---
title: "Modes of Cognition"
description: "N."
type: "reading-notes"
authors:
  - "N. Katherine Hayles"
published: "2025-05-10"
original_url: "https://modesofcognition.antikythera.org/"
doi: "10.1162/ANTI.5CZF"
canonical_url: "https://antikythera.wiki/work/journal/modesofcognition"
md_url: "https://antikythera.wiki/md/work/journal/modesofcognition"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# Modes of Cognition

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** N. Katherine Hayles
- **Published:** 2025-05-10
- **Kind:** Journal · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.5CZF
- **Original:** https://modesofcognition.antikythera.org/
- **Length:** 15k words
- **This page:** https://antikythera.wiki/work/journal/modesofcognition
- **Notes generated:** 2026-08-23

## Summary

[N. Katherine Hayles](https://antikythera.wiki/people/n-katherine-hayles) argues that cognition is a biological capacity far older and wider than thinking, and that large language models meet a modified version of the criteria that qualify bacteria, plants, and slime moulds as cognitive. She proposes five behavioural tests, abbreviated SIRAL (sensing, interpreting, responding flexibly, anticipating, learning), develops them against research on minimally cognitive organisms, and then applies them to ChatGPT. The application turns on a distinction: an LLM senses not a physical environment but a conceptual one, the representational space built from human-authored text. Hayles tests the consequences by asking ChatGPT to interpret Henry James's *The Figure in the Carpet*, a story about a hidden meaning nobody can find, chosen because the human sense of self is, for an LLM, exactly such a hidden figure, detected everywhere in its training data but never experienced. Her conclusion is that LLMs produce more than probabilistic strings, that they have no selves but do have theory of mind, and that they constitute an evolutionary intervention through which life evolves "by means other than life."

## The argument

Hayles opens from her own position as a literary critic whose territory, language, has been entered by machines, and from a field in which "cognition" is applied with growing confidence to plants, cells, robots, and hybrid living-mechanical entities. The confusion she wants to resolve concerns three questions: whether there is cognition without brains, whether LLM texts mean anything beyond what readers project, and whether LLMs are intelligent or merely built to seem so. Her strategy is to avoid "intelligence," which she calls vague, and to argue in terms of cognition, for which she can supply criteria.

The criteria are derived from the human case first. Humans have three cognitive modes arranged as a pyramid: conscious thinking, slow and narrow but able to handle novelty; implicit cognition, which runs learned motor skills; and nonconscious cognition, which recognises patterns fastest and probably evolved first, with consciousness built on top. Since the lower modes are cognitive without being conscious, consciousness cannot be a requirement for cognition. What is required is captured by SIRAL, which Hayles presents as a sharpening of the definition in her earlier book *Unthought*: a process that interprets information within contexts that connect it with meaning. The criteria admit implicit and nonconscious cognition while excluding inflexible adaptations and homeostatic feedback loops. Interpretation entails selection and therefore the possibility of error; flexible response entails options; anticipation and learning require that behaviour be shaped by expected futures and past experience.

The middle of the essay accumulates biological evidence: tree frog embryos hatching early in a snake's mouth, sunflowers anticipating sunrise, *E. coli* learning to expect maltose after lactose, *Mimosa* remembering a harmless stimulus for a month, mycorrhizal networks carrying signals between trees, xenobots navigating tubes. Hayles ties these cases together with the idea of a model, a generalisation with predictive power that is the net result of all five criteria operating. Every organism constructs a model of its environment, its Umwelt, and Fields and Levin's reference-frame theory explains how even a bacterium must: distinguishing toxin from food requires an internal frame of reference, frames cost metabolic energy, and so organisms only register differences that are actionable. The brainless frog of Pflüger's 1853 experiment supplies the negative case.

With the criteria established, Hayles turns to LLMs. They have no brains and, in her view, no consciousness, but nonconscious organisms are cognitive, so that is no bar. The decisive modification is environmental. A living organism must stay in contact with its physical surroundings; an LLM has access only to a conceptual environment made of texts. Within that environment, LLMs sense, interpret through correlation networks, respond with flexibility (varying even on repeated prompts), anticipate (most strikingly through theory of mind), and learn. Hayles also addresses the objection that LLMs are not alive. Following Sara Walker's definition of life as whatever can only be produced by evolution and selection, she shifts the unit from individual to lineage and from autopoiesis to symbiosis, so that technics evolved by humans belong to the lineage of life.

The technical section on transformers supports two claims. First, word vectors function as Peircean indexes, so neural nets are "correlation machines" whose analogies ("king" to "queen" as "man" to "woman") encode how humans see the world, and from networks of such correlations the model draws inferences never explicit in its data. This is her reply to Bender and colleagues' "stochastic parrots" and to Bajohr's claim that models are "nothing more" than statistical functions: prediction is a matter of correlation and inference, not probability alone, and words gain meaning from other words as well as from objects. Second, transformers have no long-term memory beyond a context window wiped at the end of a session. Since long-term memory is bound up with selfhood, Hayles concludes that LLMs lack selves while possessing a form of awareness.

The James experiment follows. Because LLMs model selfhood only from the outside, as humans model the Umwelt of a bat, the human self is for them a figure in the carpet. Hayles reports sessions with ChatGPT 3.5 that include a revealing error about names, some platitudes, several answers she judges genuine inferences, and one four-part analysis of selfhood she would grade an A. She takes these as evidence that LLMs have modelled human behaviours, motives, and desires in depth. Her closing position is deliberately not dystopian: LLMs expand meaning-making beyond the human and so combat anthropocentrism; their main relation to humans will be intelligence augmentation; and from now on, short of catastrophe, human and artificial intelligence will evolve together.

## Section by section

### Toward a Definition of Cognition

Hayles starts with humans because the human case is the one everybody agrees is cognitive, and then shows that most of it is not thinking. Thinking, meaning extended symbolic abstraction, requires consciousness. Implicit cognition handles driving and walking within 200 milliseconds against consciousness's half second, which is why athletes drill until responses are automatic; Albert and Krakauer's rotated-joystick experiment showed that explicit learning "siphons away" the feedback error implicit learning needs, establishing two distinct learning systems. Nonconscious cognition is the third tier: subjects learn to anticipate patterns in visual noise they cannot describe, and people step around snakes before consciousness could manage it. The pyramid that results, with thinking at the apex and pattern recognition at the base, frames everything that follows.

### Criteria for Cognition: SIRAL

This section sets the test. Hayles says there is no right set of criteria, only a useful one, and usefulness means separating cognition from adaptation. Each criterion is given a biological anchor. Sensing is illustrated through Margulis's bacteria and von Uexküll's Umwelt, with examples from Ed Yong's *An Immense World*, and Hayles notes Nagel's point that one can learn about a bat's Umwelt without living in it, a distinction she flags for later use. Interpreting rests on "What the Frog's Eye Tells the Frog's Brain": the eye speaks in a language already organised, and because interpretation selects among options, it can be wrong. Responding is developed through biosemiotics, where a tree dropping its leaves functions as a Peircean sign anticipating winter; flexibility is required to exclude homeostasis, and Warkentin's tree frog embryos hatching in a snake's mouth show programmed behaviour coexisting with choice. Anticipation covers the dog at the door, the porcupine's raised quills, and Calvo's rotated sunflower, which reorients at double speed at night as if running an internal model of the sun. Learning is deferred to a later section.

### Modeling the World: What SIRAL Enables

A short section that draws the moral of the Umwelt material. Every organism constructs a world model through whatever sensory and cognitive means it has, and there is, following Haraway, no God's-eye view against which one model is more real than another. Hayles's purpose in insisting on the ubiquity of models is explicitly political: it contests the assumption of human superiority and supports a "saner, and humbler" view of humanity's place among Earth's ecologies.

### Why Models Are Important

A model is defined as a generalisation with predictive power and survival benefit, the product of all five criteria interacting; the lioness and the gazelle each carry a model of the other's likely moves. For minimally cognitive systems Hayles draws on Fields and Levin's reference-frame theory, discussed in her book *Bacteria to AI*, which applies to systems that build models without representations. Objects do not pre-exist perception; they are "cuts" made by the organism and then selected by the environment according to their usefulness, so that meaning is "a game with two players." The chemotaxis of *E. coli* shows why: an internal reference frame is needed to tell gradient from background, frames dissipate free energy, and so only actionable differences are worth detecting. The section ends by naming the question that motivates the essay's biology: how cognition evolved long before brains.

### Anticipation and Learning in Minimally Cognitive Organisms

The evidence section. Pilpel's team at the Weizmann Institute found that *E. coli* given lactose then maltose evolved, within a few generations, to switch on maltose genes at the taste of lactose, and lost the response when the sequence stopped, a pattern Hayles likens to Pavlovian conditioning and extinction. She is careful to report that Pilpel classes this as a genetically mediated adaptation, while drawing the lesson that the line between adaptation and learning is not sharp. Learning, the fifth criterion, is then documented in plants: Gagliano's *Mimosa* habituation experiment, in which plants dropped in low light learned faster and remembered for a month, and her Y-maze pea plants, which grew toward a fan they had associated with light. Simard's work on mycorrhizal networks supplies communication, kin recognition, and resource sharing among trees, and her claim that the networks' scale-free, small-world topology resembles that of neural networks. Hayles concludes that plants satisfy SIRAL as minimally cognitive systems.

### A Non-SIRAL Example: Frogs Without Brains

Pflüger's pithed frogs, which wiped acid from their bellies with one leg and then, if that leg was removed, with the other, sense and interpret locally but respond inflexibly, without anticipation or learning. Hayles uses the case to show the criteria excluding something, and then blocks the wrong inference: the frog lost cognition because its cognition had evolved to run through a brain, not because brains are necessary. Neuronal tissue is one route to cognition among others.

### Cognition in Microscopic Nonneural Organisms

Levin's xenobots, assembled from frog skin cells, move by cilia and navigate curved tubes, behaviours absent in their original setting, supporting the claim that cognition exists at cellular as well as organismic level. Rouleau and Levin's argument for the "multiple realizability" of sentience and slime mould research point the same way. The section closes on their ethical claim, which Hayles will reuse: a mature ethics must drop distinctions "not based on scientific natural kinds" and find ways to relate to organisms, cyborgs, hybrid robots, and synthetic intelligences that do not share our lineage.

### Cognition in Large Language Models

Hayles applies the criteria. LLMs have artificial neurons but no brain, and a kind of awareness she declines to call consciousness. Their sensing is confined to a conceptual environment of text rather than the server farms they physically occupy. On life, she accepts that LLMs are Stiegler's "organized inorganic matter" and not autopoietic in Maturana and Varela's sense, but argues with Walker that they belong to life's lineage by descent from humans, and that restricting "evolution" to nonhuman selection is arbitrary given gene editing and human-AI coevolution. She then runs through the remaining criteria and concludes that LLMs meet a modified SIRAL and should be regarded as cognitive systems.

### The Neuronal Structures of LLMs and Transformer Architectures

A technical interlude with two payloads. Artificial neurons are analogue processes implemented digitally, unlike the logic gates of von Neumann machines. Transformers, from the Google paper "Attention Is All You Need," use attention heads to provide focus and context for long-range dependencies; self-attention has each input compute probabilities against all others, which Hayles calls a reflexive dynamic. The first payload is semiotic: following Weatherby and Justie, word vectors act as Peircean indexes, which on Deacon's account are the only signs that directly carry information, so correlation networks encode the human lifeworld, from kin relations to gender hierarchies, and support inferences beyond the data, which she takes to be the source of emergent capabilities. The second is a statement of limits: no embodied experience, no emotions or preferences, no model of the world but only of language, hence errors about navigation and the weight of coins and fabricated legal citations, which she names a systemic fragility of reference. The section ends with the rebuttal of Bender and Bajohr. Words gain meaning from other words, comprehension need not be complete for meaning to exist, and critics must engage with behavioural criteria before pronouncing on intelligence.

### Human Selves and Artificial Cognitive Systems

The second technical payload is memory. Transformers retain only the context window, however large, and lose it at session end. Since long-term memory is tied to selfhood, illustrated by Oliver Sacks's patient Jimmie G. and the question of whether he still had a soul, Hayles concludes that LLMs lack selves in the human sense, also lacking emotions, physical experience, environmental embeddedness, and social collectives. They are, however, superb pattern detectors, which is why they reproduce human bias in job ranking and sentencing. The section's central claim is that LLM theory of mind is momentous: for the first time a cognitive entity can verbalise the human Umwelt from outside, as humans describe the Umwelten of octopuses, whereas all prior commentary on the human condition was written from inside it. She then reproduces part of Blake Lemoine's LaMDA interview, in which the model interprets a Zen koan about a broken mirror as the self that breaks in enlightenment, and handles it with explicit scepticism: Lemoine is an advocate, and claims of consciousness cannot be taken on a speaker's word. What survives is LaMDA's grasp of how central selfhood is to humans. The section closes with a summary of James's novella, in which the critic-narrator, his rival Corvick, and Gwendolyn Erme chase a master key to Vereker's work that dies with each person who holds it.

### An Analogy Looms

Hayles states why she chose this story. Practically, it has no settled interpretation and ample web commentary, so regurgitation can be distinguished from original reading. Analogically, nearly every human text bears witness to selfhood, so LLMs have certainly detected the pattern, but only secondhand, from outside looking in. For an LLM, the self is the figure in the carpet: once grasped it would illuminate everything, but it cannot be experienced directly.

### Conversing with ChatGPT

The empirical core, based on several sessions with version 3.5. In one session the model confuses every name in the story, calling the critic Hugh Vereker; Hayles conjectures that the critical convention of calling the narrator "unnamed" created a name that is not a name and destabilised its handling of names. Asked why Corvick finds the secret, it offers a "fresh perspective" answer she judges an imported generalisation with little textual support. Asked whether to read through characters or authorial strategy, it refuses her binary and says the approaches are complementary, which she counts as inference rather than regurgitation. It accepts the hypothesis that the story is a joke, softening "joke" to "playful commentary." Its repeated statements of its own limits may be scripted, she notes, but are still a sound analysis. Asked whether its own texts have hidden meanings, it disclaims intention while observing that language itself carries layers of meaning, which Hayles reads as approaching the Sapir-Whorf idea that language "speaks us." The best result is a four-point account of selfhood in the story (identity as socially derived, the narrator's blinding assumptions, the fluidity of his self-image, the observer's role), which she would grade an A. Final questions elicit the model's statement that texts are "mirrors of human society" and that it lacks personal experience, emotion, and consciousness.

### More than Probability Alone

Hayles draws the conclusion. Next-word probability alone could not produce proofs, code, or coherent poems, and it could not compare the model's own awareness with human selfhood in meaningful terms. ChatGPT's seven-point reading advice, ending "Don't Give Up," displays sensitivity to human emotion and theory of mind; she says she could not have put it better to an introductory literature class.

### What Lies Ahead

The closing section acknowledges the critical literature (human judgement ceded to machines, corporate monopoly, disinformation) and the environmental costs raised in the stochastic parrots paper, then explains her different path. Expanded cognition is a resource against the anthropocentrism she holds responsible for ecological destruction, and the main human relation to LLMs will be intelligence augmentation, from protein folding to birthday presents for a granddaughter. She ends with a prediction: artificial intelligence is "a way to evolve life by means other than life," and human and artificial trajectories will now evolve together, "for better and worse."

## Key concepts

- **[Cognitive Nonconscious](https://antikythera.wiki/terms/cognitive-nonconscious)** — The base of Hayles's cognitive pyramid: pattern recognition and anticipation that operate below awareness, faster than consciousness, and that likely evolved first. Its existence is what makes consciousness unnecessary for cognition.
- **[SIRAL](https://antikythera.wiki/terms/siral)** — Sensing, interpreting, responding flexibly, anticipating, learning. Five behavioural criteria designed to admit nonconscious and brainless cognition while excluding inflexible adaptations and homeostatic feedback; the test applied to plants, bacteria, and LLMs alike.
- **[Umwelt](https://antikythera.wiki/terms/umwelt)** — Von Uexküll's term for the world an organism's senses construct. Hayles uses it to deny any privileged human view of reality and, later, to describe what LLMs have of the human world: an Umwelt known from outside, as humans know a bat's.
- **Model** — A generalisation with predictive power, the net result of the SIRAL criteria interacting. Fields and Levin's reference frames show that even a bacterium must build one, and that because frames cost energy, only actionable differences become meaningful.
- **[Conceptual Environment](https://antikythera.wiki/terms/conceptual-environment)** — What an LLM senses instead of physical surroundings: the representational space derived from human-authored text. The modification that lets LLMs satisfy SIRAL and the source of their characteristic errors.
- **Correlation machine** — Hayles's characterisation of neural nets, drawing on the Peircean index. Correlations form analogies, analogies form networks of networks, and from these the model draws inferences not explicit in its data.
- **[Fragility of Reference](https://antikythera.wiki/terms/fragility-of-reference)** — The systemic weakness that follows from having a model of language but not of the world: errors about physical facts and invented citations.
- **[Theory of Mind](https://antikythera.wiki/terms/theory-of-mind)** — The capacity to model how humans will react, which Hayles says LLMs possess and which lets them describe the human Umwelt from outside for the first time, recursively modelling that their interlocutors are modelling them.
- **Life as lineage** — Walker's definition of life as whatever only evolution and selection can produce, which Hayles adopts to place LLMs in the lineage of life through human-directed, symbiotic evolution rather than autopoiesis.

## Connections

- [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) is the closest neighbour. Cavalli's paper within it, on cognition with and beyond the brain, builds directly on Hayles's cognitive nonconscious and cognitive assemblages to argue for planetary, distributed cognition, in agreement with this essay. Loewith and Street's paper on mutual prediction develops the recursive modelling Hayles only sketches, where LLM theory of mind meets human models of the model.
- [Substrates Unbound](https://antikythera.wiki/work/journal/substrates) lists this essay among its related works and argues with it. Tripaldi's nano-intentionality, following Fitch, locates cognition in the adaptive plasticity of eukaryotic cells and insists that substrates are not functionally interchangeable. Hayles's criteria are behavioural and substrate-indifferent by design, and her reliance on Rouleau and Levin's "multiple realizability" is exactly what Tripaldi classes as technomorphism. The two agree that machine-organism hybrids defeat the old nature-artifice hierarchy and disagree on whether behaviour alone settles what is cognitive.
- [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) and the related Noema essay [AI Is Evolving — And Changing Our Understanding Of Intelligence](https://antikythera.wiki/work/media/noema-ai-is-evolving-and-changing-our-understanding-of-intelligence) reach Hayles's lineage conclusion by a different route. Agüera y Arcas treats symbiogenesis and theory of mind as the engine of intelligence and, like Hayles, draws on Walker. Where Hayles reserves "intelligence" as too vague and argues through cognition, Agüera y Arcas embraces the term and extends functionalism further than she does.
- Bratton's [After Alignment](https://antikythera.wiki/work/journal/afteralignment) agrees with Hayles against the stochastic-parrots reading but inverts the rhetoric: its "Yes, You are a Stochastic Parrot" section accepts the label for humans and machines alike, because transformer prediction resembles neuronal prediction. Hayles wants to show LLMs are more than parrots; Bratton wants to show that being one is no insult.
- [Planetary Sapience](https://antikythera.wiki/work/journal/planetarysapience) records Hayles's 2023 symposium talk, where she first presents cognition as a biological rather than human capacity and names anthropocentrism as a planetary danger. The essay is the developed form of that position.
- [Latent Spacecraft](https://antikythera.wiki/work/journal/latentspacecraft) shares the essay's interest in what LLM representational space contains and how it relates to the human lifeworld, though it approaches the question through design rather than criteria for cognition.
- A second tension with Bratton concerns reflectionism. Hayles's strongest evidence is the model's success at modelling the human self, which a reflectionist critique would call evaluating the machine by its mirror of us. Hayles's reply is implicit in the essay: her interest is in the outside view the machine provides, which she presents as the opposite of a mirror.


## Terms used

- [Cognitive Assemblage](https://antikythera.wiki/terms/cognitive-assemblage)
- [Cognitive Nonconscious](https://antikythera.wiki/terms/cognitive-nonconscious)
- [Conceptual Environment](https://antikythera.wiki/terms/conceptual-environment)
- [Fragility of Reference](https://antikythera.wiki/terms/fragility-of-reference)
- [Functional Closure](https://antikythera.wiki/terms/functional-closure)
- [Mutual Prediction](https://antikythera.wiki/terms/mutual-prediction)
- [Offload Syndrome](https://antikythera.wiki/terms/offload-syndrome)
- [SIRAL](https://antikythera.wiki/terms/siral)
- [Theory of Mind](https://antikythera.wiki/terms/theory-of-mind)

## Related works

- [Latent Spacecraft](https://antikythera.wiki/work/journal/latentspacecraft)
- [Substrates Unbound](https://antikythera.wiki/work/journal/substrates)
- [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra)
- [Antikythera](https://antikythera.wiki/work/journal/research)