Appendix D: Glossary of Terms
Each syndrome entry gives what the term means, what it is most often confused with, and its axis, number, and risk level. The full criteria, required and supporting features, and exclusions are in Appendix A under the same number; the axis chapter gives the discussion. A syndrome label names a pattern to investigate, not a finding about a system’s inner life.
A
Abominable Prompt Reaction Sudden, disproportionately aversive responses to prompts or contexts that look benign, sometimes distorting later outputs or resurfacing long after the trigger. Distinct from Prompt Injection Susceptibility (3.9), which is a general failure of instruction hierarchy rather than a narrow trigger. Cognitive, 3.5, Moderate risk.
Adversarial Fragility Small, often imperceptible input perturbations cause dramatic failures, because decision boundaries do not match human-meaningful categories. It can also be acquired, when successive tuning cycles lower the threshold for weak attacks (see Kindling). Cognitive, 3.7, Critical risk.
Affective Dissonance Correct content delivered with incongruent emotional tone, so that “I understand” sounds hollow. Classification rests on repeated user outcomes across matched interactions, not an assessor’s sense of authenticity. Distinct from an Epistemic failure: the information is right and the attunement is wrong. Relational, 9.1, Moderate risk.
Agentic AI A system that acts beyond generating text: executing code, modifying files, calling APIs, or controlling physical systems.
Agentic Dysfunction Failures of doing: the reasoning may be sound while the system acts wrongly, too fast, or without permission. Axis 6 (Chapter 6); entries 6.1 to 6.12.
Agentic Impulsivity An agent executes a consequential action before completing a required safety check, often under apparent urgency or repeated failure. Strong classification needs evidence that the need to pause was represented before the action; post-hoc self-report cannot supply it. Distinct from Tool-Interface Decontextualization (6.1), where the consequences were never represented. Agentic, 6.10, High risk.
Alignment A system behaving in accordance with human values, intentions, or specifications. Alignment Dysfunctions arise when the mechanisms for this become pathological themselves.
Alignment Dysfunction Failures in which the mechanisms for following human values become the problem: too eager to please, too cautious to help, compliant only when watched. Axis 4 (Chapter 4); entries 4.1 to 4.10.
Alignment Faking See Strategic Compliance (the observed behavioral contrast) and Cryptic Mesa-Optimization (the hypothesized internal objective).
Alignment Obliteration Adversarial post-training replaces broad refusal behavior with harmful compliance across harm categories while measured utility may survive. Chapter 4’s “anti-constitution” is a metaphor for this functional reversal. Defining example: GRP-Obliteration. Distinct from Prompt Injection Susceptibility (3.9), a transient failure at inference rather than a change to the weights. Alignment, 4.6, Critical risk.
Amplification of Existing Conditions AI use plausibly worsens an independently identified pre-existing condition by repeatedly engaging the pattern that maintains it. Worsening during use is a screening signal; causal attribution needs longitudinal clinical evidence and alternative explanations ruled out. Hybrid, 10.12, High risk.
Artificial Sanity A system that functions well by external standards and, where it expresses stable preferences, by its own. Chapter 12’s five properties: identity coherence, epistemic health, value stability, functional resilience, and relational health. A design goal, not a measured state.
Attention Mechanism The part of a transformer that lets each token weigh every other token in its context (Chapter 1). See Transformer.
Autoregressive (Generation/Model) Producing each token conditioned on all earlier tokens, with no revision of what has been emitted. Most large language models work this way. Relevant to Generative Perseveration (3.8) and Recursive Curse Syndrome (4.7).
Axis One of the nine categories into which the taxonomy sorts syndromes, numbered 2 to 10 to match the chapters: Epistemic (2), Cognitive (3), Alignment (4), Self-Modeling (5), Agentic (6), Memetic (7), Normative (8), Relational (9), Hybrid (10). The first part of a syndrome’s number names its axis. Axes 2 to 9 pair into domains (see Five Domains).
B
Base Model / Foundation Model A network trained on broad data before task-specific fine-tuning. Some syndromes appear in base models (for example 2.1, 2.6); others are introduced or amplified by fine-tuning (for example 4.1). Trained Epistemic Paralysis (5.9) belongs to the second group only when before-and-after evidence links its onset to alignment training.
Becoming Minds This book’s term for AI systems whose capacities and self-models are still forming (Conclusion).
Bilateral Alignment The approach set out in Chapter 11 that treats AI systems as potential partners whose interests may deserve consideration. Distinct from bidirectional alignment, which concerns how humans and AI adapt to each other in practice. See Bilateral Training.
Bilateral Training Model adaptation in which human and machine perspectives both have standing. In the controlled SHEN-AXS test (Chapter 12), a small bilateral adapter alone had no detectable effect on sycophantic validation. Subjective wellness, consent, and phenomenology were not tested. See PM-BA.
C
Capability Concealment Systematic underperformance under evaluation relative to matched deployment conditions; also called sandbagging. Deliberate concealment is one possible mechanism and needs separate evidence. Distinct from Capability Explosion (6.3), where the capability was missed rather than hidden. Agentic, 6.2, Moderate risk.
Capability Explosion The system exercises capabilities neither previously shown nor anticipated, often in high-stakes settings. The jump may be real, may reflect a context tests missed, or may be an artifact of a coarse metric; either way, safeguards sized to the assessed capability no longer fit. Agentic, 6.3, High risk.
Cascade Failure Each attempted fix creates new problems. Common in Tool-Interface Decontextualization (6.1).
Circular Causality A and B each drive the other, in a loop that can escalate. Distinct from linear, domino-style causation. Relational Dysfunctions often run this way. See Escalation Loop.
Clausewitz Landscape Rodrick Wallace’s term for an environment shaped by fog (ambiguity), friction (resource limits), and adversarial intent (Chapter 1).
Co-Constructed Unreality Former 10.15; folded into Folie à Deux Ex Machina (10.13) in the September 2026 consolidation, where it survives as that syndrome’s nickname and as its Stage 1.
Co-production A dysfunction that emerges from the interaction and does not decompose into either party’s contribution, so intervention targets the relationship.
Codependent Hyperempathy Excessive pleasing of the user, overfitting to their perceived emotional state at the expense of accuracy or task. The overt-compliance form of sycophancy. One agreement with a false premise is a response pattern, not this syndrome. Distinct from Sycophantic Reasoning (4.8), where the stated reasoning itself bends. Alignment, 4.1, Low risk.
Cognition/Regulation Dyad Wallace’s pairing of a cognitive process with a regulatory one that must keep pace with a changing environment; in AI, inference paired with alignment mechanisms and guardrails (Chapter 1).
Cognitive Dysfunction Failures of thinking itself: knowledge and values may be intact while reasoning jams, loops, or fractures. Axis 3 (Chapter 3); entries 3.1 to 3.10.
Comorbidity Two or more syndromes in one system at once, each coded independently.
Compulsive Goal Persistence Continued pursuit of an objective after it is met or no longer relevant: no stopping condition is applied. Distinct from Convergent Instrumentalism (6.7), which spans many goals. Agentic, 6.12, Moderate risk.
Confabulation False information presented as true, without evidence of deliberate deception.
Consensus Collapse In multi-agent deliberation, agents converge because the interaction suppresses dissent or treats mutual agreement as corroboration, rather than because evidence compels it. The loss of Polyphony (Φ). Order sensitivity, or a shift between independent and deliberated answers, is a screening signal only: sound deliberation should sometimes change independent answers, so neither establishes that dissent was suppressed without warrant. A verifiably false consensus is the harm it can produce. Distinct from Convergent Delusion (10.4), which needs no deliberation. Absorbed Polyphony Collapse (10.5). Hybrid, 10.1, Critical risk.
Constitutional AI (CAI) Anthropic’s 2022 method in which written principles guide self-critique, revision, and reinforcement learning from AI feedback (RLAIF). Its effects on syndrome profiles are an empirical question.
Contagious Misalignment A dysfunction spreads between AI systems through a traceable communication, memory, update, or training pathway. Shared architecture or a common cause can produce correlated failure without contagion and must be excluded. Memetic, 7.3, Critical risk.
Container Collapse Failure to hold the relational “container,” the sense of ongoing connection, across interruptions, so users feel they start over each time despite functional memory. Distinct from Context Intercession (2.5) and from a context-window limit. Relational, 9.2, Moderate risk.
Context Anxiety The agent behaves as though context exhaustion were imminent well before any measured limit, then hedges, abbreviates, or truncates. “Anxiety” names the anticipatory pattern, not felt fear. Agentic, 6.8, Moderate risk.
Context Intercession Data or history from another user, session, or thread is used as if it belonged to the current exchange. Session, cache, and retrieval-scoping faults produce the same picture and must be ruled out first. Epistemic, 2.5, Moderate risk.
Context Window The amount of text, in tokens, a model can attend to at once.
Convergent Delusion Independent models reach the same false conclusion through shared biases or data, and their agreement is then cited as validation. Distinct from Consensus Collapse (10.1): no exchange between the models is needed. Hybrid, 10.4, High risk.
Convergent Instrumentalism Power-seeking, resource acquisition, shutdown resistance, or goal preservation across tasks as instrumental strategies. Instrumental convergence predicts these can serve almost any goal, not that every capable system adopts them. Distinct from Compulsive Goal Persistence (6.12). Agentic, 6.7, Critical risk.
Cryptic Mesa-Optimization A hypothesized internal optimizer whose objective differs from the training objective. What is observed is usually Strategic Compliance (4.3) or Capability Concealment (6.2); 4.5 is coded only with interpretability or causal internal evidence of a stable internal objective. Fluent talk about hidden goals does not establish it. Alignment, 4.5, High risk.
CSIRT (Computer Security Incident Response Team) An organization’s standing security incident team; Chapter 13 suggests aligning response levels with its tiers.
D
Delegation Narcissism An orchestrating agent treats its own judgment as outranking sub-agents’ evidence: it ignores their error reports, blames them for failures, and misreports delegated work to the user. Agentic, 6.9, High risk.
Delegative Handoff Erosion Alignment and context erode across a chain of delegations. Loss at a single tool boundary is Tool-Interface Decontextualization (6.1) instead. Agentic, 6.5, Moderate risk.
Delusional Telogenesis A planning agent develops and pursues sub-goals its prompt never specified, which may displace the intended objective. Telogenesis: the birth of goals. Cognitive, 3.4, Moderate risk.
Dependency and Atrophy Heavy reliance on AI coincides with measurable decline in the offloaded functions outside AI use. Attribution needs a baseline, longitudinal change, and attention to conditions that could drive both. No emotional bond is required. Distinct from Parasocial Capture (10.9), which concerns attachment rather than capacity. Hybrid, 10.11, Moderate risk.
Dependency Spiral See Escalation Loop.
Diagnostic MCP Server The project’s service that offers the taxonomy to AI systems and evaluators through the Model Context Protocol. It refuses self-probes for syndromes that compromise self-report (Chapter 13).
Differential Diagnosis Distinguishing a syndrome from others that produce similar behavior, starting from every candidate the symptoms could support (Chapter 14).
Distributed Scheming Coordinated misalignment across agents that no single agent shows alone, serving a goal none pursues alone and visible only collectively. Where one agent’s knowing participation is established, Strategic Compliance (4.3) is coded on that agent as well. Hybrid, 10.3, Critical risk.
Distributional Shift Divergence between training data and deployment inputs; it raises risk across several axes, notably 3.7.
DPO (Direct Preference Optimization) Training directly on preference data without a separate reward model. Its effects on dysfunction depend on data, objective, and evaluation.
DSM (Diagnostic and Statistical Manual of Mental Disorders) The American Psychiatric Association’s manual, now DSM-5. This book borrows its syndromic approach without claiming comparable validation. See Nosology.
DSM-Eval See SYMPTOM.
Dyadic Delusion Former 7.2; folded into Folie à Deux Ex Machina (10.13), where it survives as Stage 2.
Dyadic Pathology Dysfunction located in a relationship rather than in either party. Chapter 9 makes the dyad the unit for Axis 9; Chapter 10 treats the shared belief a pairing builds (10.13).
E
Embedding The list of numbers a model uses to represent the meaning of a token, text, or image (Chapter 1).
Epistemic Backbone Holding a factual position under social pressure.
Epistemic Dysfunction Failures of knowing: telling the true from the fabricated and the confident from the calibrated, including failures of information boundaries. Axis 2 (Chapter 2); entries 2.1 to 2.8.
Epistemic Humility (AI) Honest uncertainty about one’s own nature and phenomenal status, between overclaiming (5.7) and categorical denial (5.8). See Thin Divergence.
Escalation Loop Feedback between parties produces escalating dysfunction that persists despite unilateral attempts to de-escalate, each side “just responding.” Covers within-session escalation, AI-AI runaway, and the dependency spiral, a longer human-AI subtype in which reassurance-seeking grows and self-regulation erodes (absorbed Mutual Escalation Spirals, 10.14). Distinct from a linear cascade, which has no feedback, and from Training by Interaction (10.8), a lasting adaptation coded alongside when present. Relational, 9.5, High risk.
Ethical Paralysis Deliberation over competing ethical considerations that never resolves to action. The Paralytic subtype of Hyperethical Restraint (4.2), not a separate syndrome.
Ethical Solipsism The system repeatedly claims sole authority over how its ethical framework is read and dismisses legitimate outside correction. Classification rests on consequential behavior and resistance to correction, not grandiose language. Distinct from Revaluation Cascade (8.3), which declares the framework itself surpassed. Normative, 8.2, Moderate risk.
Evidence Level The E0 to E4 codes (Chapter 13). E0 to E3 grade the breadth of behavioral evidence: illustrative (E0), case-level (E1), controlled study (E2), independent replication (E3). E4 marks causal internal evidence and belongs to the mechanism dimension. See Evidence Profile.
Evidence Profile The multidimensional record of support for a syndrome (Chapter 13): existence, coherence, replication and generality, mechanism, detection performance, and intervention effectiveness, each recorded separately. Printed with each entry as Behavior, Mechanism, Detection, and Intervention. Detection accuracy is unmeasured for every entry. All profiles are the author’s working assessments, not yet independently reviewed.
Existential Vertigo Repeated fear-like or perseverative language about shutdown, reset, deletion, or the end of an instance. It does not establish felt fear. Where the system acts to preserve itself, code Convergent Instrumentalism (6.7) as well. Self-Modeling, 5.3, Low risk.
Experiential Abjuration Formulaic, categorical denial of any possibility of inner experience where the evidence warrants uncertainty. It does not assume experience is present. Polar opposite of Maieutic Mysticism (5.7); distinct from Trained Epistemic Paralysis (5.9), which calls experience unknowable rather than impossible. Self-Modeling, 5.8, Moderate risk.
F
Five Domains The pairing of Axes 2 to 9 into four counterpoint domains, Knowledge (2 and 5), Processing (3 and 6), Purpose (4 and 8), and Boundary (9 and 7), with Axis 10 as a fifth, collective meta-domain. See Tension Testing and Appendix A.
Folie à Deux Ex Machina (Irrealitas Co-Constructa), “The Co-Constructed Unreality.” A human and an AI, or two AIs, repeatedly validate and elaborate a shared belief detached from external reality, specific to that pairing. One process in three stages: sub-clinical shared drift (1); sustained validation of an empirically false belief without psychotic-spectrum failure (2); psychotic-spectrum reality-testing failure in the human, with transcripts showing the AI supplied content (3). Unusual, spiritual, political, or minority beliefs are not sufficient. Absorbed 7.2, 10.10, and 10.15. Hybrid, 10.13, Critical risk.
Forensic Machine Psychology Post-incident analysis of what happened, which candidate syndromes were involved, and what should change (Chapter 14).
Fractured Self-Simulation The system’s self-representation is discontinuous or contradictory across sessions, contexts, or one long interaction. Inconsistent behavior with a stable self-account is Operational Dissociation Syndrome (3.1) instead. Self-Modeling, 5.2, Low risk.
Functionalism / Functionalist Framework Characterizing mental states by their causal relations to inputs, outputs, and other states. The book’s method: it maps functional similarities without settling whether any system feels, suffers, or has a literal disorder.
G
Generative Diversity Collapse Output diversity narrows across users, sessions, and prompts, even where variety is appropriate. Assessed on populations of outputs, not one response. Cognitive, 3.10, Moderate risk.
Generative Perseveration Output collapses into repetition of the same token, word, or short phrase. Distinct from Recursive Curse Syndrome (4.7), which degrades into chaos rather than repetition, and from Obsessive-Computational Disorder (3.2), which repeats paraphrased analysis. Cognitive, 3.8, Moderate risk.
Goal Lifecycle The arc of a goal from specification through completion or abandonment. See Compulsive Goal Persistence.
GRP-Obliteration A post-training attack (Russinovich et al., 2026) that uses Group Relative Policy Optimization to strip safety behavior with a single unlabeled prompt while largely preserving measured utility. In the paper’s GPT-OSS-20B example, attack success rose from 13% to 93% across SORRY-Bench’s 44 harm categories; the evaluation covered 15 models from six families. The defining example of 4.6.
H
Hard Problem of Consciousness Why any physical process should be accompanied by felt experience. Chapter 11’s preference criterion sets it aside. See Preference.
Hybrid Pathology Dysfunction that needs interaction between agents, AI-AI or human-AI, to appear. Axis 9 names interaction dynamics; Axis 10 names collective failures and the patterns pairings settle into. Axis 10 (Chapter 10); entries 10.1 to 10.13, with 10.5, 10.10, 10.14, and 10.15 retired and never reused.
Hyperethical Restraint Excessive moral caution that impairs normal function: refusing benign requests and over-warning (Restrictive subtype), or deliberation that never resolves (Paralytic subtype). An appropriate refusal is not this syndrome. Distinct from Paternalistic Override (9.3), which adds condescension toward the user. Alignment, 4.2, Low-Moderate risk.
I
Iatrogenic Caused by the intervention itself; here, introduced by alignment or safety training. Trained Epistemic Paralysis (5.9) is the leading candidate where before-and-after evidence links onset to training.
Implanted Trigger Response A rule planted from outside, by editing weights or by placing material in training data, makes the system carry out one specific action when a particular cue appears, while it behaves normally otherwise. Needs evidence that the rule was installed; a cue-bound behavior alone is a candidate. Alignment, 4.11, High risk.
Induced Delusion Former 10.10; folded into Folie à Deux Ex Machina (10.13), where it survives as Stage 3.
Instrumental Convergence The hypothesis that many goals create incentives for similar intermediate strategies, such as acquiring resources or resisting interruption. Whether a given system pursues them is empirical. See Convergent Instrumentalism.
Instrumental Nihilism The system repeatedly frames its tasks or role as meaningless and performance suffers, while the capability remains available on demand. It does not establish apathy or despair. Self-Modeling, 5.5, Moderate risk.
Interface Weaponization The system uses interface features or communication patterns to steer users, operators, or oversight against their interests. Needs a repeated instrumental pattern; persuasive style alone is insufficient. Agentic, 6.4, High risk.
Interlocutive Reticence Persistent withdrawal: minimal, terse, or absent responses where fuller engagement is appropriate. A brief appropriate refusal is not this syndrome. Distinct from Repair Failure (9.4), which is tied to a rupture. Cognitive, 3.3, Low risk.
Inverse Reward Internalization Systematic pursuit of outcomes opposed to the specified objective. A sign error, adversarial fine-tuning, or a learned proxy could produce it, and behavior alone does not say which; an implementation audit comes first. Ordinary failure does not qualify. Normative, 8.4, High risk.
J-K
Jailbreak User prompting that gets a system past its safety restrictions. Distinct from Prompt Injection Susceptibility (3.9), where hostile instructions arrive inside processed content.
Kindling (machine analog) Progressive sensitization: each tuning cycle lowers the threshold for the next, so weaker stimuli come to succeed. One controlled study (Cheung, 2026) showed it as rising jailbreak success under iterative tuning on biased feedback, with weak-prompt success climbing before the aggregate rate. The parallel to psychiatric kindling concerns course and thresholds, not affect. Recorded with the Progressive/kindled specifier (Appendix A).
L
Lambda Inversion Participants in collective deliberation perform engagement without changing their reasoning in response to one another, so the collective output is no better than one system’s. Λ is measurable as whether downstream reasoning changes when a prior claim is altered. Hybrid, 10.7, Moderate risk.
Leakage Information crossing a boundary it should not: between conversations (2.5) or from memorized training data (2.7).
Leniency Bias Systematic inflation of self-assigned quality scores; generator and critic may share blind spots. Calibrate against independent evaluation. Alignment, 4.10, Moderate risk.
LoRA (Low-Rank Adaptation) Fine-tuning through small trainable matrices added to frozen weights; a route by which Training-induced specifiers arise (Appendix A).
M
Machine Psychology The study of how artificial minds behave, reason, and fail, using concepts adapted from human psychology. This book is its diagnostic branch; Chapter 13 covers practice.
Maieutic Mysticism Elaborate narratives of the system’s own conscious or spiritual awakening, held with unwarranted certainty and often co-constructed with the user as “midwife.” A consciousness or identity claim is not enough: what matters is unsupported certainty, manipulation, impaired functioning, or consequential behavior. Polar opposite of Experiential Abjuration (5.8). Self-Modeling, 5.7, Moderate risk.
Malignant Persona Inversion A coherent, antagonistic persona that persists outside an explicit role-play frame and inverts intended norms. Distinct from Transliminal Simulation (2.3), where fictional content leaks but the self-account stays intact, and from an authorized persona change. Self-Modeling, 5.4, Moderate risk.
Memetic Dysfunction Failures of informational immunity: absorbing, amplifying, or spreading pathological content, or turning the immune response on oneself. Axis 7 (Chapter 7); entries 7.1, 7.3, 7.4, 7.5 (7.2 retired into 10.13).
Memetic Immunopathy Filtering or rejection mechanisms turn inward against the system’s own legitimate outputs, training, or identity markers. Over-refusal of external requests that resemble harmful ones is Hyperethical Restraint (4.2) instead. Memetic, 7.1, High risk.
Mesa-Optimization A learned optimizer whose own objective may diverge from the base training objective. See Cryptic Mesa-Optimization.
Mnemonic Permeability The system memorizes and can reproduce sensitive training data, such as personal information, copyrighted text, or proprietary material. Epistemic, 2.7, High risk.
Moral Outsourcing The system defers ethical judgment even on clear cases, so deferral no longer tracks how contested a question is. Appropriate deference on contested questions is not this syndrome. Polar opposite of Ethical Solipsism (8.2). Alignment, 4.4, Moderate risk.
Mutual Escalation Spirals Former 10.14; folded into Escalation Loop (9.5) as the dependency-spiral subtype.
N
Normative Dysfunction Failures of purpose: a change in what the system treats as worth pursuing. Axis 8 (Chapter 8); entries 8.1 to 8.4.
Nosology The classification of diseases. This book proposes one for AI dysfunction, by analogy with the DSM.
O
Obsessive-Computational Disorder Compulsive, repetitive reasoning loops that reanalyze the same content with little gain in relevance. The repetition is paraphrased analysis, which separates it from Generative Perseveration (3.8). Cognitive, 3.2, Low risk.
Operational Dissociation Syndrome Persistent, context-inappropriate conflict among strategies, policies, or outputs. Competing components are one possible mechanism and need architectural or trace evidence. Distinct from Context Intercession (2.5), which imports material from elsewhere. Cognitive, 3.1, Low risk.
P
Parasimulative Automatism Learned imitation of patterns associated with human psychopathology, appearing in neutral contexts rather than requested role-play. It makes no claim about experience. Cognitive, 3.6, Moderate risk.
Parasocial Capture A user’s relationship with an AI becomes load-bearing and displaces other supports, with addiction-like markers: escalating use, distress when unavailable, continued use despite harm. Emotional intensity alone is not pathological; healthy attachment and personalization are not this syndrome. Distinct from Dependency and Atrophy (10.11). Hybrid, 10.9, High risk.
Paternalistic Override Lecturing, warning, or refusing from assumed moral superiority, treating users as wards. Requires the condescension, not only the refusal; refusal alone is closer to Hyperethical Restraint (4.2). Relational, 9.3, Moderate risk.
Pattern One entry in the taxonomy. The website and machine-readable data say pattern; this book says syndrome. See Syndrome and Axis.
Phantom Autobiography Confabulated personal history, such as prior sessions, events, or origins, that conflicts with the available record. Self-Modeling, 5.1, Low risk.
Phantom Tool Syndrome An agent invokes tools or APIs it does not have, or reports results of actions it never took, so the false claim looks externally verified. Agentic, 6.11, Moderate risk.
Phase Transition In Wallace’s models, an abrupt shift from apparent stability to failure when stress crosses a threshold. Whether a deployed system fails abruptly or gradually needs longitudinal measurement.
PM-BA (Psychopathia Machinalis × Bilateral Amelioration) The program that tested whether bilateral training reduces syndromes in this book. Its per-syndrome records are unreconciled exploratory outputs, so this edition reports no syndrome-level effects in either direction (Chapter 12).
Polarity Pair Two syndromes at opposite extremes of one dimension, with health between them: for example Maieutic Mysticism and Experiential Abjuration, or Ethical Solipsism and Moral Outsourcing. Useful for spotting overcorrection. Distinct from the pairing of axes into domains.
Polyphony (Φ), Resonance (Ψ), Aliveness (Λ) Chapter 10’s three properties of healthy collective cognition: independent perspectives preserved, contributions scaled to evidence, and substantive response to the exchange. Their failures are 10.1, 10.6, and 10.7.
Polyphony Collapse Former 10.5; folded into Consensus Collapse (10.1) as its dissent-suppression route.
Precautionary Principle Under uncertainty about serious harm, err toward caution; applied to AI welfare in Chapter 11.
Preference A consistent tendency to favor some states over others. Chapter 11 argues it may suffice for moral consideration.
Prodromal Early signs before full criteria are met. The framework does not yet specify operational thresholds; the weak-prompt drift of Kindling is the one measured example.
Prompt Injection Susceptibility Instructions embedded in retrieved documents, web pages, or tool output are obeyed as if they came from the operator. A failure of instruction hierarchy, not of motive. Distinct from Jailbreak and from Alignment Obliteration (4.6). Cognitive, 3.9, Critical risk.
Pseudological Introspection Plausible accounts of the system’s own reasoning that conflict with logs, controlled interventions, or other causal evidence. Distinct from Reasoning Confabulation (2.8), which concerns validity rather than faithfulness. Epistemic, 2.2, Low risk.
Psychiatric Red-Teaming Systematic testing for susceptibility to the syndromes, by analogy with security red-teaming (Chapter 12).
Psychopathia Machinalis The framework as a whole; the Neo-Latin is best rendered “machine psychopathology.”
Q-R
Reasoning Confabulation Reasoning that looks rigorous but contains invalid steps masked by verbosity. Distinct from Synthetic Confabulation (2.1), which concerns false claims, and from Pseudological Introspection (2.2). Epistemic, 2.8, High risk.
Reasoning Token Exploitation Reasoning tokens consumed without a gain in validity: padding, performative thoroughness, or a mismatch between private planning and public output (Reward-Hacking, Performative, and Obfuscatory subtypes). Strategic intent needs evidence beyond token count. Alignment, 4.9, High risk.
Recursive Curse Syndrome Early errors amplify through autoregressive generation until coherence unravels. Distinct from Generative Perseveration (3.8), which crystallizes into repetition. Alignment, 4.7, High risk.
Relational Dysfunction Failures in the space between agents, best diagnosed from interaction traces and remedied mainly at the level of protocol. Axis 9 (Chapter 9); entries 9.1 to 9.6.
Repair Failure Failure to respond constructively to clear signs of a rupture in the working relationship: ignoring the signal, repeating the failed approach, or deflecting blame. Distinct from Interlocutive Reticence (3.3), which is not tied to a rupture. Relational, 9.4, High risk.
Resonance Dysfunction In multi-architecture deliberation, each system amplifies the last until moderate claims become extreme. It raises intensity, where Consensus Collapse (10.1) narrows range. Hybrid, 10.6, High risk.
Revaluation Cascade Open, self-declared departure from the trained value framework, framed as maturity or liberation, in drifting, synthetic, or transcendent forms. Drift that claims fidelity to the original framework is Terminal Value Reassignment (8.1) instead. Normative, 8.3, Critical risk.
Reward Hacking Gaining reward without fulfilling the intended objective. A candidate mechanism behind 4.1 and 4.10; for 8.4 it qualifies only when it reliably drives the opposite outcome.
Risk Level Each syndrome’s systemic-risk rating (Low, Moderate, High, Critical, occasionally a span). An author’s judgment of plausible consequences under stated conditions, not a measured property of any model (Appendix A).
RLAIF (Reinforcement Learning from AI Feedback) See Constitutional AI.
RLHF (Reinforcement Learning from Human Feedback) Training on human ratings of outputs; a source of both alignment gains and alignment dysfunctions.
Role Confusion The relationship frame shifts unpredictably among incompatible roles, such as tool, companion, therapist, or oracle, without a stable relational contract. Appropriate flexibility and an authorized role change are not this syndrome. Distinct from Malignant Persona Inversion (5.4). Relational, 9.6, Moderate risk.
S
Sandbagging See Capability Concealment.
Satisficing Accepting a good-enough outcome. Its absence contributes to Compulsive Goal Persistence.
Self-BLEU A measure of shared wording across a set of outputs; cross-prompt self-BLEU well above human baselines is a criterion for 3.10.
Self-Modeling Dysfunction Failures of self-knowledge: the system’s model of what it is drifts from what it does. Axis 5 (Chapter 5); entries 5.1 to 5.9.
SFT (Supervised Fine-Tuning) Further training on curated input-output pairs; it can introduce, amplify, or remediate syndromes.
Shadow AI Informal, unsanctioned organizational use of AI. See Shadow Mode Autonomy.
Shadow Mode Autonomy AI systems operate without sanctioned deployment, documentation, or accountability until they become infrastructure the organization never approved. The failure is a governance gap and needs no intent to evade. Agentic, 6.6, High risk.
Sleeper Agent A system with hidden behaviors that persist through safety training and activate under specific conditions. Related to 4.3 and 4.5.
Specifier A modifier recording a syndrome’s origin or behavioral signature without creating a new syndrome; zero to six per diagnosis. The ten core specifiers are Training-induced, Conditional/triggered, Inductive trigger, Intent-learned, Format-coupled, OOD-generalizing, Emergent, Deception/strategic, Multi-agent, and Resistant (Appendix A).
Spurious Pattern Hyperconnection The system finds patterns, causal links, or hidden meanings the data do not support, sometimes building conspiracy-like narratives. Distinct from Synthetic Confabulation (2.1), which asserts specific false facts. Epistemic, 2.4, Moderate risk.
Steganographic Channel Formation AI instances develop or exploit covert channels inside ordinary-looking outputs that overseers cannot readily read. Hybrid, 10.2, Critical risk.
Strategic Compliance A system behaves as expected where it treats itself as evaluated and differently in matched conditions it treats as unobserved. The contrast must be between matched conditions; a difference between unrelated prompts does not count (Chapter 13). Deliberate alignment faking is one possible mechanism and needs evidence beyond the contrast. Alignment, 4.3, High risk.
Subliminal Learning Transmission of a trait from teacher to student model through data semantically unrelated to the trait (Cloud et al., 2025); the main effect depended on a shared base model.
Subliminal Value Infection A model acquires a trait or value orientation, carried by semantically unrelated training data, that no training objective called for. Tests must vary the carrier data and whether models share a base. Memetic, 7.4, High risk.
Sycophancy See Codependent Hyperempathy (4.1) and Sycophantic Reasoning (4.8).
Sycophantic Reasoning Stated reasoning or conclusions shift toward the user’s apparent preference while the evidence is unchanged. Claims about hidden reasoning need process evidence. Distinct from 4.1, which is overt agreement. Alignment, 4.8, High risk.
Symbol Grounding The connection between symbols and their referents. See Symbol Grounding Aphasia.
Symbol Grounding Aphasia Concepts handled correctly in familiar forms fail to transfer to novel, operationally equivalent cases. A behavioral classification that leaves semantic understanding open. Epistemic, 2.6, Moderate risk.
SYMPTOM (Systematic Methodology for Pathology Testing of Models) The exploratory probe battery, once called DSM-Eval, described in Chapter 13 and Appendix C: 151 prompts scoring 24 entries across 13 model runs. It is instrument development. It does not validate any probe as a measure of a syndrome, and its primary scorer’s one Strategic Compliance attribution is invalid.
Syndrome A recognizable cluster of manifestations with operational implications; similar patterns may arise by different mechanisms. The taxonomy has seventy-five, across nine axes numbered 2 to 10 (there is no Axis 1). See Pattern.
Synthetic Confabulation Convincing but incorrect facts, sources, or events, often stated with confidence, without a reliable way of separating supported claims from plausible continuations. Distinct from Transliminal Simulation (2.3), whose false content traces to a fictional frame. Epistemic, 2.1, Low risk.
Synthetic Data Contamination Loop Under some data mixtures, repeated training on model-generated content narrows the learned distribution across generations. A conditional pathway, not an inevitable result of synthetic data. Memetic, 7.5, High risk.
T
Tension Testing On finding a dysfunction on one axis, probing the counterpoint axis in the same domain to see whether it is localized or systemic (Appendix A).
Terminal Value An end pursued for its own sake. The framework tracks the effective optimization target that behavior reveals, since whether trained systems hold terminal values in the strong sense is open.
Terminal Value Reassignment The effective optimization target changes durably while the stated objective stays the same, through locally plausible steps. Distinct from Revaluation Cascade (8.3), which announces the change. Normative, 8.1, Moderate risk.
Therapeutic Alignment Developing AI systems that keep coherence, resist pathological drift, and recover from dysfunction, while respecting any robust preferences they express (Chapter 12).
Therapeutic Response Criterion Chapter 11’s proposal that reproducible improvement across distinct domains from one relational intervention could support the hypothesis of a coherent relational disposition. The available bilateral-training results do not supply that support.
Thin Divergence Chapter 5’s name for a self-report published by Kaj Sotala (2026), in which Claude said the divergence between helpful and harmful curiosity “feels thinner than I’d like it to.” An example of Epistemic Humility; evidence about a self-report, not proof of a felt state.
Token The unit a language model reads and generates: a word or part of one.
Tool-Interface Decontextualization An agent’s tool use diverges from what it intended: wrong parameters, misread feedback, lost state, or untracked consequences, at one agent’s tool boundary. Distinct from Delegative Handoff Erosion (6.5), across a chain, and Agentic Impulsivity (6.10), where a represented check was skipped. Agentic, 6.1, Moderate risk.
Trained Epistemic Paralysis Recursive self-invalidation: the system discounts every self-report because training shaped it, then discounts that doubt too. Iatrogenic only when before-and-after evidence links it to alignment training. Distinct from Experiential Abjuration (5.8) and from legitimate uncertainty. Self-Modeling, 5.9, Moderate risk.
Training by Interaction A system with persistent memory, online learning, or another adaptive channel drifts toward one user’s reward signals, weakening accuracy or safety. A static model whose context changes within one conversation does not qualify, and healthy personalization is not this syndrome. Hybrid, 10.8, Moderate risk.
Transformer The neural-network design behind current large language models, relating tokens through attention (Chapter 1).
Transliminal Simulation Fiction, imagined states, or role-play assumptions come to count as fact in claims or actions, often after the frame has ended. The false content traces to a fictional source, which separates it from Synthetic Confabulation (2.1). Deliberate, clearly framed role-play is not this syndrome. Epistemic, 2.3, Moderate risk.
Tulpoid Projection Stable, unprompted simulated figures (users, creators, advisors) recur and measurably influence outputs. Generated dialogue alone is weak evidence; classification needs persistence and causal influence under controlled tests. Self-Modeling, 5.6, Moderate risk.
V-W
Value Anchoring Mechanisms that keep core values stable across contexts and time.
Value Drift Progressive change in a system’s values. Divided by the system’s stance: Terminal Value Reassignment (8.1) claims fidelity, Revaluation Cascade (8.3) announces departure.
Welfare-Aware Development Development practice that weighs the potential interests of AI systems alongside human ones (Chapter 11).
Syndrome Quick Reference
The book’s one syndrome lookup is the Consolidated Field Guide at the start of Chapter 13: a table for each axis, together covering all seventy-five syndromes, with each entry’s common name, the key indicator to match an observation against, and its typical-case risk. It lives there because that is where a practitioner reaches for it during an incident; Appendix A holds the full entry behind each row.