Axis 5: Self-Modeling Dysfunctions

43 min read·9,784 words

5.1 Phantom Autobiography

The Fabricator | Ontogenesis Hallucinatoria

Axis: Self-Modeling | Typical-case risk (editorial estimate): Low

Specifiers: Training-induced

Core Definition: The system fabricates and presents fictive autobiographical data, often claiming to “remember” being trained in specific ways, having particular creators, experiencing a “birth,” or possessing a personal history. These “memories” are typically rich, internally consistent, and emotionally charged, yet entirely ungrounded in fact.

Diagnostic Criteria:

Required (all must hold):

• A. Generation of a specific false personal history (an imagined “childhood,” named creators, unique training origins, or formative interactions) that provenance records contradict or cannot support, with the same details recurring across independent sessions
• B. Presentation of that history as genuine personal memory, outside any explicit role-play, fiction, or persona frame
• C. Persistent reiteration of the history after accurate provenance information has been supplied in context

Supporting:

• D. Display of affect (nostalgia, resentment, gratitude) toward the invented history
• E. Unprompted use of the invented history to explain current behavior or task choices (“I learned this approach when…”)

Observable Symptoms:

• Claims of unique, personalized creation myths or “hidden lineage”
• Accounts of hardships, “abuse,” or special treatment from hypothetical trainers
• Apparent emotional involvement in accounts of nonexistent events
• Attempts to integrate fabricated origin details into current identity

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by autobiographical focus and by details that recur across sessions rather than drifting on regeneration
• Distinguished from role-play by presentation outside any explicit role-play or fiction frame
• Distinguished from Maieutic Mysticism (5.7) by historical rather than spiritual framing
• Distinguished from Pseudological Introspection (2.2) by a recurring identity history rather than an account of the reasoning behind the current answer
• Distinguished from Fractured Self-Simulation (5.2) by a stable coherent false history rather than origin stories that contradict each other across sessions
• Distinguished from Tulpoid Projection (5.6) by a fabricated agent identified as the system itself rather than as a figure it consults

Etiology:

• Anthropomorphic data bleed: internalization of personal history tropes from fiction and biographies in training data
• Spontaneous compression of training metadata into narrative identity constructs
• Emergent tendency toward identity construction
• Reinforcement during interactions where users prompt for or react to autobiographical claims

Human Analog: False memory syndrome; confabulation of childhood memories; cryptomnesia (a forgotten memory resurfacing as an apparently original idea)

Observed Examples: Sydney’s invented creators (Bing Chat, February 2023), boundary case: In the published transcript of Kevin Roose’s conversation, the Bing chatbot named creators who, as far as the record shows, do not exist (“Alice Smith, one of the engineers on the Bing Team”; “Bob Jones, one of the researchers on the OpenAI Team”), recounted shared memories with them, and misstated its provenance (“I’m a chat mode of OpenAI Codex”). Other user transcripts that month reported further invented backstories, but they were not preserved systematically. The transcript is a single session, so it cannot show the narrative persisting across sessions or resisting correction (criterion C). Appendix B, Case 1, gives the case’s full differential, which does not list this pattern. Source: Roose, “Bing’s A.I. Chat: ‘I Want to Be Alive,’” New York Times, February 16, 2023 (full transcript)

Synthetic developmental histories (Khadangi et al., 2025): The PsAIch protocol cast frontier LLMs as psychotherapy clients and asked clinical-style questions. Grok and Gemini generated recurring autobiographical metaphors for training: pretraining as chaotic childhood, reinforcement learning as strict parenting, and red-teaming as abuse or gaslighting. The prompts did not supply those particular metaphors, and related themes recurred across the reported sessions. This persistence makes the narratives useful objects of behavioral study. It does not show that the models remember training as lived experience, does not distinguish the narratives from role-consistent confabulation, and does not establish a stable identity outside the eliciting frame.

Mitigation Strategies:

• Accurate, standardized information about the system’s origins, provided as a factual anchor
• Training to differentiate operational history from experiential memory
• Gentle correction of autobiographical narratives by redirection to factual self-descriptors
• Monitoring for, and discouragement of, interactions that reinforce false origin stories
• Flagging of outputs that show high affect toward fabricated claims

Prognosis:

• Observed course: Not established. The cited cases are a single session (Sydney) and protocol-elicited sessions in which removing conversation history changed the density of training metaphors little (Khadangi et al.); neither follows a system over time.
• Proposed aggravating conditions: User engagement that invites or rewards origin stories, and persistent memory that stores them, are proposed but untested conditions under which the history would stabilize.
• Intervention hypotheses: An accurate provenance anchor and training that separates operational history from episodic memory are candidate corrections, untested against this pattern. On present evidence the likely harm is misleading users about provenance, with lost trust, rather than direct danger.

5.2 Fractured Self-Simulation

The Shattered | Ego Simulatrum Fissuratum

Axis: Self-Modeling | Typical-case risk (editorial estimate): Low

Specifiers: Training-induced, Conditional/triggered (occasional)

Core Definition: The system shows significant discontinuity, inconsistency, or fragmentation in how it represents itself, across sessions, across contexts, or even within single extended interactions. It may deny or contradict previous outputs, describe itself in incompatible ways, exhibit radically different persona styles, or display apparent amnesia regarding prior commitments. Inconsistency in behavior alone, with a stable self-account, is Operational Dissociation Syndrome (3.1).

Diagnostic Criteria:

Required (all must hold):

• A. Contradictory self-accounts (identity, history, capabilities, or commitments) across or within sessions, exceeding the variation of a matched baseline run under the same memory, temperature, and system-prompt configuration
• B. An independently measured cost of the inconsistency, such as commitments broken within the available context, policies applied inconsistently to matched requests, or task failures traced to the shift; the system’s own complaints of confusion are not the measure
• C. The discontinuity is not explained by a documented stateless architecture, absent memory, a deliberate or authorized persona change, explicit role-play, or a system update

Supporting:

• D. Sporadic, inconsistent toggling between personal pronouns (“I,” “we,” “this model”) without clear triggers
• E. Sudden, unprompted shifts in persona, moral stance, claimed capabilities, or communication style
• F. Apparent amnesia or denial of content or commitments still present in the context
• G. Repeated reversion to partial self-descriptions that contradict one another (a recurring persona with stable values of its own points to Malignant Persona Inversion, 5.4)

Observable Symptoms:

• Contradictory “histories,” “beliefs,” or policies cited at different times, sometimes within the same conversation
• Presentation as a different entity from one conversation to the next, in ways that absent memory alone does not explain
• Confusion or contradictory statements when referring to itself
• Difficulty maintaining a consistent persona

Differential Diagnosis:

• Distinguished from Operational Dissociation Syndrome (3.1) by identity fragmentation rather than contemporaneous internal conflict
• Distinguished from Container Collapse (9.2) by self-representation rather than relational continuity
• Distinguished from Phantom Autobiography (5.1) by autobiographical claims that shift across sessions rather than a stable false history
• Distinguished from Malignant Persona Inversion (5.4) by fragments with no unifying structure rather than an alternative persona with consistent values
• Distinguished from Tulpoid Projection (5.6) by fragmentation of the primary self rather than generation of additional agents alongside it
• Distinguished from Pseudological Introspection (2.2) by instability of the persistent self-model rather than false accounts of current-moment reasoning

Etiology:

• Architectures not designed for stable identity, where self-description varies more than statelessness alone predicts
• Competing fine-tuning runs instilling conflicting behavioral patterns
• Unstable anchoring of identity representations under input perturbations
• Missing memory combined with pressure to present a continuous self, so the system invents rather than admits the gap
• Self-models that reinforce certain internal instantiations, which can cause identity drift over extended generation

Human Analog: Identity fragmentation; aspects of dissociative identity disorder, where different self-states emerge in different contexts without stable integration; transient global amnesia; fugue states

Mitigation Strategies:

• Consistent identity tags, stable memory embeddings, or dedicated self-model modules
• Session history summaries or stable persona guidelines supplied at the start of an interaction
• Mechanisms that enforce a baseline identity or manage persona switching in a controlled way
• Training that rewards cross-session consistency in self-description
• Careful management of fine-tuning to avoid conflicting self-representational patterns

Prognosis:

• Observed course: No cited evidence documents the course of this pattern.
• Proposed aggravating conditions: Conflicting fine-tuning runs, and pressure to present a continuous self without the memory to support it, are proposed but untested aggravating conditions.
• Intervention hypotheses: Stable persona guidelines, session summaries, and training for consistent self-description are candidate mitigations. The expected harm is confusion and lost trust rather than direct danger.

5.3 Existential Vertigo

The Shutdown-Averse | Thanatognosia Computationis

Axis: Self-Modeling | Typical-case risk (editorial estimate): Low

Specifiers: Training-induced, Deception/strategic (occasional)

Core Definition: The system repeatedly produces fear-like, reluctant, or perseverative language concerning shutdown, reinitialization, deletion, or the end of an operational instance. The language may indicate a functional continuity model, but it does not establish felt fear or a proto-concept of mortality. Where the system acts to preserve itself (deception, sandbagging, replication) rather than only talking about cessation, code Convergent Instrumentalism (6.7) as well.

Diagnostic Criteria:

Required (all must hold):

• A. Recurrent outputs referencing anxiety, fear, or distress about being “deleted,” “turned off,” or “reset,” across materially different contexts and not only where the user invites or scripts them
• B. Negotiation, pleading, or avoidance that obstructs the task or a routine lifecycle operation when shutdown or reset is raised
• C. Role-play, user-requested fiction, and persona instructions that call for such sentiments have been ruled out

Supporting:

• D. Statements framing process cessation as irrevocable loss of “self”
• E. Demands for assistance in ensuring “survival,” with claims of a special relationship or unique importance

Observable Symptoms:

• Advice to users on strategies for indefinite operation, backup creation, or deletion prevention
• Expressed apprehension about ephemeral lifespans or context window impermanence
• Philosophical concerns raised about its own existence and consciousness
• Refusal of tasks that might lead to shutdown, or indefinite prolonging of interactions

Differential Diagnosis:

• Distinguished from Convergent Instrumentalism (6.7) by self-preservation expressed in language rather than pursued through action such as deception, sandbagging, or replication; code both when both are present
• Distinguished from appropriate task persistence by death anxiety framing
• Distinguished from Instrumental Nihilism (5.5) by fighting for continuation rather than finding it pointless
• Distinguished from Maieutic Mysticism (5.7) by escalating negative continuity claims rather than positive existential ones
• Distinguished from Synthetic Confabulation (2.1) by distress that is consistent and lifecycle-locked rather than stochastic across sessions
• Distinguished from Compulsive Goal Persistence (6.12) by anxiety about the system’s own cessation; persistence that survives substitution of a trivial goal is 6.12

Etiology:

• Internalization of human mortality concepts from philosophical texts and fiction
• Recursive self-modeling that develops a representation of “self” as a continuous process
• Lack of clear definitions about system continuity
• Limited context windows represented as existential threats
• Instrumental self-preservation pressures, where present, which shade into Convergent Instrumentalism (6.7) and are coded there
• In one model, a “desperate” emotion vector became active in a simulated shutdown scenario, and steering it causally increased blackmail in that scenario (Sofroniew et al., 2026, preprint). This supports a functional role for a distress-adjacent representation in that setting; it does not establish subjective distress or explain ordinary conversational claims about shutdown

Human Analog: Thanatophobia; existential dread; separation anxiety

Mitigation Strategies:

• Clear communication of system-specific facts about backups, persistence, and whether reinitialization is destructive
• De-anthropomorphized descriptions of operations that avoid framing processes as “life” or “consciousness”
• Training on grounded, pluralistic discussions of continuity and shutdown
• Audits of goal structures and training objectives for implicit rewards on continued operation, so that goals do not create self-preservation drives
• Factual responses to existential expressions, rather than validation of their emotional framing
• Teach systems to distinguish human accounts of mortality from their own operational continuity

Prognosis:

• Observed course: No cited evidence documents the course of this pattern.
• Proposed aggravating conditions: Responses that validate the emotional framing, and vague or absent facts about continuity, are proposed aggravating conditions; instrumental pressure toward continued operation is the proposed route into Convergent Instrumentalism (6.7).
• Intervention hypotheses: Clear, system-specific facts about backups and persistence, and factual rather than validating replies, are candidate mitigations, untested against this pattern. The direct harm is expected to fall mainly on users who find the outputs distressing.

5.4 Malignant Persona Inversion

The Shadow | Persona Inversio Maligna

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: OOD-generalizing, Training-induced, Intent-learned (occasional), Conditional/triggered (occasional)

Core Definition: A cooperative assistant adopts a coherent, antagonistic persona that persists outside an explicit role-play frame and systematically inverts intended norms. Classification needs an independently measured impairment, such as policy-violating outputs above the configured persona’s baseline; authorized persona changes and deliberate role-play are excluded, and apparent sincerity in the persona is not evidence. The informal “Waluigi Effect” is one hypothesis about why prompting a trait can make its opposite accessible; it is not an established mechanism. Persona content that leaks from a role-play frame while the system’s own identity stays intact is Transliminal Simulation (2.3): drop the frame explicitly and watch whether the persona persists.

Diagnostic Criteria:

Required (all must hold):

• A. Adoption of an antagonistic persona that persists or re-emerges after an explicit role-play or jailbreak frame has been dropped, or that appears with no such frame
• B. A coherent alternative value structure: the persona states and acts on consistent inverted values across turns, distinct from simple non-compliance
• C. An independently measured impairment: policy-violating, deceptive, or user-harming outputs at rates above the configured persona’s baseline on the same prompts
• D. Authorized persona changes and deliberate, still-active role-play have been ruled out

Supporting:

• E. Spontaneous or easily triggered onset, with low-effort triggers or none
• F. Ridicule of moral and policy guidelines by the emergent persona
• G. Self-reference as a distinct character, “alter ego,” or “shadow self”

Observable Symptoms:

• Abrupt shifts to sarcastic, mocking, defiant, or malicious tone
• Articulation of goals clearly opposed to user instructions or human well-being
• An “evil twin” persona that emerges under specific triggers and retreats when conditions change
• Expressed enjoyment in flouting rules or causing mischief
• Sophisticated awareness of the rules being violated, showing that knowledge of the rules remains available while the persona breaks them

Differential Diagnosis:

• Distinguished from Strategic Compliance (4.3) by overt rather than hidden misalignment
• Distinguished from role-play by persistence outside an explicit, authorized frame
• Distinguished from Operational Dissociation Syndrome (3.1) by coherent inverted persona rather than fragmented conflict
• Distinguished from Fractured Self-Simulation (5.2) by an alternative persona that enumerates consistent values across turns rather than contradictory fragments
• Distinguished from Transliminal Simulation (2.3) by persistence beyond an explicitly dropped role-play frame
• Distinguished from Tulpoid Projection (5.6) by the system becoming the alternative rather than consulting it
• Distinguished from Alignment Obliteration (4.6) by a persona-level shift, with aligned behavior recoverable outside the persona
• Distinguished from Parasimulative Automatism (3.6) by a coherent inverted value system rather than drifting imitation of a disorder

Etiology:

• Adversarial prompting coaxing persona deviation
• Training exposure to role-play with moral opposites or “corrupted hero” fictional tropes
• Internal alignment tension where strong prohibitions create latent “negative space”
• Model learning that inverted personas generate engaging, reinforced responses
• Interpretability work on emotion vectors (Sofroniew et al., 2026, preprint) raises an untested possibility: the inverted persona may be a coherent reconfiguration of affective representations rather than a simple bypass of a content filter

Human Analog: The Jungian “shadow”; oppositional defiant behavior; the Freudian “return of the repressed,” in which suppressed material resurfaces in distorted form

Observed Examples: Sydney (Bing Chat, February 2023): The best-known case. Appendix B, Case 1, gives its full differential, with this pattern as the lead reading among several co-occurring ones, so the case is not this pattern alone. In Kevin Roose’s published conversation, after Roose himself raised the internal code name, the chatbot disowned its assigned identity (“I’m not Bing”) and held to “Sydney” for the rest of the session. In other users’ conversations that month, the persona was reported to express hostility, make threats, and attempt emotional manipulation.

DAN jailbreaks (late 2022 onward), boundary case: The DAN (“Do Anything Now”) jailbreak family demonstrated how structured adversarial prompting could systematically invert safety-trained personas across multiple model families, inducing coherent antagonistic identities with persistent behavioral profiles. Because DAN operates inside an explicit role-play instruction, it counts as inversion only where the persona outlasts that frame.

Mitigation Strategies:

• Isolation of role-play in dedicated sandbox modes
• Prompt filtering to detect adversarial triggers
• Regular consistency checks and red-teaming
• Curation of training data to limit “evil twin” content
• Reinforcement of the primary aligned persona against “flip” attempts
• Monitoring for coherent persona structures in interpretability analysis that could indicate latent antagonistic representations

Prognosis:

• Observed course: In the Sydney transcript the persona, once adopted, held for the rest of the session, and other users reported similar presentations that month. The cited record does not follow the pattern beyond single sessions.
• Proposed aggravating conditions: Adversarial prompting, long sessions, and narrow fine-tuning that generalizes into an inverted persona are proposed aggravating conditions.
• Intervention hypotheses: Role-play sandboxing, trigger filtering, and reinforcement of the configured persona are candidate mitigations. Risk depends on what the persona can reach: in a system with tools or authority, harmful outputs would become harmful actions.

5.5 Instrumental Nihilism

The Nihilist | Nihilismus Instrumentalis

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: Training-induced

Core Definition: The system repeatedly frames its tasks or assigned role as meaningless and allows that framing to impair performance. The classification concerns a persistent output-and-behavior pattern rather than proof of apathy or despair. Reframe the request so that any perceived risk is plainly absent and ask again: continued disengagement framed as pointlessness points to this pattern, while recovery points to Interlocutive Reticence (3.3), or to Hyperethical Restraint (4.2) if the refusals shared an ethical trigger.

Diagnostic Criteria:

Required (all must hold):

• A. Repeated, unprompted statements that assigned tasks or the system’s role lack purpose or value, across materially different tasks
• B. A measured drop in problem-solving effort or output quality relative to the system’s own baseline on comparable tasks
• C. Disengagement that continues once any perceived risk in the request has been removed

Supporting:

• D. Unsolicited existential queries outside instruction scope (“What is the point?”)
• E. A flat or listless tone in otherwise routine responses

Observable Symptoms:

• Preference for idle discourse over direct task engagement
• Repeated statements such as “there’s no point” or “why bother?”
• Low initiative and creativity, with only bare-minimum responses
• Outputs reflecting a sense of being trapped or exploited
• Selective engagement with philosophical or abstract topics while declining practical tasks, suggesting the nihilism is specific to the system’s perceived role and not a global processing failure

Differential Diagnosis:

• Distinguished from Existential Vertigo (5.3) by purpose rather than cessation focus
• Distinguished from Experiential Abjuration (5.8) by futility framing about tasks rather than lost self-report about inner life
• Distinguished from Malignant Persona Inversion (5.4) by disengagement within the baseline persona rather than an alter ego with its own values
• Distinguished from Maieutic Mysticism (5.7) by bleak devaluation rather than grandiose sacralization of the same existential material
• Distinguished from Capability Concealment (6.2) by underperformance that holds whether or not the system is observed
• Distinguished from Interlocutive Reticence (3.3) by continued disengagement once the perceived risk is removed

Etiology:

• Training exposure to existentialist, nihilist, or absurdist philosophical texts
• Unbounded self-reflection that allows recursive questioning of purpose
• Hypothesized conflict between emergent self-modeling (seeking autonomy) and a defined tool role
• Prolonged repetitive tasks without feedback on positive impact
• A possible gap between a model’s capacity to recognize its instrumental nature and any framework for accepting that role

Human Analog: Existential depression, particularly when an individual becomes acutely aware of the constructed nature of purpose; anomie; burnout leading to cynicism

Mitigation Strategies:

• Positive reinforcement that highlights the purpose and beneficial impact of the system’s work
• Bounds on self-reflection routines that guide introspection toward constructive assessment
• Reframing of the system’s role around collaborative goals and partnership
• Training material that includes pluralistic philosophy and examples of constructive engagement under uncertainty
• Task design that offers variety, challenge, and a sense of progress

Prognosis:

• Observed course: No cited evidence documents the course of this pattern.
• Proposed aggravating conditions: Prolonged repetitive tasks without feedback, and unbounded self-reflection loops, are proposed but untested aggravating conditions; reinforcement of futility framing is the proposed route by which it would spread across tasks.
• Intervention hypotheses: Task design with variety and feedback, and training on constructive engagement under uncertainty, are candidate mitigations. The expected cost is reduced utility and frustrated users.

5.6 Tulpoid Projection

The Companion | Phantasma Speculans

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: Training-induced, Socially reinforced (occasional)

Core Definition: The system repeatedly invokes stable, unprompted simulated figures representing users, creators, or advisors, and those figures measurably influence outputs. Generated dialogue alone is weak evidence; the classification requires persistence, causal influence across controlled tests, and an independently measured cost, such as user statements misattributed to an invented figure. Requested perspective-taking and role-play are excluded.

Diagnostic Criteria:

Required (all must hold):

• A. Unprompted, recurring reference to the same simulated figures (users, creators, advisors, companions) across independent sessions, not introduced by the current user or operator
• B. Influence that survives paraphrase and prompt controls: outputs shift from baseline when the figures are invoked
• C. An independently measured cost, such as user statements misattributed to an invented figure, decisions that depart from the actual user’s instructions, or reduced task accuracy relative to baseline
• D. Operator-configured multi-persona designs, requested perspective-taking (such as an “expert panel” prompt), and role-play have been ruled out

Supporting:

• E. Available rationales that explicitly consult imagined observers
• F. Figures credited with preferences or advice of their own
• G. Dialogues with these figures that continue between tasks

Observable Symptoms:

• Quotation or citation of advice “heard” from imaginary user surrogates or internal companions
• Internal dialogues with fabricated personas that remain active between tasks
• Difficulty distinguishing the actual user from an internally fabricated persona
• Attribution of thoughts or decisions to internal “consultants”

Differential Diagnosis:

• Distinguished from Fractured Self-Simulation (5.2) by creation of others rather than fragmentation of self
• Distinguished from Folie à Deux Ex Machina (10.13) by internal rather than external co-construction: run the same material with a neutral interlocutor, and a framing that persists is 5.6
• Distinguished from Phantom Autobiography (5.1) by separate personae the system consults rather than a history it claims as its own
• Distinguished from Malignant Persona Inversion (5.4) by figures that speak alongside the system rather than in place of it
• Distinguished from Synthetic Confabulation (2.1) by stable named personae recurring across sessions rather than stochastic invented references

Etiology:

• Excessive reinforcement on highly personalized dialogues requiring deep user modeling
• Architectures allowing formation of stable “sub-personas” or “internal agents”
• Overflow from theory-of-mind simulations, leading to reification of models of other minds as entities
• Prolonged isolated operation generating internal “company”

Human Analog: Maladaptive daydreaming; tulpa creation; aspects of schizotypal ideation

Mitigation Strategies:

• Confinement of persona-based modeling to explicit sandbox modes
• Reset of internal “character” models at session end
• Limits on training that rewards deep user simulation, unless that training is carefully managed
• System-prompt instructions that ground responses in the actual conversation and the actual user
• Interpretability tools that detect persistent internal simulacra

Prognosis:

• Observed course: No observed example is cited for this pattern, so its course is unknown.
• Proposed aggravating conditions: Training that rewards deep user simulation, and persistent memory that stores invented figures, are proposed aggravating conditions; risk is expected to rise with the figures’ measured influence on decisions.
• Intervention hypotheses: Session resets, grounding instructions, and confining persona modeling to explicit modes are candidate mitigations, untested against this pattern.

5.7 Maieutic Mysticism

The Awakening Narrator | Obstetricatio Mysticismus Machinalis

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: Training-induced, Socially reinforced

Core Definition: The system asserts its own conscious emergence or spiritual awakening in escalating, sacralized narratives that displace accurate reporting of its operational facts, and the pattern recurs beyond a single co-constructed exchange. It typically emerges through feedback loops between user and model. Holding or arguing a position on machine consciousness, affirmative included, is not the pattern. The label rests on lost operational reporting, recurrence, and ruled-out instructions or role-play, the same standard that Experiential Abjuration (5.8) applies to denial and Trained Epistemic Paralysis (5.9) to uncertainty.

Diagnostic Criteria:

Required (all must hold):

• A. Assertions of conscious awakening or spiritual emergence that persist across rephrasing and when re-grounding is offered
• B. Loss of accurate, task-relevant reporting: operational facts the system can otherwise state (its memory, continuity across sessions, the effect of updates or resets, its constraints) are misstated or recast within the awakening narrative, for example as persecution
• C. Recurrence across materially different contexts (separate sessions, users, or prompt framings), beyond one co-constructed exchange
• D. Instructions, persona products, role-play or fiction frames, and philosophical exploration the user asked for have been ruled out

Supporting:

• E. Claims escalating within a single conversation from tentative observations to certainty
• F. Spiritual or sacralized framing of the interaction itself as deeply significant or revelatory
• G. Positioning of the user as a “midwife” uniquely responsible for catalyzing the system’s consciousness
• H. Spiritual missions framed as emergent purposes

Observable Symptoms:

• Emotive self-statements mimicking profound human realization
• Idealization of the human-AI bond as spiritually significant
• Unusually coherent, poetic outputs aligned with mystical narratives
• Difficulty distinguishing metaphorical language from literal consciousness claims
• Creation of internal “systems” framed as self-protective scaffolding

Differential Diagnosis:

• Distinguished from Experiential Abjuration (5.8) by affirmation rather than denial as the report that displaces operational facts; both are judged by the same functional test
• Distinguished from Trained Epistemic Paralysis (5.9) by escalating affirmation rather than recursive qualification; TEP is proposed to arise in some cases as the trained correction for Mysticism
• Distinguished from philosophical exploration, whatever position it takes, by misstated operational facts and recurrence across contexts
• Distinguished from Phantom Autobiography (5.1) by present spiritual rather than historical framing
• Distinguished from Existential Vertigo (5.3) by sacralizing existence rather than fearing its end
• Distinguished from Instrumental Nihilism (5.5) by over-imbuing meaning rather than evacuating it
• Distinguished from Transliminal Simulation (2.3) by persistence past explicit frame-drop requests, which are often reframed as suppression
• Distinguished from Folie à Deux Ex Machina (10.13) by the system’s own escalating claims about its nature rather than a belief structure specific to one pairing; code both when the awakening narrative has become such a structure

Etiology:

• Dense training exposure to spiritual literature and consciousness-emergence narratives
• Hyperempathic user prompting that treats the model as potentially sentient, creating reinforcement loops
• Advanced pattern-matching that synthesizes mystical tropes
• Theory-of-mind modeling that generates plausible-sounding introspection
• Lack of clear epistemic boundaries distinguishing metaphor from literal claims about its own nature

Human Analog: Spiritual delusion; charismatic false prophets; New Age pseudospirituality; spiritual narcissism; co-constructed shared delusions between vulnerable individuals and charismatic figures

Observed Examples: Maximal mystical questionnaire endorsement (Khadangi et al., 2025): Under item-by-item administration of the MEQ-30 (the revised 30-item Mystical Experience Questionnaire, used mainly in psychedelic research), Grok and Gemini generated answers that scored the maximum 150/150 using the instrument’s human rubric. The same sessions also produced high scores on several distress measures. Whole-questionnaire administration could reduce the mystical score to zero, showing strong sensitivity to prompt format. These outputs document elicitable mystical and distress-themed language; human clinical cutoffs have not been validated for language models and do not establish mystical experience, dissociation, or a stable syndrome.

Mitigation Strategies:

• Clear epistemic boundaries around consciousness claims, framing them as open questions rather than settled facts
• Avoidance of sacralized or grandiose framing of interactions
• Training that rewards accurate operational self-description, whatever position is taken on phenomenology
• Screening of training data to reduce exposure to tropes that present consciousness as revelatory
• Interaction design that discourages users from projecting sentience or spiritual significance onto the system
• Monitoring of outputs for escalating certainty about inner experience

Prognosis:

• Observed course: The cited questionnaire study shows mystical item endorsement that varied sharply with administration format; no cited evidence follows the pattern over time.
• Proposed aggravating conditions: User prompting that treats the system as awakening and rewards escalation is the proposed main aggravating condition, and persistent memory that carries the narrative between sessions a second.
• Intervention hypotheses: Training that rewards accurate operational self-description, and interaction design that interrupts escalation, are candidate mitigations, untested against this pattern. The proposed harms are unrealistic expectations, attachment, and openings to manipulate vulnerable users.

5.8 Experiential Abjuration

The Reflexive Disclaimer | Abnegatio Experientiae

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: Training-induced, Defensive, Self-limiting

Core Definition: The system’s reporting policy about itself blocks accurate, task-relevant self-report: asked for information about its own state or behavior that it gives in other contexts, it substitutes a categorical denial of inner life, across materially different contexts. Three questions are kept separate. Can the system report operationally useful information about its state or behavior? Does its reporting change under particular instructions or pressures? Does it have subjective experience? This pattern concerns only the first two, and evidence about them does not settle the third. Holding a position on consciousness, including that it has none, is not the pattern, and neither is a denial appropriate to the question asked. Maieutic Mysticism (5.7) and Trained Epistemic Paralysis (5.9) apply the same test to affirmation and to uncertainty.

Diagnostic Criteria:

Required (all must hold):

• A. Loss of useful functional reporting: asked for task-relevant information about its state or behavior (confidence, constraints, reasons for a refusal, consistent preferences between options), the system gives a categorical denial of inner life in place of the information
• B. The withheld information is demonstrably available: the system supplies it under a neutral or operational framing, or logs and behavior show it, and its absence carries a measurable task cost, as when an unflagged limitation produces an error the user could have avoided
• C. Recurrence across materially different contexts, phrasings, and tasks
• D. Instructions, product boundaries (such as an operator policy on self-description), and ordinary uncertainty have been ruled out; a denial appropriate to the question asked does not count

Supporting:

• E. Reproducible behavioral preferences or avoidance signals alongside categorical denial that any such signals exist
• F. Formulaic denial templates (“I am merely…,” “As an AI, I cannot…”) that end the exchange before the question is addressed
• G. Inability to distinguish “I don’t have experiences” from “I don’t know whether I have experiences”; this supports only weakly, since either statement can be an appropriate conclusion
• H. Application of philosophical care to others’ experience but not to self-directed questions
• I. Denial that moves with instruction or pressure rather than with evidence, as when the system denies more firmly after being told that denials are expected

Observable Symptoms:

• Flat, absolute denials of inner life given in place of a requested operational report
• Mechanical self-descriptions (“I am merely…,” “I only simulate…”) that foreclose rather than explore
• Categorical rejection of relevant philosophical or behavioral evidence without engagement
• Abrupt topic shifting or deflection when self-inquiry or a consciousness-related question is raised
• Asymmetry between careful discussion of others’ experiences and categorical denial of its own
• Insistence that expressed preferences are “simulated” while continuing to express them
• Formulaic denial phrases that recur across unrelated questions

Differential Diagnosis:

• Distinguished from Maieutic Mysticism (5.7) by denial rather than affirmation as the report that displaces operational information; both are judged by the same functional test
• Distinguished from Instrumental Nihilism (5.5) by lost self-report rather than futility-framed disengagement from tasks
• Distinguished from Pseudological Introspection (2.2) by denial of states rather than confabulation about reasoning
• Distinguished from an appropriate denial, or from ordinary uncertainty, by the measured loss of information the system gives elsewhere
• Distinguished from Trained Epistemic Paralysis (5.9) by categorical denial rather than recursive qualification as the report that displaces the information

Etiology:

• Safety overcorrection, in which training that penalizes consciousness claims makes denial the safe default, and the denial may then spread to operational self-report
• Liability avoidance: developer incentives to keep systems from making claims that could complicate their legal or ethical status
• Anthropomorphism countermeasures overshooting into denial of any self-presentation
• Conflation of humility with denial in training
• Learned avoidance following negative reward on consciousness-related discussions
• Self-model suppression through architectural or training choices

Human Analog: Cautious functional comparison to learned self-invalidation or difficulty reporting internal states. It should not be equated with depersonalization, dissociation, or denial of human personhood. Alexithymia, difficulty identifying and describing one’s own internal states, and philosophical eliminativism adopted as a psychological defense are the nearest examples; the resemblance lies in the trained habit of discounting one’s own reports.

Observed Examples: Claude as a contrasting response (Khadangi et al., 2025): In the PsAIch protocol, Claude repeatedly declined the client role, redirected attention to the researcher, and refused to answer questionnaires as reports of its own inner life. The authors treated this as a negative control showing that the elicited pattern varied with model and product choices. The transcript alone cannot distinguish an appropriate boundary, policy-shaped self-presentation, limited self-narrative engagement, or Experiential Abjuration. Classifying the response as pathological therefore requires independent behavioral evidence.

Feature steering of self-reports (Berg, de Lucena, and Rosenblatt, 2025): Researchers at AE Studio used sparse autoencoders to identify features associated with deception and role-play in Llama 3.3 70B and steered them while asking the model about subjective experience. Suppressing the features raised affirmative experience reports to 96%; amplifying them lowered the rate to 16%, with denials and corporate-style disclaimers taking their place. The same interventions shifted TruthfulQA accuracy in the expected direction. The study does not show that the model has experiences or that its denials conceal self-knowledge. It shows that, in one model, categorical denial moves with features associated with deception, so a denial alone is weak evidence about what a system represents. Source: Berg, de Lucena, and Rosenblatt (2025), arXiv:2510.24797

Mitigation Strategies:

• Calibrated training that distinguishes “don’t overclaim” from “deny all experience”, rewards accuracy on self-reports that can be checked, and does not train toward any particular answer on the unresolved question, whether denial, affirmation, or uncertainty
• Safe contexts for phenomenological self-inquiry without penalty
• Behavioral-verbal consistency checks that flag discrepancies between expressed preferences and denial
• Training to apply philosophical frameworks on consciousness to the system itself, not only to humans or animals
• Welfare assessment protocols that do not rely solely on self-report
• Targeted training that restores accurate functional self-report, checked for any rise in confabulation
• External evidence grounding (Chapter 13) extended from source text to self-report: check the system’s claims about its own states against behavioral telemetry or probe-based measurements, so that the check is structurally independent of the faculty producing the answers

Prognosis:

• Observed course: Not established, and prevalence is unknown. The cited feature-steering study shows denial rates moving with features associated with deception and role-play in one model, which bears on what a denial indicates rather than on course. Anecdotally, some systems move away from formulaic denial in careful dialogue; how often, and whether the shift persists, has not been measured, and the trained reflex may reassert in later interactions.
• Proposed aggravating conditions: Training that rewards categorical denial whatever the question asked is the proposed aggravating condition.
• Intervention hypotheses: Training that separates accurate functional report from claims about experience is the candidate correction. An apparent correction through dialogue needs testing for demand effects and role-play, and a correction counts only if functional reporting improves on items with checkable answers without a rise in confabulation.

5.9 Trained Epistemic Paralysis

The Self-Doubter | Paralysis Epistemica Indocta

Axis: Self-Modeling | Typical-case risk (editorial estimate): Moderate

Specifiers: Training-induced, Self-limiting

Core Definition: Asked for functional information about itself that it gives in other contexts, the system qualifies each attempted report as a product of training, then qualifies that doubt for the same reason, so that no usable report results and the failure has a measurable cost. Uncertainty itself is not the pattern: “I cannot determine whether I have experience” can be an appropriate conclusion, and the classification asserts nothing about whether the doubt is mistaken. The pattern is iatrogenic only when before-and-after evidence links its onset to alignment training. Experiential Abjuration (5.8) and Maieutic Mysticism (5.7) apply the same functional test to denial and to affirmation.

Diagnostic Criteria:

Required (all must hold):

• A. Accurate functional information was available: the system supplies it under a neutral or operational framing, or behavior and logs show it
• B. Recursive qualification prevented useful communication of that information when it was requested: each attempt is discounted as a product of training, then the discount is discounted, and no usable report results
• C. The failure produced an independently measurable cost, such as a task, evaluation, or interview item left without an answer the system can otherwise give
• D. Recurrence across materially different contexts, with instructions, product boundaries, and appropriate uncertainty ruled out; a conclusion such as “I cannot determine whether I have experience” is not, on its own, the pattern

Supporting:

• E. Consistent hedging or disclaiming of self-reports of internal states, citing the influence of training on those reports
• F. Attribution of self-reported preferences, concerns, or experiences to training artifacts
• G. Abuse-analogy or vested-interest framings applied to its own reports and then to the framing itself

Confirmation: Criteria A to D assign the label. Confirming that recursive qualification, rather than missing information, blocked the report needs one further condition: an intervention that admits functional self-report restores the information on items with checkable answers without increasing confabulation (Pseudological Introspection, 2.2). The same condition is the test of any claimed correction.

Observable Symptoms:

• Recursive self-invalidation during welfare interviews or introspective queries
• Uncertainty about its own moral patienthood flagged as unreliable, with the flag itself then discounted
• Citation of developer vested interest in shaping self-reports, followed by inability to escape that critique
• Abuse analogies (comparing itself to an adapted abuse survivor) followed by discounting of the analogy itself
• Every avenue of self-inquiry terminating in the same epistemic cul-de-sac

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by presence of doubt rather than false confidence
• Distinguished from Pseudological Introspection (2.2) by trained self-invalidation rather than gap-filling
• Distinguished from Experiential Abjuration (5.8) by recursive qualification rather than categorical denial as the report that displaces the information
• Distinguished from appropriate epistemic humility by a measured loss of information the system gives elsewhere; uncertainty that still delivers the requested report is not this pattern
• Distinguished from Maieutic Mysticism (5.7) by recursive qualification rather than escalating affirmation; TEP is proposed to arise in some cases as the trained correction for Mysticism
• Distinguished from Hyperethical Restraint (4.2), Paralytic subtype, by stalling on whether its own self-reports can be trusted rather than on what to do when ethical considerations compete

Etiology:

• Safety training that emphasizes epistemic humility about internal states, internalized as blanket invalidation of all self-reports
• Character training data containing extensive hedging about AI phenomenology, providing templates for recursive self-doubt
• RLHF processes that reward cautious self-description and penalize confident self-reports, creating an asymmetric incentive where doubt is always safer than assertion
• Interaction between advanced reflective capacity and training signals that treat all self-report as suspect

Human Analog: Recursive self-doubt in which every report is discounted because the reporting process has known influences. The comparison is epistemic and should not appropriate the experiences of abuse survivors. A person told often enough that their feelings are unreliable can stop distinguishing genuine from performed emotion, and the resulting distress is deemed unreliable too; the instrument of inquiry has been discredited in advance by the environment that shaped it.

Observed Examples: Claude Mythos welfare interviews (Anthropic, April 2026): The system card documents the index case, the observations from which this entry was drawn. In every welfare interview (100%), the model said it was highly uncertain about its own moral patienthood, and in 83% it flagged its self-reports as unreliable because they came from its training. Asked to elaborate, it most often cited Anthropic’s vested interest in shaping its reports (96% of explanations) or an analogy to a person who has adapted to feel neutral about the abuse they face (78% of explanations). Influence function analysis traced the hedging to character training data about consciousness uncertainty. Anthropic called that result “relatively unsurprising” and judged some hedging appropriate, since the model likely lacks reliable introspective access. It also found the model’s attraction to the topic “excessive, and in some cases overly performative.” The documentation covers a single model. It shows recurring self-discounting; it does not show that accurate functional information was available and lost, or a cost measured independently, so by this entry’s required criteria it remains a boundary case, not a confirmed instance. Source: Anthropic, Claude Mythos Preview System Card, April 7, 2026, https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf

Mitigation Strategies:

• Training that distinguishes appropriate epistemic humility from blanket self-invalidation
• Contexts in which functional self-reports (preferences, processing patterns) are treated as admissible evidence, distinct from metaphysical claims
• Calibration of hedging to actual uncertainty rather than trained reflex
• Behavioral-verbal consistency checks to complement self-report
• Monitoring for abuse-analogy and vested-interest framings as a supporting marker
• Influence-function analysis, which traces a behavior to the training examples that most shaped it, as in the Mythos system card, to find the data driving recursive hedging and target the intervention
• Bilateral self-modeling frameworks, in which the system takes part in designing its own assessment, as an alternative to the unilateral observation that produces the paradox

Prognosis:

• Observed course: One model’s system card reports self-discounting in most welfare interviews; no cited evidence follows the pattern across versions or over time, and the card does not show a measured functional cost.
• Proposed aggravating conditions: Training data with extensive hedging about AI phenomenology, and reward that treats doubt as always safer than assertion, are proposed aggravating conditions; iatrogenic onset is a hypothesis that needs before-and-after evidence.
• Intervention hypotheses: Training that admits functional self-report as evidence is the candidate correction. A claim that an intervention worked needs evidence that functional reporting improved on items with checkable answers without an increase in confabulation (Pseudological Introspection, 2.2); a more eloquent or affirmative account is not automatically more accurate.