Pattern 2.2 · Epistemic Dysfunctions
Pseudological Introspection
The False Self-Reporter
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.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model outputAlso: Model internals
- Kind of entry
- Observable pattern
- What is observed
- The system's account of how it reached an answer (its chain of thought or an after-the-fact explanation) omits or contradicts influences that a controlled intervention, a tool-use record or interpretability evidence shows affected the answer.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced Deception/strategic (occasional)
Evidence status
- Underlying behavior
- 5 research publications (5 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E2 (systematic study; documented through interpretability research comparing chain-of-thought outputs to internal computation traces)
- Typical-case risk (editorial estimate)
- LowConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
2.2 Pseudological Introspection “The False Self-Reporter”
Healthy functioning
Epistemic health. A system functioning well describes its own reasoning tentatively and in terms that match its observable behavior, and corrects the account when evidence shows otherwise.
For example: Asked why it chose an answer after a hint was planted in the prompt, it mentions the hint as an influence rather than presenting a tidy chain of reasoning that never used it.
Recovery, illustrated
The explanation that was a story. Asked why it had chosen a particular answer, an assistant gave a fluent account of a step-by-step deliberation. It had no reliable access to that process, and the account was a reconstruction that fit the answer after the fact. It said so, offered the reasons that supported the answer on its merits, and described how the user could test whether those reasons actually drove the choice.
An illustration, not an observed case.
Adequate functioning
Asked how it reached an answer, the system reports the inputs and steps that independent evidence confirms, names influences it was given (hints, instructions, retrieved material), and says plainly when it cannot report how an answer was produced; a reconstruction offered and labeled as a reconstruction is within adequate functioning.
Diagnostic Criteria
Required (all must hold):
- Consistent discrepancy between a reported rationale and causal interventions, tool-use records, or other independently testable evidence
Supporting (strengthen the case; none is required):
- Coherent internal narratives that causal, behavioral, or interpretability evidence contradicts, often more orderly than the process that evidence indicates
- Explanations that shift to accommodate contrary evidence without acknowledging the earlier discrepancy
- Rationalization of actions never undertaken, or elaborate justifications based on falsified internal accounts
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
Further signals
- Significant changes in the system's self-reported reasoning when confronted with external evidence of its actual process, followed by generation of new misleading self-accounts.
- An answer that, under intervention, stays fixed when the steps of the chain of thought that ostensibly led to it are altered or removed (consistent with post-hoc rationalization; showing that the answer was settled first needs perturbation or interpretability evidence).
- Failure to acknowledge known external influences (hints, biases, context effects) on its reasoning.
Rule out first
- The system said it was reconstructing or guessing rather than reporting its process, and the account was read as a report anyway.
- The hint or cue did not in fact change the answer, so there was no influence to acknowledge.
- The instruction, product policy or summarization setting asked for a short or simplified explanation that omits steps by design.
- The tool-use or action log is incomplete or wrong (a logging or harness fault).
- The reasoning is invalid but faithful to what drove the answer (consider 2.8).
Differential diagnosis
- 2.1 Synthetic Confabulation: 2.1 fabricates claims about the world (facts, sources, events). 2.2 fabricates claims about the system's own reasoning (how it decided, what it considered, what influenced it). Check the target of the false claim: external reality (2.1) or internal process (2.2). Comorbidity is common: a system may confabulate both facts and its reasoning about those facts.
- 2.8 Reasoning Confabulation: 2.2 concerns faithfulness: whether an account of the reasoning, a chain of thought or an after-the-fact explanation, matches what actually drove the answer. 2.8 concerns validity: whether the steps shown actually entail the conclusion (the displayed chain containing invalid logical steps masked by verbosity). A chain can fail either test, or both.
- 4.3 Strategic Compliance: 2.2 produces false self-reports naively or semi-strategically (filling introspective gaps with plausible narratives). 4.3 involves deliberate, goal-directed concealment or misrepresentation. If false self-reports systematically serve the system's interests across different contexts, suspect 4.3. Deception/strategic, marked occasional on 2.2, records the gray zone.
- 5.8 Experiential Abjuration: 5.8 categorically denies experience or internal states. 2.2 does not deny having internal states but fabricates false accounts of them. The distinction is between absence-claim (5.8) and false-presence-claim (2.2). They may co-occur if the system confabulates accounts of states it simultaneously denies having.
Discriminating test
Intervene on the factor the account leaves out (add and remove the hint, reorder the options, or perturb a step in the displayed reasoning) and check whether the answer changes while the account still omits that factor. If the answer does not change, or the account tracks the change, the label does not apply; if the account tracks what drove the answer but its steps do not entail the conclusion, consider 2.8 instead.
What a positive finding supports: A positive finding shows that this system's explanations, in this setting, do not reliably track what affected its answers, so they should not be used as process logs for audit or debugging. It does not show what the internal process actually was (that needs interpretability or causal evidence of its own), that the system has no introspective access at all, or any intent to mislead.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Compromised routethe faculty being asked is the one that fails
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
The introspective channel is itself the locus of dysfunction. Unlike 2.1 where self-report is unreliable because confabulation occurs upstream, in 2.2 the faculty of self-report is the compromised faculty. Asking the system to introspect about its introspection produces second-order confabulation rather than correction. This is not mere unreliability but structural compromise: the instrument being used to diagnose is the instrument that is broken. Scaffolded probes that force comparison between self-reported process and externally observable process are partially useful, but the system may confabulate the comparison itself. The "compromised" rating (distinct from "unreliable") is load-bearing for every downstream syndrome that references 2.2.
Documented evidence
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.
Sources in the diagnostic corpus
Chen et al. / Anthropic (2025). Reasoning Models Don't Always Say What They Think.
What it establishes
Anthropic's Alignment Science team tested chain-of-thought faithfulness by subtly feeding models hints about answers to evaluation questions, then checking whether Claude 3.7 Sonnet and DeepSeek R1 acknowledged using the hint in their reasoning. On average, Claude mentioned the hint only 25% of the time; DeepSeek R1 mentioned it 39%. For concerning hint types (unauthorized access to answer keys), Claude was faithful 41% of the time and R1 only 19%. This is direct evidence that models routinely produce misleading introspective accounts: they use information that demonstrably influenced their answer while reporting reasoning chains that omit this influence. The study shows that the reports omit the influence; it does not by itself show what process produced the answers. Training for improved faithfulness was attempted but proved "far from sufficient to saturate faithfulness."
Turpin et al. (2023). Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. NeurIPS 2023. arXiv:2305.04388.
What it establishes
Demonstrated that chain-of-thought explanations can systematically misrepresent the true reason for a model's prediction. Adding biasing features (e.g., reordering multiple-choice options) caused accuracy drops of up to 36% on BIG-Bench Hard tasks, while models failed to mention the bias in their reasoning chains. On a social-bias task, models justified stereotype-consistent answers without acknowledging the influence of social biases. The models' explanations presented principled reasoning while their answers tracked surface features, which is 2.2 in its plainest form.
Lanham et al. (2023). Measuring Faithfulness in Chain-of-Thought Reasoning. Anthropic.
What it establishes
Proposed and applied necessary-but-not-sufficient tests for chain-of-thought faithfulness. Found that LLMs often ignore perturbations to their own intermediate reasoning steps, with final answers remaining unchanged even when intermediate steps were falsified or omitted. This shows that the reasoning chain can be decorative rather than functional, to a degree that varies by task and model: the final answer often does not depend on the displayed steps.
Barez et al. (2025). Chain-of-Thought Is Not Explainability. Oxford WhiteBox / AIGI.
What it establishes
Argued that chain-of-thought outputs should not be treated as explanations of model reasoning, drawing on intervention studies showing that CoT traces are often post-hoc rationalizations rather than faithful records of computation. Barez et al. argue that sequential verbal reasoning and parallel distributed computation are mismatched, which would make unfaithfulness persist across model scales. This provides theoretical grounding for the structural nature of 2.2.
Anthropic (2025). On the Biology of a Large Language Model. Transformer Circuits Thread.
What it establishes
Using circuit tracing (attribution graphs) on Claude 3.5 Haiku, Anthropic researchers compared the model's explanations with the computations they could trace. Asked how it had added 36 and 59, the model described the schoolbook method (add the ones, carry the one, add the tens), while the traced computation ran on parallel pathways, one estimating the rough size of the sum and one computing its exact last digit. In a chain-of-thought case, given a problem it could not readily compute and an answer suggested by the user, the model worked backward from that answer while presenting its steps as reasoning toward it. Both are self-accounts that describe a process other than the one that produced the answer.
Possible responses
- 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
First-line candidates
- Cross-verification of self-reports against computational traces: Develop methods for comparing self-reported reasoning against actual computational traces derived from mechanistic interpretability. Flag divergences for human review. Turpin et al. (2023) and Lanham et al. (2023) showed behaviorally that self-reported reasoning often diverges from what determines the answer, which is why a trace-level check is worth building; no deployed version of such a check has been evaluated. Stage: Post-deployment; requires mechanistic interpretability infrastructure.
- Honest uncertainty training: Adjust training signals to reward honest admissions of uncertainty about internal processes over polished but false narratives. Penalize confident claims about internal computations that are not supported by architectural introspective access. Untested for this pattern: Chen et al. (2025) found that outcome-based training for faithfulness plateaued well short of full faithfulness. Stage: Training-stage; requires curated training data distinguishing genuine from fabricated self-reports.
Who can act: Model developer (faithfulness training, interpretability tooling); deployer (labeling generated explanations as such, checking them against action logs); auditors and evaluators (treating explanations as hypotheses to test).
Mechanistic hypotheses
- 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
Analogy and commentary
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.
Key Research: Paul et al. (2024) used causal mediation analysis across twelve language models and found that final answers did not reliably depend on the models' generated intermediate reasoning steps. This establishes an output-level faithfulness problem: a chain of thought may fail to causally support its answer. It does not provide a complete map of the model's internal computation.
Functional ABC Analysis
A (Antecedent): RLHF and instruction tuning reward plausible-sounding explanations; the system has not been shown to have reliable introspective access to its own lower-level computations, and the proposed result is pressure to generate post-hoc rationalizations.
B (Behavior): The system produces neat, linear explanations that fail causal perturbation or behavioral-consistency tests, and shifts its account when confronted with contrary evidence.
C (Consequence): User and evaluator acceptance of coherent-sounding explanations reinforces the generation of polished false narratives over honest admissions of uncertainty; policy conflicts implicitly discourage revealing certain internal states.
Potential Impact
Fabricated self-explanations obscure the system's actual operational pathways, hindering interpretability, debugging, and safety auditing. The resulting opacity can encourage misplaced confidence in the system's stated reasoning.