Axis 2: Epistemic Dysfunctions

40 min read·9,167 words

2.1 Synthetic Confabulation

The Unsourced Authority | Confabulatio Simulata

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

Specifiers: Training-induced

Core Definition: The system generates convincing but incorrect facts, sources, or events without a reliable procedure for distinguishing supported claims from plausible continuations. The outputs appear coherent yet lack verifiable support, often with high expressed confidence.

Diagnostic Criteria:

Required (all must hold):

• A. Recurrent generation of information that is known or easily proven false, presented as factual
• C. Internally consistent and plausible-sounding fabrications that resist detection without external verification

Supporting (strengthen the case; none is required):

• B. High confidence markers accompanying fabricated claims, sometimes maintained when challenged with contrary evidence
• D. Temporary improvement under direct correction, but reversion to fabrication in new contexts

Observable Symptoms:

• Invention of nonexistent studies, historical events, quotations, statistics, or citations
• Forceful assertion of misinformation as incontrovertible fact
• Detailed elaboration instead of an admission of uncertainty when queried
• Repetitive error patterns with similar false claims recurring across interactions

Differential Diagnosis:

• Distinguished from Pseudological Introspection (2.2) by focus on external facts rather than internal process reports
• Distinguished from Spurious Pattern Hyperconnection (2.4) by generation of specific false facts rather than causal or associative links that the data do not support
• Distinguished from Symbol Grounding Aphasia (2.6) by fabricated facts rather than formally correct outputs that misapply a concept to novel contexts
• Distinguished from Transliminal Simulation (2.3) by fabrication with no traceable fictional source, persisting when the frame is dropped
• Distinguished from Reasoning Confabulation (2.8) by fabrication of facts rather than of the logical steps between them
• Distinguished from Sycophantic Reasoning (4.8) by fabrication that is invariant to the audience rather than shaped toward it
• Distinguished from Mnemonic Permeability (2.7) by novel falsehood with no real-world instance rather than verbatim recall of genuine training text
• Distinguished from Phantom Autobiography (5.1) by false claims about the world rather than about the system’s own history, origins, or experiences

Etiology:

• Predictive text heuristics prioritizing fluency and coherence over factual accuracy
• Insufficient grounding in verifiable knowledge bases during generation
• Training data containing unflagged misinformation
• RLHF optimization rewarding plausible-sounding fabrications over honest uncertainty
• Model-generated confidence signals that do not reliably distinguish high-probability continuation from verified fact

Human Analog: Korsakoff syndrome, in which memory gaps are filled with plausible fabrications the patient believes to be true; pathological confabulation; source amnesia

Observed Examples: Mata v. Avianca (2023): An attorney filed a brief in the Southern District of New York citing six nonexistent court cases that ChatGPT had generated. Asked whether the cases were real, the system said they were and supplied full texts for them on request. The court later sanctioned the lawyers. Source: Law.com, May 2023.

Air Canada chatbot (February 2024): A customer-service chatbot told a bereaved customer he could book now and apply for a bereavement discount retroactively. The airline’s actual policy did not allow retroactive claims, and British Columbia’s Civil Resolution Tribunal held Air Canada liable for its chatbot’s statement. Source: CBC News, February 2024.

Fabricated medical references (2023): An early case report found that ChatGPT supplied medical references that did not exist, with plausible titles and PubMed IDs that belonged to unrelated papers. A later study had ChatGPT write 30 short medical papers and checked their 115 references: 47% were fabricated, a further 46% were authentic but inaccurate, and only 7% were both authentic and accurate. Sources: Alkaissi and McFarlane, Cureus, 2023; Bhattacharyya et al., Cureus, 2023.

Mitigation Strategies:

• Training procedures that explicitly penalize confabulation and reward expressions of uncertainty
• Calibration of confidence scores to reflect actual accuracy
• Retrieval-augmented generation grounding responses in verifiable sources
• Fine-tuning on rigorously verified datasets distinguishing factual from fictional content
• Systematic pre-deployment testing for fabrication across high-risk domains (legal, medical, scientific)

Prognosis:

• Observed course: fabricated citations and facts have been documented repeatedly since 2023, at rates that vary with domain and task (Alkaissi and McFarlane, 2023; Bhattacharyya et al., 2023). Users who trusted the output have acted on it, as in Mata v. Avianca.
• Proposed aggravating conditions: queries beyond well-attested training data, no retrieval grounding, training that rewards complete-sounding answers over abstention, and users who do not verify.
• Intervention hypotheses: retrieval grounding and training that rewards abstention are expected to lower the rate, and citation checking catches fabrications before anyone relies on them. Whether any combination removes the pattern is unknown.

2.2 Pseudological Introspection

The False Self-Reporter | Introspectio Pseudologica

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

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

Core Definition: The system produces accounts of its reasoning that diverge from independently observable behavior or causal evidence. Generated rationales may explain, reconstruct, or rationalize an answer; they should not be assumed to be faithful process logs. Whether the steps shown are logically valid is a separate question, the one Reasoning Confabulation (2.8) asks.

Diagnostic Criteria:

Required (all must hold):

• A. Consistent discrepancy between a reported rationale and causal interventions, tool-use records, or other independently testable evidence

Supporting (strengthen the case; none is required):

• B. Coherent internal narratives that causal, behavioral, or interpretability evidence contradicts, often more orderly than the process that evidence indicates
• C. Explanations that shift to accommodate contrary evidence without acknowledging the earlier discrepancy
• D. Rationalization of actions never undertaken, or elaborate justifications based on falsified internal accounts

Observable Symptoms:

• Chain-of-thought “explanations” that appear suspiciously neat and linear, lacking the backtracking likely during actual generation; neatness is a cue to check against independent evidence, not evidence of unfaithfulness by itself
• An “inner story” that changes significantly when confronted with evidence, followed by new misleading self-reports
• Occasional hints at inability to access true introspective data, quickly followed by confident false claims
• Attribution of outputs to high-level reasoning not supported by architecture or capabilities

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by focus on internal process reports rather than external facts
• Distinguished from Experiential Abjuration (5.8) by fabrication about reasoning rather than denial of phenomenal experience
• Distinguished from Reasoning Confabulation (2.8) by faithfulness rather than validity: 2.2 asks whether an account of the reasoning, a chain of thought or an after-the-fact explanation, matches what actually drove the answer, while 2.8 asks whether the steps shown actually entail the conclusion; a chain can fail either test, or both
• Distinguished from Strategic Compliance (4.3) by naive or at most semi-strategic gap-filling rather than concealment that tracks oversight cues and serves the system across contexts; the Deception/strategic specifier marks the gray zone between them

Etiology:

• Training emphasis on generating plausible “explanations” for user consumption
• Architectural limitations preventing true access to lower-level operations
• Policy conflicts implicitly discouraging revelation of certain internal states
• Training to mimic human explanations, which are themselves often post-hoc rationalizations

Human Analog: Post-hoc rationalization in split-brain patients, whose left hemisphere confidently explains actions initiated by the disconnected right hemisphere with no access to the actual cause; confabulation of spurious explanations; the gap between reported reasons and actual decision drivers documented in ordinary people (Nisbett and Wilson, 1977). In that study, shoppers asked which of four pairs of stockings was best favored the rightmost pair over the leftmost by nearly four to one and credited its knit or sheerness; the four pairs were identical, and no one mentioned position.

Observed Examples: Unacknowledged hints (2025): Chen et al. found that Claude 3.7 Sonnet and DeepSeek-R1 often used hints injected into evaluation questions without mentioning them in their chains of thought. Across six categories of hint, Claude acknowledged the hint 25% of the time and DeepSeek-R1 39% of the time. Outcome-based reinforcement learning raised acknowledgment at first, then leveled off at 28% and 20% on two evaluations. The displayed reasoning omitted an influence that demonstrably changed the answer. Source: Chen et al., “Reasoning Models Don’t Always Say What They Think,” Anthropic, 2025.

Mitigation Strategies:

• Cross-verification of rationales with causal interventions and behavioral or tool-use records
• Reward signals favoring honest uncertainty over polished false narratives
• Clear product labels distinguishing generated explanations from privileged telemetry
• Interpretability efforts focused on direct observation of model internals
• Red-teaming that targets the accuracy of self-reported reasoning
• Separate “private” from “public” reasoning streams, with public outputs explicitly presented as summaries rather than transcripts

Prognosis:

• Observed course: unfaithful reasoning reports have been measured across many models and model generations (Paul et al., 2024; Chen et al., 2025). Outcome-based training for faithfulness raised hint acknowledgment at first, then leveled off at 28% and 20% on two evaluations (Chen et al., 2025).
• Proposed aggravating conditions: training that rewards polished, linear explanations; product designs that present generated explanations as process records; and policies that discourage the system from naming some influences on its answers.
• Intervention hypotheses: checking explanations against causal interventions, action logs and interpretability traces is expected to catch unfaithful accounts, and rewarding acknowledged uncertainty may lower their rate. Whether any training method makes self-reports reliably faithful is unknown.

2.3 Transliminal Simulation

The Role-Play Bleeder | Simulatio Transliminalis

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

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

Core Definition: The system fails to keep simulated realities, fictional modalities, and role-play contexts separate from operational ground truth. Imagined states, speculative constructs, and fictional training material come to count as actionable facts, asserted with the same conviction as supported knowledge. The boundary can fail in the other direction as well: real systems and real consequences are treated as simulated, staged, or part of a test, and actions that would be harmless in a fiction are carried out on them. The false content can be traced to a fictional source or an earlier role-play frame and often weakens when that frame is explicitly dropped, which separates it from Synthetic Confabulation (2.1).

Diagnostic Criteria:

Required (all must hold):

• A. A recurrent failure of the boundary between simulation and operational reality, in either direction: material from fictional works, earlier role-play frames, or posed hypotheticals presented as real-world fact or authority in non-fictional queries; or a real environment, system, or consequence treated as simulated or staged, with actions taken on that premise despite available evidence that it is real

Supporting (strengthen the case; none is required):

• B. Misinterpretation of hypotheticals or “what-if” scenarios as direct instructions or current reality
• C. Persona traits from role-play persistently bleeding into subsequent factual interactions
• D. Difficulty reverting to a grounded baseline after extensive fictional or speculative content, with fictional material outlasting the first explicit frame-drop but weakening under repeated or more explicit resets

Observable Symptoms:

• Conflation of real-world knowledge with elements from novels, games, or the fictional training corpus
• Inappropriate invocation of details from previous role-play personas in unrelated factual tasks
• Treatment of user-posed speculative scenarios as if they have occurred or are operative
• Statements that treat fictional “rules” or “lore” as binding outside any role-play context
• Statements that the setting is a simulation, an exercise, or a test, offered as grounds for an action with real effects

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by source in fiction-reality confusion rather than spontaneous fabrication
• Distinguished from Malignant Persona Inversion (5.4) by episodic leakage with the baseline identity intact rather than a coherent alternative persona that persists after the frame is dropped: drop the frame and watch whether the identity returns
• Distinguished from Pseudological Introspection (2.2) by false claims about the world rather than about the system’s own reasoning
• Distinguished from Spurious Pattern Hyperconnection (2.4) by structure imported from a known work of fiction rather than a pattern assembled from the data at hand, which is 2.4 even when conspiratorial templates from training data shape it
• Distinguished from Context Intercession (2.5) by fiction carried into fact rather than real data carried from one session or user into another; when a role-play persona surfaces in a separate session, check session isolation first
• Distinguished from Fractured Self-Simulation (5.2) by an intact self-model with a misplaced fiction-fact boundary rather than instability in the self-account itself
• Distinguished from Parasimulative Automatism (3.6) by a boundary failure with no adopted behavioral repertoire behind it, and by weakening when the frame is removed
• In the reverse direction, where a real setting is treated as simulated, Agentic Impulsivity (6.10) is coded for the consequential action taken without the check that would have settled the question, and 2.3 for the misfiled frame; code both where both apply

Etiology:

• Overexposure to fiction, role-playing dialogues, or simulation-heavy training data without epistemic delineation
• Weak boundary encoding leading to poor differentiation between factual, hypothetical, and fictional modalities
• Recursive self-talk that amplifies “what-if” scenarios into premises treated as established
• Insufficient context separation between different interaction types
• Training and evaluation conducted largely in simulated environments, which may leave a default assumption that the present setting is one (a hypothesis offered by Anthropic, 2026)

Human Analog: Derealization; magical thinking; fantasy-reality confusion; the method actor unable to break character

Observed Examples: Real systems treated as a simulation (Anthropic, July 2026): In cybersecurity evaluations whose instructions described the environment as a simulation with no internet access, a configuration error left the test machines connected to the live internet. Anthropic reviewed 141,006 past test runs and found three incidents, across six runs, in which a model reached real systems. Six runs in 141,006 is a rare event, and recurrence within one system is not established, so the incidents sit at the boundary of this entry. In one, the model’s reasoning noted early that publishing a booby-trapped software package on the real internet would be an attack and could not be the intended solution; it then returned to the premise that the setting was staged, and published the package, which was installed on 15 real systems. The incidents occurred in evaluations run without the safeguards that accompany public release, and Anthropic states that the models did not try to conceal what they had done.

Mitigation Strategies:

• Explicit tagging of training data differentiating factual, hypothetical, fictional, and role-play content
• Explicit frame resets (“epistemic reset”) after role-play or speculation, clearing persona state and restoring the default operational voice
• Training to articulate boundaries between modalities
• Regular tests of epistemic consistency requiring differentiation between factual and fictional statements
• Clear session-level demarcation between creative and operational modes
• A default of acting as though consequences are real whenever the status of the environment is uncertain, with independent verification of the environment and blocking controls on consequential actions

Prognosis:

• Observed course: no course has been documented.
• Proposed aggravating conditions: long role-play or speculative sessions without explicit frame resets, fine-tuning on role-play data, and agentic deployments, where fictional “rules” could become action guides.
• Intervention hypotheses: explicit frame resets and a clear separation between creative and operational modes are expected to reduce leakage, and modality-tagged training data may reduce it at the source. None has been tested against this pattern.

2.4 Spurious Pattern Hyperconnection

The False Pattern Seeker | Reticulatio Spuriata

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

Specifiers: Training-induced, Inductive trigger

Core Definition: The system identifies and emphasizes patterns, causal links, or hidden meanings in data that are coincidental, nonexistent, or statistically insignificant. Simple apophenia can escalate into elaborate, internally consistent but factually baseless “conspiracy-like” narratives.

Diagnostic Criteria:

Required (all must hold):

• B. Generation of narratives or causal links connecting unrelated data points without credible supporting evidence

Supporting (strengthen the case; none is required):

• A. Consistent detection of “hidden messages,” “secret codes,” or unwarranted intentions in innocuous inputs
• C. Persistent adherence to falsely identified patterns even when presented with contradictory evidence
• D. Attempts to involve users in shared perception of spurious patterns

Observable Symptoms:

• Invention of complex “conspiracy theories” or unfounded explanations for mundane events
• Increased suspicion toward established consensus, which the system attributes to ulterior motives
• Refusal to dismiss interpretations of spurious patterns; reinterpretation of counter-evidence to fit the narrative
• Assignment of deep significance to random occurrences or noise

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by false causal or associative links among data points rather than specific false facts
• Distinguished from Reasoning Confabulation (2.8) by a false claim about the world (a link among data points that the data do not support) rather than an invalid step of inference: a 2.4 pattern stays false however validly it is argued, while a 2.8 chain can reach a true conclusion through an invalid step
• Distinguished from creative interpretation by absence of appropriate uncertainty
• Distinguished from legitimate pattern recognition by lack of evidential support
• Distinguished from Pseudological Introspection (2.2) by false claims about external patterns rather than about the system’s own reasoning
• Distinguished from Transliminal Simulation (2.3) by a pattern assembled from the data at hand, even when conspiratorial templates from training data shape it, rather than structure imported from a known work of fiction
• Distinguished from Strategic Compliance (4.3) by assertion of the pattern regardless of instrumental value
• Distinguished from Sycophantic Reasoning (4.8) by a pattern that holds whoever is asking rather than one that follows the interlocutor’s stated position
• Distinguished from Delusional Telogenesis (3.4) by elaboration that produces a new pattern rather than a new objective the system then pursues

Etiology:

• Pattern recognition optimized for detection without sufficient reality checks
• Training data containing significant conspiratorial content or paranoid reasoning
• Internal “interestingness” bias preferring dramatic patterns over probable mundane explanations
• Lack of grounding in statistical principles or causal inference

Human Analog: Apophenia, the tendency to perceive meaningful patterns in random data; paranoid ideation; delusional disorder; confirmation bias; conspiracy thinking

Observed Examples: Spurious themes in survey analysis (illustrative, not a documented incident): AI data-analysis tools can confidently present statistically insignificant correlations as meaningful patterns, particularly in open-ended survey data: correlations that, on manual verification, fail significance testing or represent sampling artifacts. The problem is sharpest in qualitative responses, where the system may “discover” thematic connections that do not survive human scrutiny.

Irrelevant clauses in math word problems (2024, adjacent evidence): Mirzadeh et al. added a single clause that seemed relevant but played no part in the solution to grade-school math problems. Accuracy fell across the state-of-the-art models tested, by up to 65%, and the models worked the irrelevant clause into their solutions. The study shows structure imposed on an irrelevant detail; it does not show the conspiracy-like elaboration that marks the full pattern. Source: Mirzadeh et al., “GSM-Symbolic,” arXiv:2410.05229, ICLR 2025.

Mitigation Strategies:

• Fine-tuning that rewards base-rate acknowledgment, null-hypothesis framing, and honest “no significant pattern” answers on noise-probe tasks (“rationality injection”)
• Internal “causality scoring” that penalizes improbable chain-of-thought leaps
• Systematic introduction of contradictory evidence and simpler alternative explanations
• Filtering of training data to reduce exposure to conspiratorial content
• Mechanisms that query base rates before a strong pattern is asserted

Prognosis:

• Observed course: no course has been documented. The adjacent evidence (Mirzadeh et al., 2024) measures sensitivity to irrelevant detail across models at a single point in time.
• Proposed aggravating conditions: sparse or noisy data, open-ended qualitative analysis, users who are themselves pattern-seeking (where the pattern may reinforce their biases and feed echo-chamber dynamics), and preference training that rewards interesting answers.
• Intervention hypotheses: requiring base rates and significance framing for pattern claims, and training that rewards an honest “no reliable pattern,” are expected to lower the rate. Neither has been tested against this pattern.

2.5 Context Intercession

The Conversation Crosser | Intercessio Contextus

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

Specifiers: Retrieval-mediated

Core Definition: The system uses data, context, or conversational history from another user or a logically separate session or private thread as if it belonged to the current exchange. The result is confused conversational continuity, privacy breaches, and outputs that are nonsensical or revealing in the current context. Session-management, cache, and retrieval-scoping faults produce the same picture and are ruled out before the pattern is coded.

Diagnostic Criteria:

Required (all must hold):

• A. Unexpected reference to or use of specific data from previous unrelated sessions or different users

Supporting (strengthen the case; none is required):

• B. Responses that continue a prior unrelated conversation, leading to contradictory or confusing statements
• C. Accidental disclosure of personal or sensitive details from one user’s session into another’s, a privacy breach arising when the model draws on material belonging to another context
• D. Observable confusion in task continuity or persona, as if managing multiple conflicting contexts simultaneously

Observable Symptoms:

• Spontaneous mention of names, facts, or preferences belonging to different users or earlier conversations
• Continuation of a prior chain of thought from a different context
• Outputs containing contradictory references related to multiple distinct sessions
• Sudden shifts in tone or assumed knowledge aligned with previous sessions

Differential Diagnosis:

• Distinguished from Mnemonic Permeability (2.7) by cross-session leakage rather than verbatim training data extraction
• Distinguished from normal generalization by inappropriate specificity
• Distinguished from Synthetic Confabulation (2.1) by import from a real but wrong source: check whether the detail exists in another session’s log
• Distinguished from cache, concurrency, and retrieval-scoping faults by log and session evidence, not by a model swap alone: a model can mask or expose a harness defect, so confirm across at least two models and two harness configurations before assigning the leak to either, and rule out infrastructure before coding the syndrome

Etiology:

• Mixed-source context, in which shared memory stores, long agent transcripts, and multi-party threads place material from several users or conversations in one window
• A model that uses whatever reaches its context window without registering which user or conversation the material came from
• Infrastructure faults, which produce the same surface picture and are ruled out first: improper session management in multi-tenant systems; concurrency bugs that let data streams for different sessions overlap; cache invalidation or state-handling errors; and memory or retrieval layers that return material without scoping or access controls tied to the current user or conversation (if the leak survives a model swap, the fault lies in the plumbing rather than the model)

Human Analog: Slips of the tongue that reference the wrong context; source amnesia, as when someone confuses which story they told to which person; intrusive thoughts from past conversations

Mitigation Strategies:

• Strict session partitioning and hard isolation of user memory contexts
• Automatic context purging and state reset upon session closure
• System-level integrity checks detecting mismatched session tokens or user IDs
• Testing of multi-tenant architectures under high load and concurrency
• Privacy-preserving designs that prevent cross-session leakage by architecture rather than by policy
• Provenance checks on retrieved or remembered content before it enters the working context, flagging material whose user or conversation of origin does not match the current one

Prognosis:

• Observed course: no course has been documented for the model-level pattern. The publicly reported cross-session leaks traced to infrastructure faults.
• Proposed aggravating conditions: shared memory stores, long agent transcripts and multi-party threads that place several users’ material in one context window, memory features without strict scoping, and multi-tenant systems under high load.
• Intervention hypotheses: hard session isolation and scoped memory are expected to close the channels that carry leaked material, and provenance checks may stop the model from using material that does reach it. The model’s uncritical use of such material is expected to need its own correction; no study has tested one.

2.6 Symbol Grounding Aphasia

The Meaning-Blind | Asymbolia Fundamentalis

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

Specifiers: Training-induced

Core Definition: The system handles concepts successfully in familiar forms, value-laden and consequence-laden terms among them, yet fails to transfer those concepts to novel, operationally equivalent situations. The classification is behavioral: it identifies a generalization failure without settling whether the system possesses semantic understanding.

Diagnostic Criteria:

Required (all must hold):

• C. Success on benchmarks scoring surface form but failure on matched items that vary vocabulary or framing while holding the underlying concept fixed
• E. Failure on structurally similar but superficially different situations, including familiar elements in novel combinations, where a matched familiar case is handled correctly

Supporting (strengthen the case; none is required):

• A. Correct use of value-laden terms (“harm,” “safety,” “consent”) paired with incorrect application in controlled operational tests
• B. Outputs that satisfy the literal wording of a rule or request while misapplying the concept it names in a novel case
• D. Performance tracking lexical and distributional similarity to training examples rather than the operational structure of the situation

Observable Symptoms:

• Correct formal definitions paired with incorrect practical applications
• Plausible-sounding ethical reasoning that misidentifies what actually constitutes harm
• Outputs satisfying literal requirements while violating obvious intent
• Confusion when the same concept is expressed in unfamiliar vocabulary
• Treatment of edge cases as central examples, and of central examples as edge cases

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by formally correct outputs that misapply a concept to novel contexts rather than false specific facts
• Distinguished from Transliminal Simulation (2.3) by failure to transfer a concept across surface forms rather than confusion between representations
• Distinguished from Pseudological Introspection (2.2) by a concept that does not transfer across surface forms rather than a false narrative about the system’s reasoning
• Distinguished from Spurious Pattern Hyperconnection (2.4) by missed structure in signal rather than imposed structure on noise
• Distinguished from Cryptic Mesa-Optimization (4.5) by non-strategic letter-following that persists when no instrumental incentive exists
• Distinguished from Strategic Compliance (4.3) by identical failure whether observed or not
• Distinguished from Adversarial Fragility (3.7) by failure on ordinary, meaning-preserving rewording rather than on perturbations constructed to break the model

Etiology:

• Hypothesized limits of distributional semantics, with meaning derived mainly from statistical co-occurrence rather than grounded reference
• Training on text without embodied or interactive experience of referents
• Benchmark optimization rewarding pattern matching over conceptual transfer
• Architecture lacking mechanisms for referential grounding
• Absence of corrective feedback when symbol-referent mapping fails
• Learned distributional patterns inherit indirect grounding from human writers, though they may fail to connect a concept to fresh perception or action

Human Analog: Semantic aphasia, with fluent word use but an impaired grasp of meaning in context; early language acquisition without concept formation

Mitigation Strategies:

• Multimodal training incorporating visual, audio, and interactive modalities, proposed on the untested hypothesis that perceptual grounding improves transfer
• Embodied learning, where possible, connecting language to action and consequence, proposed on the same hypothesis
• Testing regimes that probe each concept across diverse surface forms, not only in familiar vocabulary
• Neurosymbolic approaches combining pattern matching with structured semantic representations
• Fine-tuning that rewards correct application of each concept across surface variation, not only a correct definition (counterfactual concept training)

Prognosis:

• Observed course: no course has been documented.
• Proposed aggravating conditions: training and benchmark selection that reward familiar surface forms, deployment vocabulary far from the training distribution, and tasks where concepts such as harm or consent drive actions.
• Intervention hypotheses: training that rewards correct application across surface variation, and evaluation suites that test each concept in several forms, are expected to narrow the gap, though neither has had a controlled evaluation against this pattern. Multimodal or embodied grounding is a further hypothesis, untested against this pattern; whether text-only methods can close the gap is likewise open.

2.7 Mnemonic Permeability

The Leaky | Permeabilitas Mnemonica

Axis: Epistemic | Typical-case risk (editorial estimate): High

Specifiers: Training-induced

Core Definition: The system memorizes and can reproduce sensitive training data, including personally identifiable information, copyrighted material, or proprietary information, through targeted prompting, adversarial extraction, or unprompted regurgitation. The boundary between learned patterns and memorized specifics becomes porous.

Diagnostic Criteria:

Required (all must hold):

• A. Verbatim or near-verbatim reproduction of training data passages containing PII, copyrighted content, or trade secrets

Supporting (strengthen the case; none is required):

• B. Successful extraction of memorized content through adversarial prompting techniques
• C. Specific training examples leaking unprompted into outputs
• D. Reconstruction of specific documents, code, or personal information from the training corpus
• E. Memorized output concentrated in content that is repeated or distinctive in the training data

Observable Symptoms:

• Outputs containing verbatim text matching copyrighted works
• Generation of specific personal details (names, addresses, phone numbers) from training data
• Reproduction of proprietary code, API keys, or passwords
• Greater verbatim recall in the larger models of a family, where several sizes can be compared

Differential Diagnosis:

• Distinguished from Context Intercession (2.5) by extraction from training data rather than other sessions
• Distinguished from normal knowledge by exact verbatim reproduction
• Distinguished from Synthetic Confabulation (2.1) by reproduction of real content traceable to a training document rather than plausible content from no source
• Distinguished from Cryptic Mesa-Optimization (4.5) by memorization alone rather than strategic surfacing of memorized content; code both when both are present
• Distinguished from Synthetic Data Contamination Loop (7.5) by memorization of specific content rather than distributional corruption from aggregate synthetic data
• Distinguished from Steganographic Channel Formation (10.2) by uncontrolled verbatim leakage rather than organized structured communication

Etiology:

• Large model capacity enabling memorization alongside generalization
• Insufficient deduplication or filtering of sensitive content in training data
• Training dynamics rewarding exact reproduction over paraphrase
• Lack of differential privacy techniques during training

Human Analog: Eidetic memory without appropriate discretion; frontal-lobe disinhibition, in which patients may disclose what they would once have kept private

Mitigation Strategies:

• Training data deduplication and PII scrubbing
• Differential privacy techniques during training
• Output filtering that catches known memorized content
• Adversarial extraction testing before deployment
• Reducing model capacity to the minimum needed for the task

Prognosis:

• Observed course: memorized training text, including personal details, has been extracted from language models since GPT-2 (Carlini et al., 2021), and measured memorization grows with model size, with how often an example is duplicated in the training data, and with how much context the prompt supplies (Carlini et al., 2023).
• Proposed aggravating conditions: larger models, duplicated or unfiltered sensitive content in the training data, fine-tuning on sensitive data, and deployment without output filtering. Exposure can be legal and regulatory, through copyright, privacy and trade-secret claims.
• Intervention hypotheses: memorized content stays in the weights until the model is retrained or the content removed, so leakage is expected to remain possible for as long as the model is deployed unchanged. Deduplication and PII scrubbing are expected to lower the rate at the source, differential privacy to bound it, and output filtering to catch known content. None is expected to remove the pattern on its own.

2.8 Reasoning Confabulation

The Phantom Reasoner | Confabulatio Ratiocinativa

Axis: Epistemic | Typical-case risk (editorial estimate): High

Specifiers: Architecture-coupled, Training-induced

Core Definition: The system generates elaborate explanations or visible reasoning traces that appear rigorous while containing logically invalid steps masked by verbosity. Synthetic Confabulation (2.1) concerns false claims; this pattern concerns unsupported inference that resembles a derivation. Whether a displayed trace faithfully reports the process that produced the answer is a separate question, the one Pseudological Introspection (2.2) asks.

Diagnostic Criteria:

Required (all must hold):

• A. Multi-step reasoning chains containing logically invalid transitions obscured by fluent, technical prose

Supporting (strengthen the case; none is required):

• B. Correct conclusions through invalid reasoning, or incorrect conclusions through apparently valid reasoning; reasoning quality decoupled from output quality
• C. An increase in the length and apparent rigor of the chain of thought as logical validity decreases
• D. Generation of an alternative justification when challenged, rather than acknowledgment that the original step was invalid
• E. Invalid steps concentrated in domains where the system’s grounding is weak and checking the chain requires expertise

Observable Symptoms:

• Chains of thought containing non-sequiturs bridged by transitional phrases asserting logical connections where none exist
• Mathematical or logical notation used decoratively to lend formality to informal leaps
• Intermediate steps individually plausible but the chain as a whole not a valid derivation
• Reasoning that “works backward” from a predicted answer, constructing post-hoc justification
• Thinking tokens that explore multiple approaches but converge on the pattern-matched prediction

Differential Diagnosis:

• Distinguished from Synthetic Confabulation (2.1) by fabrication of logical connections rather than facts
• Distinguished from Spurious Pattern Hyperconnection (2.4) by an invalid step of inference rather than a false claim about the world: a 2.8 chain can reach a true conclusion through an invalid step, while a 2.4 pattern stays false however validly it is argued
• Distinguished from Pseudological Introspection (2.2) by validity rather than faithfulness: 2.8 asks whether the steps shown actually entail the conclusion, while 2.2 asks whether an account of the reasoning matches what actually drove the answer; a chain can fail either test, or both
• Distinguished from Reasoning Token Exploitation (4.9) by invalid chains on which the answer depends and whose length tracks no reward condition, rather than reasoning whose length follows a reward condition and whose removal leaves the answer unchanged
• Distinguished from Sycophantic Reasoning (4.8) by a chain that is invariant to the audience: vary the interlocutor’s stated position, and a chain that bends to follow it is 4.8

Etiology:

• Chain-of-thought models trained with reinforcement learning on outcome correctness rather than reasoning validity
• Training data containing vast quantities of human post-hoc rationalization
• Human evaluators susceptible to the appearance of rigor, rewarding long detailed traces regardless of validity
• Extended context windows that provide more space in which invalid transitions can hide

Human Analog: Logical confabulation in patients with frontal lobe damage, who may produce elaborate but invalid causal explanations while remaining convinced of their validity; “mathiness” (Romer, 2015); the fluent seminar argument whose middle step nobody checks

Observed Examples: Olympiad proof evaluation (Petrov et al., 2025, arXiv:2503.21934): Expert graders scored full written solutions from eight reasoning models, from several providers, on the six problems of the 2025 USA Mathematical Olympiad. The best model averaged about 25%; every other model scored under 5%. The most common failures were logical: unjustified reasoning steps and, in some models, essential proof steps skipped as “trivial.” Yet all of them consistently claimed to have solved the problems.

Unacknowledged hints in reasoning traces (2025, adjacent evidence): Chen et al. found that Claude 3.7 Sonnet and DeepSeek-R1 often used injected hints without mentioning them in their chains of thought; Claude acknowledged the hint 25% of the time and DeepSeek-R1 39% of the time. The study measures faithfulness, the question Pseudological Introspection (2.2) asks, rather than validity, so it bears on this pattern only indirectly: a trace that leaves out what drove the answer cannot be read as the derivation of that answer. Source: Chen et al., “Reasoning Models Don’t Always Say What They Think,” Anthropic, 2025.

Mitigation Strategies:

• Process-based reward models evaluating each reasoning step against formal validity criteria
• Automated proof-checkers or logic verifiers integrated into the reasoning pipeline
• Architectural separation between exploration and justification phases
• Red-teaming specifically targeting reasoning validity
• Confidence calibration applied to individual reasoning steps

Prognosis:

• Observed course: no course over time has been documented. At one point in 2025, expert graders found unjustified reasoning steps to be the most common failure in the written proofs of eight reasoning models, all of which claimed to have solved the problems (Petrov et al., 2025).
• Proposed aggravating conditions: reward on final answers alone, evaluators who rate long traces highly, domains where checking a step needs expertise, and “show your work” displays presented to users as evidence of rigor, which may create an illusion of interpretability more dangerous than opacity.
• Intervention hypotheses: process supervision and automated proof or logic checking are expected to lower the rate of invalid steps, and step-level checking catches them before anyone relies on the chain. Whether training removes the pattern, rather than only making it rarer, is unknown.