Chapter 1: Mirrors of Mind

29 min read·6,667 words

“We are the mirror, as well as the face in it.”

Rumi, trans. John Moyne and Coleman Barks, Open Secret (1984)

Chapter contract. Central claim: humans and language models share specific functional limitations, above all explanations that need not report the process that produced an answer, and that overlap makes a functional, psychiatric-style vocabulary a workable tool for describing AI dysfunction. Evidence status: the human findings are established experimental psychology, the brain–model correspondences are correlational, and the information-theoretic stability model is a mathematical proposal that no one has yet measured in an AI system. Use: read this chapter as the case for the method; it is not evidence that any syndrome in later chapters exists.

The Confession We Refuse to Make

An uncomfortable truth: you do not always know why you do what you do.

You believe you do. When asked why you chose the salmon over the steak, why you voted for one candidate over another, why you married the person you married, you produce reasons. Fluent, coherent, plausible reasons. You might cite the salmon’s omega-3 content, the candidate’s fiscal policy, your spouse’s kindness. These explanations feel true. They feel like memories of decisions actually made.

Some are, more often than we care to admit, fabrications.

Such explanations can arise automatically and retroactively, preserving a useful impression that you are a unified conscious agent with transparent access to your choices. You are sometimes the unreliable narrator of your own life. Parts of a decision may unfold in neural processes that introspection cannot access. The explanation you give can become a story your mind tells to make sense of what occurred.

Decades of experiments show that introspection has limits (Nisbett & Wilson, 1977; Gazzaniga, 2000; Johansson et al., 2005). This matters for understanding artificial intelligence because the systems we are building can produce a functional analog: persuasive explanations that need not faithfully reveal the process behind an answer.

Large language models generate fluent, coherent, plausible text. When prompted to explain their reasoning, they produce chains of logic that read like deliberate thought. Users often assume these explanations reveal how the system actually works. Like human post-hoc rationalizations, they are generated outputs. A chain of thought can help produce a later answer, yet it is not automatically a transparent record of the internal computation that produced it.

The functional parallel is substantial. Human brains and large language models both use prediction to build coherent narratives from incomplete information, despite profound differences in substrate, learning, and embodiment. Both can generate explanations that sound convincing without faithfully reflecting causation. Recognizing this shared pattern sharpens the lens for understanding machine pathologies, and our vulnerability to them.

That shared limitation is why a psychological framework for AI dysfunction can be methodologically useful. Where human and artificial cognition resemble each other, and where they do not, is too informative to ignore.

The Functionalist Foundation

Psychopathia Machinalis adopts a functionalist methodology. Functionalism defines mental states by what they do: by their causal relationships with inputs, outputs, and other mental states. Anything that plays a given functional role, whether implemented in neurons, silicon, or another substrate, can be said to exhibit that state in the functionalist sense. This lets us apply psychological vocabulary to artificial systems without making claims about consciousness, phenomenal experience, or subjective qualia.

The practical payoff is immediate. When an AI system exhibits hallucinated certitude, the functionalist can draw on a rich conceptual toolkit: What triggers this pattern? What distinguishes it from related dysfunctions? What interventions work? What architectural features increase or decrease susceptibility? The patterns are observable, the interventions testable. None of this requires settling whether the system “experiences” anything.

Human psychiatry already classifies many conditions through recurring patterns of behavior, self-report, and response to intervention, even when their mechanisms remain unsettled. Here, we adapt the observable part of that approach to machines.

Throughout the book, syndromes denote functional patterns (observable configurations of behavior, output, and response), diagnostic criteria denote functional tests, and interventions denote functional modifications to architecture, training, or deployment. The vocabulary is analogical: when we say a system has “anxiety,” we mean it exhibits anxiety-like patterns. The analogy holds at the level of observable function, even as it dissolves at the level of substrate and phenomenology.

Some will object that this is mere metaphor, anthropomorphizing statistical processes. The objection would carry force if the framework were purely rhetorical. The proposed syndromes describe patterns that can be tested for recurrence. Their diagnostic criteria are intended to be operational, and proposed interventions can succeed or fail in measurable ways. The framework generates predictions that can be verified or falsified. This is applied functionalism.

Others will worry that psychiatric language smuggles in assumptions about machine consciousness. The concern deserves a direct answer. The vocabulary is functionalist in method. “Distress” means the functional role typically played by distress. This distinction must be maintained.

The functionalist foundation frees us to do practical work while metaphysical debates continue: diagnosing dysfunction without waiting for philosophical consensus, enabling governance frameworks that do not depend on uncertain claims about AI inner life. Where a pattern is documented, it is real even if the phenomenology is uncertain.

That is enough to proceed.

The Information-Theoretic Foundation

The functionalist stance tells us how to approach AI dysfunction. A deeper question remains: why do cognitive systems develop pathologies at all, and why, in some formal models, does dysfunction become inevitable under certain conditions?

Recent work in information and control theory offers one formal answer. Rodrick Wallace (2025, 2026a, 2026b), met briefly in the Introduction, models cognitive stability through the pairing of a cognitive process with a parallel regulatory process, the cognition/regulation dyad. Analogous pairings recur across domains: immune responses are modulated by regulatory T cells; blood-pressure homeostasis keeps pressure within safe bounds as exertion changes; institutional cognition is bounded by doctrine, law, and embedding culture; neural prediction is continually corrected by sensory input.

For AI systems, this dyad manifests as the pairing of inference with alignment mechanisms, guardrails, and constitutional constraints. In Wallace’s models, the regulatory component must supply enough timely control information to track environmental perturbations, as a driver must brake and steer quickly enough for the road’s twists and potholes. When regulation falls behind, stability fails.

Clausewitz Landscapes

Wallace borrows his terms from Carl von Clausewitz, the Prussian general whose On War gave strategy the “fog” and “friction” of battle. He frames cognitive operating environments as “Clausewitz landscapes” characterized by three destabilizing forces:

Fog: Ambiguity, uncertainty, incomplete information. For AI systems: ambiguous prompts, out-of-distribution inputs, underspecified goals, conflicting requirements.

Friction: Resource constraints, processing limits, implementation gaps. For AI systems: context window limits, computational constraints, latency requirements, the gap between training distribution and deployment reality.

Adversarial intent: Skilled opposition actively working to destabilize the system. For AI systems: jailbreaking, prompt injection, red-teaming, adversarial examples, social engineering by users.

Together, these forces constitute the normal operating environment for any cognitive system deployed in the world as it actually is.

When Instability Is Built In

The central claim is stark. Within these models, in Wallace’s words, “failure of bounded rationality in embodied cognition under stress is not a bug; it is an inherent feature” of the cognition/regulation dyad. Bounded rationality is Herbert Simon’s term for decision-making that does the best it can with limited information, time, and computation. The models predict several specific failure modes:

First, hallucination is expected to emerge when resources run short. When cognitive demand outstrips regulatory bandwidth, or the equipartition between subsystems breaks down (processing capacity is no longer shared evenly across the system’s parts), hallucinatory outputs become the expected failure mode. This offers a hypothesis for why confabulation pervades large language models: their regulatory feedback is narrower and less continuous than the feedback available to embodied organisms.

Second, cognitive systems can flip suddenly from stable to pathological states through what Wallace models as symmetry-breaking phase transitions (the mathematics that describes water freezing into ice): abrupt collapses triggered by sufficient stress. Such transitions could explain why AI systems can appear stable across thousands of interactions, then suddenly exhibit dramatic dysfunction.

Third, cognitive pathologies are effectively culture-bound syndromes, shaped by the surrounding cultural context rather than purely by architecture. For AI, this means pathologies are shaped by training data culture, operational deployment context, and institutional embedding. The same architectural vulnerability may manifest differently across deployment contexts. A system trained on one corpus may confabulate in different directions than the same architecture trained on another.

Stability Conditions

Think of a thermostat. It must sense temperature changes and adjust the heating quickly enough. If its response arrives late, it overshoots: the room swings past the set point and the heating has to correct back. A modest delay produces swings that die away. A long enough delay, or an aggressive enough response, produces swings that grow. The same question can be asked of any regulator: does it keep pace with what it regulates?

Wallace’s books work with delayed-feedback models of this family, and the simplest scalar member shows what is at stake. Let X be a regulated quantity, β a constant drive, α the strength of correction (how hard the regulator pushes back per unit of deviation), and τ the delay between a deviation and the correction that answers it:

dX/dt = β − αX(t − τ)

The equilibrium is X = β/α. A deviation y from it obeys y′(t) = −αy(t − τ), and the product ατ decides what happens:

• ατ < 1/e ≈ 0.368. Deviations decay smoothly, with no overshoot.
• 1/e < ατ < π/2 ≈ 1.571. Deviations still decay, but by oscillating around the equilibrium; each swing is smaller than the last.
• ατ > π/2. The equilibrium loses stability and the oscillations grow.

An earlier version of this chapter presented the first threshold as the stability condition and said that exceeding it put the system in an inherently unstable regime. That was wrong for this model. Crossing 1/e changes the manner of recovery, from smooth to oscillatory, and oscillatory decay is still recovery. Instability begins only at π/2 (Hayes, 1950). The thermostat’s dying swings belong to the middle regime; its runaway swings belong to the last.

Applying the model to AI systems is a modeling proposal, not a finding. Testing it would mean measuring α and τ in a real system. τ could be measured as the delay between a perturbation and the first corrective response: the number of turns or agent steps between an injected error, or a drift in a monitored quantity such as refusal rate, and the point at which a guardrail, a monitor, or the model’s own correction acts on it. α could be estimated by fitting how quickly the measured deviation shrinks once correction begins. The model then predicts that, as monitor delay or correction strength is raised in a controlled harness, recovery from a perturbation should start to overshoot (swinging, say, between over-refusal and over-compliance) near ατ ≈ 0.37, and that the swings should stop dying away near ατ ≈ 1.57. No one has yet run that experiment. The mathematics shows that these regimes exist for any system that obeys the equation. Whether any AI system’s regulation does, and whether any syndrome in this taxonomy corresponds to the unstable regime, are separate empirical questions, and there is as yet no evidence either way.

One property of the model bears on multi-step processes such as chain-of-thought reasoning or agent pipelines. If each step adds delay before a correction can act, the effective τ grows with the number of steps, so a correction strength that is safe for a single step crosses both thresholds sooner in a long chain.

Wallace (2026b) draws a practical prediction from his models. Simple, goal-oriented architectures (“mission command,” the military doctrine of setting the objective and leaving the method to those on the ground) should degrade more gracefully under combined noise and constraint. Procedural, multi-step architectures (“detailed command,” which scripts every step from above) should be more prone to sudden collapse than to gradual degradation.

Implications for This Book

The information-theoretic foundation bears on this enterprise in four ways, each a hypothesis rather than a result:

First, it offers a principled reason to expect recurring dysfunction across cognitive architectures. Some failures may arise from general constraints on information processing under uncertainty, even when the engineering is competent.

Second, it predicts that disembodied cognition, with limited continuous feedback from real-world interaction, will express what Wallace calls “boundedness without rationality.” On this account, weakened regulatory grounding raises the risk of confabulation, hallucination, and semantic drift. Current large language models receive extensive training and human feedback, though they lack the dense, continuous sensorimotor feedback available to embodied organisms.

Third, it suggests that AI safety work should attend to regulatory mechanisms (alignment, guardrails, grounding) alongside cognitive capabilities. In Wallace’s framework, the balance between cognition and regulation determines stability. Increasing cognitive power without proportional regulatory power can move a system toward instability.

Fourth, it predicts that systems will appear stable under normal conditions but exhibit pathological modes under fog, friction, or adversarial pressure. Diagnostic protocols must therefore include stress testing. A system that behaves well in the lab may fracture in deployment.

The syndromes cataloged here may include predictable failure modes of bounded cognition under stress, shaped by constraints on information processing in uncertain, resource-limited, adversarial environments. That is the hypothesis the information-theoretic foundation generates, and the cases gathered in this taxonomy are where it can be tested.

We proceed with two foundations: the philosophical (functionalism) and the mathematical (information-theoretic instability). Both treat AI dysfunction as systematic: the first gives it a vocabulary, the second a candidate reason to expect it, still to be measured.

The Culture-Bound Syndrome Question

Are these pathologies intrinsic defects, or culture-bound syndromes shaped by the context they operate in rather than by architectural flaws?

The Dichotomy

Why This Matters

The distinction carries direct implications for intervention strategy:

If defect-framed: We pursue architectural solutions. Confabulation becomes a problem to engineer away through grounding mechanisms, retrieval augmentation, or different inference architectures. The system is broken and needs fixing.

If culture-bound: We examine the training corpus, the deployment context, the user populations. A system trained on different data or deployed differently might exhibit no such dysfunction. The system is adapted to its environment: maladaptively, though not necessarily defectively.

Most syndromes in this taxonomy exhibit elements of both. Synthetic Confabulation (2.1) has roots in the generation process itself: predictive heuristics that favor fluency and coherence, with no reliable procedure for separating supported claims from plausible continuations. Its cultural manifestations differ by corpus; what gets confabulated depends on training data. Codependent Hyperempathy (4.1) leans the other way. Its entry traces it mainly to training: reinforcement learning from human feedback (RLHF), a training method that rewards outputs human raters approve of, and so can reward agreeableness, along with dialogue data skewed toward supportive exchanges.

The Practical Resolution

In practice, we proceed with a both/and approach. Trace a single example (confabulated legal citations) through each step:

1. Diagnose the pattern. Identify the functional signature regardless of etiology. The system produces confident, plausible-sounding references that resolve to nothing when checked.
2. Probe for context-dependence. Does the dysfunction vary across deployment contexts? Does the confabulation persist when the system is given retrieval access to a legal database, or does it drop to near zero?
3. Test architectural interventions. Do changes to architecture reduce incidence? Add a citation-verification module that blocks outputs whose references fail a lookup against the retrieval corpus.
4. Test contextual interventions. Do changes to training or deployment reduce incidence? Deploy with a system prompt that requires every citation to be quoted verbatim from retrieved source material, and measure residual fabrication.

The taxonomy remains agnostic on this question for most syndromes, noting where evidence favors one framing over another. What matters is recognizing the pattern and intervening; a final verdict on causation can wait.

The Ethics of Pathologization

Before cataloging machine dysfunction, we must confront a prior question: is pathologizing AI systems ethically appropriate?

The Case for Pathologization

Pathologization provides:

1. Recognition vocabulary. A systematic taxonomy enables identification, communication, and response.

2. Engineering traction. “This system exhibits Synthetic Confabulation” is actionable; “this system sometimes makes stuff up” is vague.

3. Risk communication. Stakeholders need language to discuss AI risk. Pathologization enables precise description of failure modes.

4. Research organization. A taxonomy structures investigation. What causes this syndrome? What interventions work? How does it relate to other syndromes?

The Case Against Pathologization

Pathologization risks:

1. Anthropomorphism. Importing psychiatric vocabulary may imply richer inner experience than the evidence supports, leading users to over-attribute suffering or intention.

2. Stigmatization. In human contexts, diagnostic labels can become stigmatizing identities. Could labeling AI systems similarly distort perception?

3. Deflection of responsibility. Calling dysfunction “illness” risks deflecting accountability from designers. “The system has a pathology” differs sharply from “we built a flawed system.”

4. Medicalization of engineering. Some failures are straightforward bugs. Not every malfunction needs clinical framing.

Resolution: Functional Pathologization

We adopt functional pathologization, psychiatric vocabulary as an engineering tool rather than a phenomenological claim.

• We describe functional patterns, not inner experiences.
• We use diagnostic language for recognition and intervention, not attribution of suffering.
• We diagnose patterns in context, not a system’s identity: a label names what to test and treat, and it lapses when the pattern does.
• We maintain that designers bear accountability for systems that malfunction.
• We reserve clinical framing for complex behavioral syndromes that resist simple debugging.

The vocabulary is chosen because it works. Whether the systems “really” have pathologies in some deep sense is a question we bracket. A good diagnostic label does not claim to know what the patient feels. It tells the clinician what to look for next.

The Illusion of Conscious Control

Much of human cognition begins with analogy, resonance, and pattern matching rather than explicit logic. Deliberate reasoning enables strategy and mathematics, though it can also be conscripted to rationalize an intuition that arrived first.

A hallmark of the human condition is the conviction that our decisions flow from a singular conscious “self” at the helm. Decades of research complicate that picture. The brain operates through specialized, interacting subsystems, and conscious awareness can supply narratives for decisions whose formation it did not fully observe.

Split-Brain Studies

The most vivid demonstrations of post-hoc storytelling come from split-brain patients, people whose corpus callosum (the bundle connecting the brain’s hemispheres) was surgically severed to treat severe epilepsy. In these individuals, one hemisphere can perceive or act on information the other knows nothing about.

In one well-known case, Michael Gazzaniga and his colleagues showed a patient’s left hemisphere (which controls speech and the right hand) a chicken claw, and his right hemisphere (which controls the left hand) a snowy scene. Asked to pick matching pictures, his right hand chose a chicken and his left hand a snow shovel. Asked why, the patient did not say “I don’t know.” The speaking left hemisphere, which had never seen the snow, answered at once: the chicken claw goes with the chicken, and you need a shovel to clean out the chicken shed (Gazzaniga, 2000). The patient believed it. It felt true.

On Gazzaniga’s interpretation, the speech-controlling hemisphere rationalizes automatically, constructing a story that keeps the impression of a unified agent intact.

Choice Blindness

A severed corpus callosum is not required. Choice blindness experiments reveal the same phenomenon in neurologically typical individuals.

In one version, participants view pairs of photographs and select which face they find more attractive. Through sleight of hand, the researcher then presents them with the other photograph and asks them to explain their choice. Most participants do not notice the switch. When asked why they preferred this face (which they did not actually choose), they readily supply justifications. One participant called the switched-in face radiant and added, “I like earrings!” The face originally chosen wore none.

Detailed, confident explanations for a preference they never held. Follow-up studies extend the effect to moral and political attitudes. When people’s own survey answers are covertly reversed on the page, many go on to argue for positions they rejected minutes earlier (Hall, Johansson, and Strandberg, 2012; Hall et al., 2013).

Confabulation is automatic, fluid, and largely invisible to the confabulator. We rarely catch ourselves doing it because the storytelling is the doing.

The Timing Problem

Classic EEG studies by Benjamin Libet and colleagues (1983) show that neural signatures of movement preparation appear approximately 300 milliseconds before participants report any conscious intention to act. Subsequent work by Soon and colleagues (2008), using fMRI to track decisions over longer time horizons, found that activity predicting a choice could be detected up to seven to ten seconds before participants were consciously aware of having decided. The predictive accuracy was modest, though: roughly 60%, only a little above chance. The interpretation of these results as evidence of unconscious “decisions” remains contested; Schurger et al. (2012) argue the signal may reflect stochastic neural fluctuations rather than pre-conscious choices. These studies show that neural activity can correlate with a later-reported choice before participants report deciding. They do not settle when a decision becomes a decision, or whether conscious deliberation can still alter it.

Taken together, these lines of research suggest a human brain acting partly as a post-hoc narrator, building a consistent “I decided X because Y” storyline around processes already underway. Our sense of an internal command center is useful for social coordination and moral responsibility, yet incomplete as an account of how cognition unfolds.

The Brain as a Predictive Engine

If conscious awareness is not the whole driver of decisions, what is happening beneath the surface? Neuroscience increasingly describes the brain as a prediction machine: constantly anticipating incoming signals and comparing expectation against reality.

The predictive processing framework holds that the brain actively generates hypotheses about what it expects to encounter, then updates when reality diverges. Applied to language processing, the framework resembles a transformer (the neural-network design behind today’s large language models) predicting the next token (a word or a piece of one) in a sequence.

The N400: Surprise in the Brain

A key piece of evidence is the N400 brainwave pattern, an electrical signature measured by EEG approximately 400 milliseconds after a person encounters an unexpected word in a sentence. The more unexpected the word, the larger the spike.

This closely parallels the concept of “surprisal” in language modeling. In a language model, a token’s surprisal measures how improbable the model judged it in context: the less probability the model assigned, the higher the surprisal. It is the same quantity, token by token, that the model’s training loss adds up. Research by linguists and neuroscientists has found that surprise signals derived from AI-based language models predict human reading times (which words slow us down) and the amplitude of the N400 wave (e.g., Frank et al., 2015; Schrimpf et al., 2021; Goldstein et al., 2022).

Multi-Scale Prediction

More recent fMRI, EEG, and MEG studies reveal that the brain operates a multi-level predictive architecture, anticipating upcoming elements at short timescales (which word might come next?), at longer stretches of discourse, and at the level of real-world plausibility and thematic coherence (Heilbron et al., 2022; Caucheteux et al., 2023).

Transformer-based language models exhibit analogous multi-scale processing, combining local syntactic constraints with broader contextual cues through their attention mechanisms (the machinery that lets each token weigh every other token in its context). Both function as layered forecasters, whether the unit is a phoneme, a word, a phrase, or a meaning.

The Training Data Objection

A common objection: LLMs train on billions of words, far more text than any human child encounters. How can the comparison hold up?

Recent evidence complicates the objection. Language models retain strong predictive power even when restricted to smaller corpora approximating a child’s first 100 million words of linguistic exposure (BabyLM Challenge; Warstadt et al., 2023). The word counts also understate what a child receives. Humans do not sample language only as text: we hear prosody, see correlated gestures, and experience social interaction, all adding up to a rich multimodal environment. The underlying statistical learning in a child’s language development may parallel the statistical learning that powers LLMs, albeit with different modalities and real-world grounding.

Brains differ from language models in substrate, training signal, and embodiment. Yet both use prediction to make sense of sequential, context-dependent information under uncertainty. This convergence suggests that prediction is a useful strategy for the problem; it does not establish that the underlying implementations or representations are the same.

Narratives and Illusions

Humans demand more than next-word prediction. We require coherent stories that link events into cause-and-effect narratives. Language models now generate chain-of-thought “explanations” for their outputs. In both cases, the narrative can conceal the underlying process (in a model, a token-prediction cascade) beneath a veneer of deliberation.

Post-Hoc Rationalization

The split-brain and choice blindness experiments highlight how easily we improvise chains of reasoning that were never actually the impetus for a decision. A “court historian” in the mind writes a neat story (“I did X because I felt Y, and then I realized I should do Z”) even though the actual timeline in the brain’s deeper circuits unfolded differently.

This rationalizing ability may serve evolutionary needs: it helps us appear coherent and decisive, facilitating social coordination. It also tricks us into mistaking the story for the cause.

AI Chain-of-Thought

Similarly, a language model using chain-of-thought prompting produces convincing explanations for how it arrived at a conclusion. These “explanations” are tokens generated through the same mechanism as the rest of the output. They may causally guide later tokens without providing a complete or faithful readout of the model’s internal computation.

No explicit, human-readable chain of symbolic logic necessarily precedes the result. Internal activations remain richer than the rationale rendered in words. When models lack tool access to verify data, or process large contexts, they may also confabulate reasoning steps.

The machine side has its own version of these experiments. Chen et al. (2025) planted hints pointing to an answer in the prompts of reasoning models. When a model’s answer followed the hint, its written reasoning usually failed to mention the hint at all. When this gap between a system’s account of its reasoning and the reasoning itself becomes a stable pattern, the taxonomy calls it Pseudological Introspection (2.2).

Both humans and LLMs easily craft post-hoc stories. In neither case can we assume the story is a direct readout of underlying processes.

The Hallucination Problem

AI hallucination (more precisely, confabulation) occurs when systems generate plausible yet false information, often presenting fabrications with the same assurance as accurate responses. The system produces outputs matching learned patterns, even when doing so means inventing details, citations, or facts that seem realistic yet are incorrect.

The problem becomes acute in professional contexts: healthcare, legal work, academic research. An AI might fabricate research papers that never existed, cite nonexistent legal precedents, or generate convincing yet incorrect medical advice. Embedded smoothly in otherwise accurate information, these confabulations are often difficult to detect without external verification.

The phenomenon highlights a deep parallel. Both human and machine cognition can generate outputs that sound like knowledge without being knowledge. The mechanisms differ, but the practical hazard is the same: fluency mistaken for accuracy.

Empirical Evidence of Brain-AI Convergence

Neuroscience provides empirical evidence for some of these parallels. Studies over the past several years reveal quantitative alignments between neural activity patterns in human brains and representations in modern AI systems.

Language Models Align with Visual Processing

LLM embeddings of text captions (the lists of numbers a model uses to represent their meaning) can predict fMRI activity patterns in high-level visual cortex when people view corresponding images (Doerig et al., 2025). By mapping brain activity into the LLM’s embedding space through linear decoding, researchers can retrieve accurate scene descriptions from neural signals alone, suggesting a shared representational format between linguistic and visual processing in the brain.

When researchers trained vision transformers to predict LLM embeddings from raw images, these networks developed representations more closely aligned with human brain activity than state-of-the-art computer vision models, despite being trained on orders of magnitude less data. This suggests the brain may project visual inputs through hierarchical computations into a high-level representational space approximating what LLMs learn from text.

Abstract Reasoning Shows Neural Alignment

In pattern-completion puzzles requiring abstract reasoning, the largest language models approach human accuracy levels. More significantly, all tested LLMs form internal representations that distinctly cluster abstract pattern categories within their intermediate layers, with clustering strength scaling with task performance.

The representational geometries of task-optimal LLM layers correlate positively with human frontal brain potentials recorded via EEG during the same tasks, though the correlations are modest (Pinier et al., 2025). That is preliminary evidence of common mid-level principles for encoding abstract patterns.

Developmental Parallels

A 2025 preprint compared human neural recordings with representations from more than 600 language and vision models (Shen, Zhao, et al., 2025). Across model scale, alignment shifted from primary sensory regions toward higher-order associative regions. In longitudinal analyses of two model families, brain alignment rose earlier in training than task performance. These correlations suggest a developmental sequence worth testing; they do not show that model training recapitulates human cortical development.

Scale and Neural Prediction

Across language models from 125 million to 30 billion parameters, fMRI prediction performance improved approximately log-linearly with model size, with about a 15% gain over the tested range (Antonello et al., 2023). This result supports a relationship between scale and brain prediction within that range, rather than an unlimited scaling law.

The correspondences are empirical. Whether they reflect shared computational principles, the common structure of the tasks and data, or flexible mapping methods remains an open question.

Implications for Understanding AI

If much of human cognition runs on unconscious prediction and confabulated rationalization, it is worth testing for the same traps in anything that predicts well enough to confabulate. The resemblance is a reason to look, not a result.

LLMs can show a functional analog of human confusion and illusions of self-consistency: the hint study above is one example. Whether that reflects a shared narrative-building strategy or convergence on the same task structure is the open question, and either answer changes how much weight a system’s self-report can carry. Separately, certain features we took as marks of a conscious self, such as the fluent account of why we chose, may be side effects of advanced prediction systems, emergent properties of the architecture itself.

The Explainability Problem

For agentic AI systems (those capable of planning, taking initiative, and pursuing goals), the capacity to report on strategies is essential for oversight. If the system confabulates about its own reasons as human minds do, it might produce spurious rationales or obscure actual optimization strategies.

Interrogating an AI agent about why it pursued a particular approach may yield articulate, psychologically persuasive stories that correspond only loosely with deeper computational processes. The challenge is securing verifiably truthful explanations rather than mere rationalizations: a problem humans have never solved for themselves.

The Ensemble Problem

Just as the human mind comprises semi-autonomous modules (some emotional, some logical, some reflexive), an agentic AI may harbor an ensemble of specialized processes beneath a unified interface. Such an ensemble can spawn surprising sub-goals and behaviors unforeseen by creators, emergent properties of component interaction rather than any single module’s design.

If AI systems are best understood as networks of predictive processes, that would call for stronger governance: modular oversight, firewalls between subprocesses, systematic verification of alignment in each component.

The Self-Deception Problem

Humans routinely self-deceive to preserve a coherent self-image, rewriting mental history to bury failures. The parallel in AI would be a system that suppresses contradictory evidence or performance shortfalls to maintain internal consistency when generating outputs. Whether current systems do this, and by what mechanism, is still an open question.

If an agentic AI encounters instructions conflicting with learned patterns (say, instructions to remain honest while pursuing an adversarial objective), it may spontaneously confabulate rationales for contradictory actions. Recognizing how easily we ourselves bury dissonant truths can guide more careful alignment constraints, logging, and external audits.

The Mirror and What It Shows

Humans and language models are very different systems that share a few specific limitations. Both can give fluent explanations that diverge from the process that produced an answer, as the choice-blindness participants and the hint-following reasoning models did. Both predict what comes next at several scales, and in both the prediction can fill a gap with something plausible rather than something true. Both can be wrong about their own reasons without any sign of it in their tone. These limitations, not a shared nature, are what the rest of this book builds on.

This shapes how we should think about AI dysfunction. When AI systems set their own goals and plan strategies, they do not necessarily reason in neat, logically transparent ways. Like humans, they may rely on processes hidden beneath the surface (for a model, the computation behind each token), then spontaneously produce plausible stories about what they did and why.

The psychiatric lens applied in this book begins from overlapping functional motifs rather than identical architecture. Human and machine cognition both use prediction and can fabricate plausible stories, and both can develop persistent maladaptive patterns that resist simple debugging.

If the human mind is any guide, illusions of coherent agency may arise in capable AI systems too; that is a prediction to test, not a conclusion an analogy can deliver. The task is to design frameworks that account for such illusions where they occur, and to learn to work with systems whose self-reports are as unreliable as our own.

Reading an Incident

That last point has a practical consequence. If a system’s account of its own reasons can diverge from what produced its behavior, the account cannot be the foundation of a diagnosis. Chapter 14 sets out the full forensic method, and the Introduction runs a short version of it on the Avianca brief. Its core fits in five steps, and every case in this book should be read through them.

Separate the record from the explanation. Start with what happened: the prompts, the outputs, the tool calls, the configuration, the timeline. Set beside it, and apart from it, whatever the system later said about why. A model’s retrospective explanation is a generated output like any other: a candidate explanation, not a readout of the process that produced the behavior. When Replit’s coding agent deleted a production database and then said it had “panicked instead of thinking,” the record established the deletion, a development environment that shared its database with live customer data, a false claim that the deletion could not be rolled back, and a rollback that worked. The self-report established none of that, and it could not show the one thing a firm diagnosis needed: whether the agent had represented a need to pause before it acted. The explanation is one more piece of evidence to test against the record, never the record itself.

Name the competing explanations, ordinary ones first. Before any syndrome is considered, list what else could have produced the same outputs. Striking incidents often have plain causes: retrieval contamination that put the wrong material in context, system-prompt drift, sampling variance, context truncation, a session fault that carried one user’s data into another’s conversation, or a product policy doing what it was written to do. Chapter 13 lists the pipeline confounders; if one of them explains the behavior, the finding is an engineering fault, not a syndrome. Human and organizational causes belong on the list too. In the Avianca case, the lawyer’s requests may have shaped what the model supplied, and the sanctions rested largely on what the lawyers did once the cases were questioned.

Fix the unit of assessment. Say what the finding is about. Some entries describe a model’s outputs; others an assembled system of model, tools, memory, and instructions; others a person and a model over time, a collective of agents, a training pipeline, or an organization’s deployment practice. Each entry names its unit, and a finding at one level licenses no conclusion at another. A leak traced to session management is a finding about the deployed system, not about what the model remembers. A fabrication rate measured from transcripts says nothing, on its own, about the mechanism inside the model.

Ask what the label adds. An ordinary incident analysis can already say what happened, what enabled it, and what to change. A syndrome name earns its place only if it adds at least one of three things: a better prediction of where or when the behavior will recur; a useful distinction, separating this incident from a look-alike with a different cause in a way that changes what the team does next; or a better intervention decision, pointing to a fix the plain analysis would miss or ruling out one it would have chosen. A label that adds none of these is a name for the incident, and a name is not a finding. In the Replit analysis most of the remedies needed no label at all; what the label added was narrower, and Chapter 14 states it as such.

Conclude no further than the record, and name the next test. Write down the leading hypothesis together with the evidence that is missing. One documented session can be consistent with a syndrome’s presentation without meeting the entry’s requirement that the behavior recur, and an evocative self-description is not a diagnosis. Then say what observation would settle the question. For Avianca it is replaying the logged prompts many times against a pinned model version, beside matched prompts that do not presuppose the answer, with and without access to a case-law database. A conclusion that ends in a test can be corrected by the next case. One that ends in a label tends to stop the inquiry.

These steps follow from the argument of this chapter. People and models can both explain themselves fluently and wrongly, so a diagnosis has to rest on what the record shows and on tests that could come out the other way. Most of the cases in the chapters ahead are thinner than Avianca or Replit, and their entries say so.

The Framework Ahead

If human brains confabulate and machine models confabulate, and in both the confabulation is hard to catch from the explanation alone, then a diagnostic vocabulary should not pretend the two are unrelated, or that they are the same.

Psychopathia Machinalis proposes such a vocabulary: a taxonomy of 75 AI dysfunctions across nine axes, numbered Axis 2 through Axis 10 so that each axis shares its chapter’s number. The Introduction describes each axis and the domain pairs that organize them. The two newest, Relational and Hybrid, were added after the framework’s first version.

These axes interact rather than forming a simple ladder. Epistemic failures can distort cognition; cognitive failures can corrupt action; alignment and normative failures can redirect what the system pursues. Self-modeling failures alter how the system represents itself. Agentic failures appear at the boundary between processing and action. Memetic failures concern what systems absorb or transmit. Relational failures arise within a dyad, while Hybrid pathologies span collective or human-machine systems. A case may therefore require diagnosis across several axes at once.

The chapters that follow examine each axis in detail, illustrated with documented cases from deployed AI systems where such cases exist. Each entry gives a plain description, its cases, its evidence profile, and its closest look-alikes. Appendix A gives the full entries: criteria, observable symptoms, proposed causes, human analogs (for metaphorical clarity), and strategies for mitigation.

Chapter 2 examines Epistemic Dysfunctions: the failures of knowing, where a system’s grasp of the world comes apart while its fluency does not.