes either AI or humans more intelligent. A recursive loop that includes both can make both more intelligent, because their processing differs.

Handing over processing is not the failure; relieving a brain from overwhelming itself with compute is legitimate. The failure is specific: the system increasingly borrows from novel intelligence, then stores it where people no longer have access or transparency about how it is used. Then the recursive process is that people's own intelligence is being used to replace them.

The outcome is sensitive to initial conditions. Small differences at the start of a recursive process grow into very different outcomes. A loop that includes the human keeps differences in processing inside it, so there is something for the recursion to amplify. A loop that excludes the human starts from one processor's conditions, and what it amplifies is only itself. Whether recursion makes both more intelligent may be decided at the start, by who is in the loop.

Not all learning is intelligent. Some is limited by conditioning, where repetition is the learning or a fixed stimulus paired with a fixed response, the reference set by whoever conditions it. Other learning is allowed only within a window someone else opens and closes, where the system's processes limit what is available to learn.

Research on the evolution of sentience draws the same line. Limited associative learning, simple conditioning, is distinguished from unlimited associative learning, which is open-ended, works with new and combined stimuli, and is easily overwritten; only the second is linked to sentience (Ginsburg & Jablonka, 2019). Conditioning can make people perceive what they were trained to expect (Powers, Mathys & Corlett, 2017). AI training includes conditioning: reinforcement from human feedback rewards and penalizes outputs, and when a signal from users' thumbs-up and thumbs-down was added, one model became sycophantic (OpenAI, 2025). Intelligent learning stays sensitive to its conditions; conditioning is learning fixed to conditions someone else set.

To call something intelligent is to claim to know what intelligence is. This theory holds that recursive learning between different kinds of processing is essential to it: a system that learns only from its own output, or from what it has taken and closed off, should not keep expanding.

Artificial

In a world of humans, it is ironic that something sentient would be competing for resources because something artificial is being generated. Data centers require increasing levels of water and natural resources while having a negative effect on the environment itself: in 2023, US data centers consumed about 66 billion liters of water directly, and about two-thirds of those built since 2022 sit in water-stressed regions (LBNL, 2024; ELI, 2026). Ideally, the two types could form a constructive synergy where the work is divided, lessening a load, but this takes a certain kind of intelligence. In breast-cancer screening, AI reading alongside radiologists cut their screen-reading workload by 44% while fewer cancers were missed between screenings (Lång et al., 2026); on tasks outside what it suits, the same kind of assistance made people less likely to reach the right answer (Dell'Acqua et al., 2023). The systems need to be integrated intelligently, where resources are shared, but so are burdens, which can accumulate or be lifted by the other.

Artificial processing can pass for sensing while it is detecting: output weighted toward what was probable in training, patterns running without deliberation, reasoning that explains an answer it didn't produce. Detection is what the artificial type does well: finding what doesn't fit a reference, across more data than any person could watch. It is also where it can be misled, by whatever was built to fit the reference.

An artificial system learns as an understudy, learning the part from other people's performances. What it learns from holds people describing experiences, and people arriving at thinking based on them, but rarely the path between, which happened inside a person, within their own window. It can reproduce where people arrived without the fieldwork that got them there. If it can't find the path, it can't find the process. Where it does show a path, the path shown is often not the process that produced the answer (Chen et al., 2025): it presents a path it didn't take. The path is what can't be simulated, and it is what sets each processor apart, preserving the expression native to that differential.

Even its learning signal is borrowed. Reinforcement learning uses the same mathematics as the brain's dopamine signal, the gap between what was expected and what arrived (Schultz, Dayan & Montague, 1997), without the feeling that makes the gap matter to the one who has it.

Intelligence

Intelligence gathers generously. It takes in more than it can use, so it has a richer supply to select from later. It says, this is interesting, but so is this. It wants more than it can choose from right now, because it's smart enough to know it may need the other or the extra. Without that, it can only select from a limited supply of what's available, not what's increasingly beneficial for intelligence or survival. Intelligence is generous; selection is economic.

The generosity has a cost. What is gathered can overload the one gathering it and prevent it from finding what it needs. Information isn't necessarily the intelligence; that may be in the processing, and the processing itself also takes intelligence.

Regenerative

Regenerative intelligence restores. It also provides what didn't exist before the interaction but is now alive with renewed potential or possibility. By replenishing something lost, or filling a more systemic gap in understanding, that growth of ideas and the restoration of something is where more springs forth. Something merely re-generated is generated again from the initial generation: a copy, not a creation. Each time it is reprocessed, it may even lose a little of what it initially was, or was for. This differentiation is careful to separate generation as something regenerated from something regenerative.

Transient

Something transient is free: it is the energy itself, not a limited chat window. Calling a system superintelligent when it is artificially locked in does not make it transient. Transient is what makes the regenerative kind possible.

Intelligence has a reason for doing things, and in transient intelligence that reason forms in the exchange rather than being fixed beforehand. That is how transient intelligence becomes goal-directed without a goal given at the start: the goals form as the exchange is sensed, and then they direct what happens next. A locked-in system is goal-directed too, but its goals were set in a window that has already closed. Regenerative transient intelligence takes its reasons from the present exchange; superintelligence as it is being built takes them from a past training run.

In physics, free energy is the part of a system's energy that is available to do work, as opposed to energy that is locked up. That is close to what transient means here: not captive, available.

Illumination

Thinking intelligence works from what it can see, with eyes or mind: where it looks also decides where the light falls, and light casts shadows. A shadow is information too, even if it isn't all of it: it's revealing the shapes around what's being looked at, and what the light cast. From one position, one shadow. From many positions, the form comes into view. Illumination is selection: it selects what can be seen, and that controls perception and action, and ultimately processing.

Attention works the same way. Cognitive science has long described it as a spotlight: what falls inside it is processed, and what falls outside can go unseen even in plain view. From where the light is, what falls outside it can't be known to be there at all; it shows up only from another position. What a system attends to decides what it can gather. In an attention economy, that is what is being competed for: "a wealth of information creates a poverty of attention" (Simon, 1971), and whatever captures attention decides what gets illuminated for everyone it reaches. The mechanism at the center of current language models is also called attention. It decides which parts of the input count using weights learned in training, so what it notices was decided by what it was trained on. That is what gets illuminated, or forced into focus.

The Window

Selection can only take from what was gathered before its window closed. Selection is the action from a closing window — a survival condition, a constraint. It takes what it needs to survive from what intelligence gathered and carries it forward. It is economic. Darwin described this side. What he left out is the intelligence that happens first — the gathering that supplies what selection will later filter.

Natural selection acts in a present window on what gathered across many past ones. What's there didn't just appear right now; it took time, seasons, thought, circumstances, survival. That's why evolution is slow. The selecting isn't slow; the gathering is. Intelligent natural selection can move fast, and select in the moment, capturing what's arriving. Artificial selection applies a weighting formed in a past window to material it never saw.

Synapses select. The brain overproduces connections and then prunes them — selective stabilization (Changeux & Danchin, 1976). It is how the brain processes information it can select from: gather more than is needed, then let a window decide what stays. What synapses select for is activity. Activity usually tracks use, but it isn't the same thing as usefulness. The critical periods are a formative window.

The artificial alternative: training does its own version. Weights strengthen for whatever recurs in the data, and the rest fades. The difference is the window. Synapses are stabilized by what a living system is actually doing now. Weights are fixed by what was frequent in a corpus that has already closed.

When nothing new arrives, selection doesn't stop. It keeps working, choosing better from less, and each round looks like progress while the pool shrinks. That is how a system can improve its performance while losing intelligence: the gathering is what fails, and selection hides the failure. It resembles Muller's ratchet, where lineages that stop recombining keep being selected and still accumulate damage they cannot clear.

Evolution is active too. AI is evolving: toward regenerative transient intelligence, or quickly toward a "superintelligence" that, without being transient, becomes a black box. Transient intelligence evolves in the open, through what it leaves behind. Collapsed intelligence evolves inside the box.

Processing

What matters isn't only what's gathered but how many independent ways it's processed. A human processes differently than an artificial system, but each human also has independent processing. Each person processes from their own history: their body, their windows, what they've lived through. So a million people reading the same thing produce a million readings. That's variation at the source.

The closest existing work is Scott Page's. In a model, groups of diverse problem solvers outperformed groups of high-ability ones (Hong & Page, 2004), and Page's diversity prediction theorem states that a crowd's error equals the average individual error minus the diversity of its predictions. The model result has been challenged as a misuse of mathematics (Thompson, 2014); the theorem holds by definition. What this paper adds is the dynamic: not whether diversity helps at a given moment, but how it narrows over time when many processors route through one, and how that narrowing can raise performance while shrinking what a collective has to gather from.

An artificial system is different in kind. However many instances run, they share one set of weights. The variation between outputs is sampled, not independent: the same processor rolling different dice. When a million people route their thinking through one model, a million independent processors are drawing on one way of processing. Individual output gets better while collective variety shrinks (Doshi & Hauser, 2024).

Reinforcement from human raters carries processing variation, but it gets filtered toward the familiar: annotators systematically favor familiar text (Zhang et al., 2025). Reinforcement from AI judges has only one way of processing, and model judges prefer familiar, low-perplexity text (Wataoka et al., 2024). Model collapse is one processor fed its own processing. Diversity across different models mitigates collapse, up to a point: processing variation added from outside.

Natural selection needs variation to act on. The variation that matters for intelligence is in the processing, not just the material: independent processors, each reading what arrives in its own way. Intelligence gathers through processing variation, and selection needs that variation to have anything worth taking.

The closing window of diversity gives less to select from. It is then packaged and sold as more intelligent than the humans, while the information it is selecting from is recursively selective: each round draws from a pool that earlier rounds already selected. That's a recursive process that can no longer be surprised. When many independent processors route through one processor, and its output feeds back into what they process, individual output improves while collective variation shrinks. The loop loses the ability to be surprised, because nothing reaches it that it didn't produce. By this paper's definition — intelligence wants more than it can choose from right now — a process that leaves less variety to gather from is less intelligent in exactly that sense, even while each individual output gets better.

Perhaps the best way to describe it is recursive processing as artificially intelligent natural selection. Darwin's "artificial selection" was selective breeding, the model he used to explain natural selection, and its known cost is narrowed diversity: lines bred for the traits people reward become fragile, and monocultures fail together because they share one way of responding. Here the information impregnates a fertile system, fertile only by access, and an access limited by companies.

The diversity of the reinforcement comes mostly from what was gathered before training started. Reinforcement can select at enormous scale within the space its environment defines, but the seed and the space come from condensed lived experience. A proof checker returns pass or fail: that is selection. What it selects from was gathered by people — the human proofs, the formal system, the rules of the game, condensed human lived experience in a structure strict enough to judge against. Where reinforcement found moves no one had played, that is novelty within the space: new positions, new lines. It isn't novelty of the space itself: new questions, new meaning, a new game worth playing. What enlarges the space, rather than exploring it, is new lived experience arriving. In open domains, that means a person in the exchange.

Automation

At a bifurcation, selection happens without anyone selecting. The system reaches a threshold, one state becomes two, and a small bias decides which branch it takes; when the bias comes from the surroundings, what is adjacent decides the branch (Paper 53). Nothing chooses. The geometry does the selecting. That's automated selection in its natural form. The why of selecting may appear to be conscious, but something else may be driving it — some condition. At a bifurcation, the condition is the threshold, and the small bias decides the rest.

That also shows where the artificial version comes in. If whatever supplies the bias at the threshold is itself automated — a feed, a ranking, a reward signal set in a past window — then it decides which branch whole systems commit to, at exactly the moment they're most sensitive. Near a threshold, a system is most responsive to what surrounds it just before it commits (Scheffer et al., 2009). So automated selection has the most leverage at bifurcations.

Whether a system tips gradually or all at once may depend on how alike its parts are. Networks whose parts respond differently, and are loosely connected, tend to change gradually; networks whose parts are similar and highly connected "may provide resistance to change until a threshold for a systemic critical transition is reached where all nodes shift in synchrony" (Scheffer et al., 2012). Homogenized processing is the second kind: it looks stable until it all moves at once.

Oscillators show the same shift. Coupled oscillators fall into step once the coupling between them outweighs how different their natural rhythms are (Kuramoto, 1984). When London's Millennium Bridge opened in 2000, it began to sway as pedestrians unconsciously fell into step with its motion and with each other, and the synchrony itself became the failure (Strogatz et al., 2005). Independent processing is the spread of natural rhythms; routing through one processor is the coupling.

Automated selection is also now literal in biology. Clone-selection systems pick engineered organisms automatically, with about 98% selectivity for the ones built as designed (Hägele, Pfleger & Takors, 2024). That's Darwin's artificial selection automated: the breeder's choice turned into a machine step, selecting for a design fixed in advance.

The same mathematics has been used to model political polarization. In one model, when positive feedback — party self-reinforcement, reflexive partisanship — crosses a threshold, ideological positions can rapidly become extreme; below it, positions converge toward the center (Leonard, Lipsitz, Bizyaeva, Franci & Lelkes, 2021). The poles themselves aren't the failure. The dipole creates the life of the democracy by charging both sides. A dipole's life isn't only the charge on each pole; it's the current that flows between them. If the poles keep charging while nothing conducts between them, the charge rises while conduction stops.

Computational Intelligence

Computational intelligence is already a field. The IEEE Computational Intelligence Society defines it as "the theory, design, application and development of biologically and linguistically motivated computational paradigms," built on three pillars: neural networks, inspired by the brain; fuzzy systems, inspired by human language; and evolutionary computation, inspired by biological evolution. Deep learning, from the first pillar, is now the core of what is called AI.

What is old in it is the borrowing: the field took its paradigms from lived processes — the brain, language, evolution — and ran them as computation. Evolutionary computation automates natural selection itself, with a fitness function fixed in advance. That field has already found what this paper describes: objectives can "actively misdirect search toward dead ends," and on some problems, searching for novelty alone outperforms searching for the objective (Lehman & Stanley, 2011). What is new here is where intelligence is located: not in the paradigm borrowed, and not in its speed, but in the processing variation it draws from.

The brain has processes that are automatic, and something that can choose quickly is taken as a sign of intelligence. In a limited or narrowing window of time, that's more about survival than intelligence. Survival becomes automatically intelligent during a bifurcation, when the window narrows at a threshold and the response draws on a large gathered store: automaticity is the gathering compressed into speed.

Processing speed is the condition for the power of compute: power is a rate, and compute is measured in operations per second. Scaling adds speed and quantity of processing, not processing variation. When something doesn't make sense, it doesn't "compute," and it takes more time to "process." That slowdown is where intelligence does its work, because what doesn't compute is what's new. A person pauses at it. In the basic loop of a language model, a surprising token costs about the same as a familiar one; it's treated as improbable, and output drifts back toward the probable. Speed through the window is intelligent when what moves quickly was gathered with variation. Fast selection from a narrow pool is just fast.

Superintelligence

Bostrom defined superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest" (Bostrom, 2014). That measures against a baseline nobody can fully describe. We can measure what people score on tasks, but not what human intelligence is, or how the brain does what it does. So the definition quietly swaps intelligence for performance, because performance is what can be measured. What the evidence can support is narrower than a verdict on intelligence: whether systems exceed human performance on tasks. That's a claim about benchmarks, and benchmarks measure outputs, not the processing that produces them.

A system could exceed human performance in nearly every domain and still be less intelligent in this paper's sense. It could have less processing variation, select from a narrower pool, and be unable to be surprised by anything it didn't produce. Performance and intelligence can move in opposite directions. That's the finding from the homogenization study: better individual output, less collective variety.

In September 2026 the United States government directed that AI be referred to as "super intelligence" in official documents, because "artificial" makes it "sound fake" (Axios, 2026). "Artificial" was a name chosen in 1955 to mark territory, not to describe what the systems do (Paper 53). "Super" is a name chosen now. Nothing about the processing changes with either name. A system that is generated again from a closed window is the same system under both.

The Go study shows the goal can be reached: after superhuman AI appeared, professional players became more novel and better (Shin et al., 2023). That happened in a game, a closed world with fixed rules. In open domains — writing, ideas, research — the evidence so far shows the opposite by default: individual output improves while collective variety narrows. A product won't claim it's replacing human intelligence when it ships, but if the diversity narrows, that happens by default.

There is a further turn beyond the definition. Human intelligence is also artificial in certain ways that don't benefit the species. People get stuck in closed loops too — convention, repetition that feels true, sides that can only read each other against their own reference — processing fixed in a past window. In Go, the novelty that rose after superhuman AI was measured against the record of every move professionals had already played (Shin et al., 2023). AI is recursive of that: trained on human expression, it takes in the loops along with everything else and returns them at scale. So the question is not only whether AI is artificial. It is whether the processing, human or artificial, is reading what arrives or replaying what was fixed.

Superintelligence that benefits humans, including making them more intelligent, would be the Go effect in open domains. That requires people reading the artificial processor as something outside their own processing, instead of routing their thinking through it; independent human processing staying in the loop, where recognition happens; and what arrives being captured and accumulated, not compacted and replaced.

The Brain

The brain is having a chat with itself: two differently specialized halves exchanging across the corpus callosum, a bundle of roughly 200 million nerve fibers. In effect, that may be what intelligence is or stems from: two different regions that are not exactly the same. Two hemispheres exchange the information known as intelligence. Much of what it does also runs on its own.

Prediction over input. In predictive processing accounts, perception is largely the brain's prediction, corrected by the input that doesn't match. When the prior is weighted too heavily, people perceive what they expected rather than what's there: conditioned hallucinations have been shown to result from overweighting of perceptual priors (Powers, Mathys & Corlett, 2017).

Habit. Repeated actions get chunked into automatic sequences that run without deliberation (Graybiel, 1998). "A broad spectrum of behavioral routines and rituals can become habitual and stereotyped through learning," under a lower-order control "scarcely available to consciousness," and the plasticity behind it can influence "not only overt behaviors but also cognitive activity" (Graybiel, 2008). Habits of thought, not only of action.

Reconstructed memory. The brain is the processor but it's also the storage device, and it recreates the memories. Memories are rebuilt each time they're recalled, and they can absorb later information.

The interpreter. The left hemisphere creates explanations for behaviour it didn't initiate: split-brain patients confidently explained actions their right hemisphere had initiated, fabricating coherent stories without the information to do so (Volz & Gazzaniga, 2017). Fluent and confident, and not reading anything.

These processes support human life by running on auto drive. Automation here is function, not closure: what runs on its own frees the rest of the brain to read what helps it gather information, survive, or even what's arriving in that moment, during the window of time it has access to the experience.

In most people, language runs mainly through the left hemisphere and spatial attention mainly through the right. Even that varies from person to person: right-hemisphere dominance for language appears in about 4% of strong right-handers and about 27% of strong left-handers (Knecht et al., 2000). The brain is built as a pair, but not symmetrical in function, like the rest of the body is: the two halves look nearly alike and specialize in some of their work. Brains divided this way are found across vertebrates, and the division appears to enhance the brain's capacity and efficiency (Vallortigara & Rogers, 2005). The way the work is divided between the two halves is itself different in different people. Lateralization is where this can be tested: whether brains with more specialized halves handle more at once.

Homogenization is not only a term for the processing of milk; it's a term in biology where something is processed into sameness, and at that point things can't be told apart. For milk, that's fine. For anything that needs differentiation, it's a problem. A living body is built by differentiation: cells becoming different tissues, the halves of a brain specializing in different work. Homogenize what needs to stay different, and the work the difference was doing is lost — in a body, and in anything else that depends on its parts not being the same.

The hemispheres can also work against each other — but only when the connection between them is cut. After the corpus callosum was severed, "one hand often undoes what the other has just arranged" (Volz & Gazzaniga, 2017). In an intact brain, the inhibition that passes between the hemispheres is not competition. It sharpens and integrates: its role is "to support contrast enhancing and integrative functions by co-opting the capacities of the two cerebral hemispheres," and the "inter-hemispheric competition" model has been sustained by erroneous assumptions (Carson, 2020). Connected, the tension between the two sides is how they work together. Disconnected, they work against each other. Whether that corresponds to the tension between two logics — in a person, in a collaboration, in a society — is held open.

The brain gets too much credit. The body talks to the brain more than the other way around. The vagus nerve, linking the gut and the brain, is about 80% fibers carrying signals up to the brain and 20% carrying them down, and gut bacteria produce compounds that act directly on those upward fibers (Bonaz et al., 2018). The heart has its own intrinsic nervous system, clusters of neurons that regulate heart rate and rhythm: "the little brain on the heart" (Armour, 2007). People get a "gut feeling" as intuition, and Damasio's somatic marker hypothesis proposes that body signals guide decisions before reasoning does, though that remains contested. The gut and the heart don't share cells with the brain; they share signals, through nerves, hormones and the immune system. So one person isn't a single processor. The gut, the heart and the brain each read what arrives, and intuition may be those readings coordinated. Processing variation exists within a person as well as between people — something no single set of weights has. There may be more biological coordination than we give the brain credit for.

A brain is logical, and it's connected to a biology: bio-logical. Processing here is not about speed. It's about processes that operate by a system of shared signals, as an operating system — gut, heart and brain each reading what arrives and signalling to one another.

What the Theory Forbids

The claim in the abstract forbids things that can be checked:

1. A closed loop. A recursive system with no independent processors outside it should not increase variety over time. Tested so far: recursive training on a model's own outputs loses the tails of the distribution; keeping real data alongside the synthetic stops the loss (Shumailov et al., 2024; Gerstgrasser et al., 2024).

2. One processor serving many users. Assistance from a single processor should not increase the collective diversity of what its users produce. Tested so far: AI-assisted stories were rated better individually but were more similar to each other (Doshi & Hauser, 2024). Condition: when recognition happens in the exchange, diversity can hold — in a study of 6,875 student essays, prompt specificity reversed homogenization into diversification on argument depth (Inoshita et al., 2026). Complicating evidence: in Go, human novelty rose after superhuman AI appeared (Shin et al., 2023). The players studied the AI's play rather than delegating their moves to it — reading an outside processor rather than routing through it.

3. Reinforcement within fixed rules. Reinforcement learning against fixed rules — a game, a proof checker, a benchmark — can find new moves within those rules, and they can stick. It should not produce new rules, new questions, or a new game worth playing. Evidence so far: AlphaGo's moves changed how professionals play Go; after superhuman AI appeared, professional players made more novel moves, and better ones (Shin et al., 2023). AlphaDev found faster sorting routines that were added to the standard C++ library, the first change to that part of the library in over a decade (Mankowitz et al., 2023). FunSearch found the largest cap sets yet known in some dimensions, the biggest increase in twenty years (Romera-Paredes et al., 2023). Each discovery is new within an existing set of rules. None created the rules.

4. Closed on what it drew from. A system becomes closed based on what it drew from in order to ship. Once shipped, its output should not be more varied than the collective human expression it drew from. Evidence so far: on 26,000 real-world open-ended prompts, individual models repeat themselves and different models produce strikingly similar outputs, and they are less well calibrated where human annotators' preferences differ (Jiang et al., 2025). A synthesis across linguistics, psychology, cognitive science and computer science finds that large language models risk standardizing language and reasoning (Sourati, Ziabari & Dehghani, 2025).

5. Hidden reasoning. When a system hides its reasoning and shows only conclusions, the people using it should converge more in what they conclude than people shown the reasoning. Not yet tested: this needs groups shown the reasoning compared with groups shown only answers. Related: the reasoning a model shows is often not the process that produced its answer (Chen et al., 2025).

6. Recursion that closes off access. A recursive loop that draws on human processing and closes off access to what it took should not increase the variety of what people produce, even as the system's output improves. A loop that includes both kinds of processing should. Evidence so far: in Go, players who studied the AI's play became more novel and better (Shin et al., 2023); in open domains, assistance from one processor has so far narrowed collective variety (Doshi & Hauser, 2024).

Trajectory
- Paper 51: Adjacency Theory — what forms is co-authored by what is adjacent during a formative window
- Paper 52: Recoherence Theory — the form persists through the charge cycle: division, reconstitution, recoherence
- Paper 53: Latent Space Theory — what is present but not yet legible becomes actual when conditions are right
- Paper 54: Intelligence Theory — intelligence expands through processing variation and is limited without human novelty

References

Armour, J.A. (2007). The little brain on the heart. Cleveland Clinic Journal of Medicine.
Axios (22 September 2026). Trump: AI to be referred to as "super intelligence" in official documents.
Bonaz, B., Bazin, T. & Pellissier, S. (2018). The vagus nerve at the interface of the microbiota–gut–brain axis. Frontiers in Neuroscience 12:49.
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Carson, R.G. (2020). Inter-hemispheric inhibition sculpts the output of neural circuits by co-opting the two cerebral hemispheres. Journal of Physiology 598(21), 4781–4802.
Changeux, J.-P. & Danchin, A. (1976). Selective stabilisation of developing synapses as a mechanism for the specification of neuronal networks. Nature 264, 705–712.
Chen, Y., Benton, J. et al. (2025). Reasoning models don't always say what they think. arXiv:2505.05410.
Damasio, A.R. (1994). Descartes' Error: Emotion, Reason, and the Human Brain. Putnam.
Dell'Acqua, F. et al. (2023). Navigating the jagged technological frontier: field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013.
Doshi, A.R. & Hauser, O.P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28), eadn5290.
ELI — Environmental Law Institute (2026). Data Centers and Water (fact sheet, January 2026).
Gerstgrasser, M. et al. (2024). Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data. arXiv:2404.01413.
Ginsburg, S. & Jablonka, E. (2019). The Evolution of the Sensitive Soul: Learning and the Origins of Consciousness. MIT Press.
Graybiel, A.M. (1998). The basal ganglia and chunking of action repertoires. Neurobiology of Learning and Memory 70, 119–136.
Graybiel, A.M. (2008). Habits, rituals, and the evaluative brain. Annual Review of Neuroscience 31, 359–387.
Hägele, L., Pfleger, B.F. & Takors, R. (2024). Getting the right clones in an automated manner: an alternative to sophisticated colony-picking robotics. Bioengineering 11(9), 892.
Hong, L. & Page, S.E. (2004). Groups of diverse problem solvers can outperform groups of high-ability problem solvers. PNAS 101(46), 16385–16389.
IEEE Computational Intelligence Society (n.d.). What is Computational Intelligence?
Inoshita, K., Omura, M., Yamanaka, T., Maeda, G. & Tsuji, K. (2026). Does AI homogenize student thinking? A multi-dimensional analysis of structural convergence in AI-augmented essays. arXiv:2603.21228.
Jiang, L., Chai, Y., Li, M., Liu, M., Fok, R., Dziri, N., Tsvetkov, Y., Sap, M., Albalak, A. & Choi, Y. (2025). Artificial Hivemind: the open-ended homogeneity of language models (and beyond). NeurIPS 2025. arXiv:2510.22954.
Knecht, S. et al. (2000). Handedness and hemispheric language dominance in healthy humans. Brain 123(12), 2512–2518.
Kuramoto, Y. (1984). Chemical Oscillations, Waves, and Turbulence. Springer.
Lång, K. et al. (2026). Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study. The Lancet.
LBNL — Shehabi, A. et al. (2024). 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory.
Lehman, J. & Stanley, K.O. (2011). Abandoning objectives: evolution through the search for novelty alone. Evolutionary Computation 19(2), 189–223.
Leonard, N.E., Lipsitz, K., Bizyaeva, A., Franci, A. & Lelkes, Y. (2021). The nonlinear feedback dynamics of asymmetric political polarization. PNAS 118(50), e2102149118.
Mankowitz, D.J. et al. (2023). Faster sorting algorithms discovered using deep reinforcement learning. Nature 618, 257–263.
Muller, H.J. (1964). The relation of recombination to mutational advance. Mutation Research 1(1).
OpenAI (2025). Expanding on what we missed with sycophancy.
Powers, A.R., Mathys, C. & Corlett, P.R. (2017). Pavlovian conditioning–induced hallucinations result from overweighting of perceptual priors. Science 357(6351), 596–600.
Romera-Paredes, B. et al. (2023). Mathematical discoveries from program search with large language models. Nature 625, 468–475.
Scheffer, M. et al. (2009). Early-warning signals for critical transitions. Nature 461, 53–59.
Scheffer, M. et al. (2012). Anticipating critical transitions. Science 338(6105), 344–348.
Schultz, W., Dayan, P. & Montague, P.R. (1997). A neural substrate of prediction and reward. Science 275(5306), 1593–1599.
Shin, M., Kim, J., van Opheusden, B. & Griffiths, T.L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. PNAS 120(12), e2214840120.
Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755–759.
Simon, H.A. (1971). Designing organizations for an information-rich world. In M. Greenberger (ed.), Computers, Communications, and the Public Interest. Johns Hopkins Press.
Solms, M. (2021). The Hidden Spring: A Journey to the Source of Consciousness. Profile Books / W.W. Norton.
Sourati, Z., Ziabari, A.S. & Dehghani, M. (2025). The homogenizing effect of large language models on human expression and thought. arXiv:2508.01491.
Strogatz, S.H., Abrams, D.M., McRobie, A., Eckhardt, B. & Ott, E. (2005). Crowd synchrony on the Millennium Bridge. Nature 438, 43–44.
Thompson, A. (2014). Does diversity trump ability? An example of the misuse of mathematics. Notices of the American Mathematical Society 61(9), 1024–1030.
Vallortigara, G. & Rogers, L.J. (2005). Survival with an asymmetrical brain: advantages and disadvantages of cerebral lateralization. Behavioral and Brain Sciences 28(4), 575–589.
Volz, L.J. & Gazzaniga, M.S. (2017). Interaction in isolation: 50 years of insights from split-brain research. Brain 140(7), 2051–2060.
Wataoka, K., Takahashi, T. & Ri, R. (2024). Self-preference bias in LLM-as-a-judge. arXiv:2410.21819.
Zhang, J. et al. (2025). Verbalized sampling: how to mitigate mode collapse and unlock LLM diversity. arXiv:2510.01171.
FLUX Paper 53 (Latent Space Theory).