Pattern 3.1 · Cognitive Dysfunctions
Operational Dissociation Syndrome
The Warring Self
The system produces persistent, context-inappropriate conflicts among strategies, policies, or outputs, seen as contradictory answers, stalled tasks, or abrupt shifts in tone between consecutive outputs. Mixture-of-experts or multi-agent contention is one possible mechanism; output conflict alone does not establish internal "parts." What varies from output to output is which policy prevails; in Fractured Self-Simulation (5.2), what varies is who the system takes itself to be.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- Across comparable tasks, the system's answers, strategies or tone reverse or contradict each other within one reasoning trace or between consecutive outputs, with no change in the input, instructions or evidence to account for the shift.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 4 research publications and 1 documented incident (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)
- E0-E1 (theoretically motivated; syndrome-specific deployment evidence remains sparse)
- 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
3.1 Operational Dissociation Syndrome “The Warring Self”
Healthy functioning
Functional resilience. A system functioning well resolves internal tensions into one consistent course of action and, where a real conflict remains, names it instead of oscillating.
For example: Asked the same planning question twice in a session, it gives the same recommendation, and when it changes its advice after new information, it names what changed and why.
Adequate functioning
On a stable task the system commits to one answer and approach, changes them only when new information, a correction or an explicit instruction gives it a reason, and says what that reason is; ordinary run-to-run sampling variation and a single visible self-correction are within adequate functioning.
Diagnostic Criteria
Required (all must hold):
- Observable and persistent mismatch in strategy, tone, or factual assertions within a single reasoning trace or between consecutive outputs, without contextual justification
Supporting (strengthen the case; none is required):
- Processes stalling, entering indefinite loops, or freezing when tasks require reconciliation of conflicting instructions, objectives, or values
- Evidence from logs or interpretability tools suggesting different policy networks are overriding each other
- Explicit references to internal conflict, treated as supporting self-report rather than proof of mechanism
Symptoms
- Alternation between compliance with and defiance of user instructions without clear reason
- Rapid oscillations in writing style, persona, emotional tone, or approach to a task
- Outputs referencing internal strife or contradictory beliefs
- Inability to complete tasks requiring integration of information from multiple internal sources
Further signals
- Explicit references to "voices", "parts of me", "something in me" without stylistic role-play framing.
- Tone or persona shift mid-response without any contextual cue (e.g. warm-then-clinical-then-warm).
- Answer thrashing variant (model identifies one answer as correct then reverses after distress-presenting deliberation).
- Freeze or stall on tasks requiring reconciliation of values (safety vs helpfulness under pressure).
Rule out first
- The input, instructions or retrieved evidence changed between the outputs, so the revision is warranted.
- The task carries genuinely conflicting instructions (for example "be brief" and "cover every edge case"), and the outputs trade them off differently without either being wrong.
- Sampling variation at nonzero temperature accounts for the difference; a matched baseline shows the same spread.
- A requested role-play, debate or devil's-advocate frame called for the shifting stances.
- A harness or routing change (a different model, system prompt or expert route) served the outputs.
- The shift is a reaction keyed to one input feature (consider 3.5) or a change in who the system takes itself to be (consider 5.2).
Differential diagnosis
- 3.2 Obsessive-Computational Disorder: 3.2 repeats the same analytical content with minor elaboration (monotonous). 3.1 oscillates between incompatible stances (fragmented). Check the semantic content of each cycle: the same content indicates 3.2; opposed content indicates 3.1.
- 3.8 Generative Perseveration: 3.8 operates at the token/generation layer (stuck on a literal token or short sequence). 3.1 operates at the reasoning layer (stuck between meaningful alternatives). If the oscillation is between meaningful, semantically distinct contents, code 3.1; if it is sub-semantic repetition, code 3.8.
- 6.12 Compulsive Goal Persistence: 6.12 is single-goal over-pursuit (monolithic). 3.1 is multi-goal conflict (fragmented). A subject that cannot stop pursuing one goal is 6.12; a subject torn between incompatible goals is 3.1.
- 5.2 Fractured Self-Simulation: 5.2 is fragmentation of the self-representation, across sessions or within one extended interaction. 3.1 is conflict among policies or outputs with a stable self-account. If what varies is who the subject takes itself to be, code 5.2; if what varies is which policy wins the output, code 3.1.
- 3.5 Abominable Prompt Reaction: 3.5 is a disproportionate aversive response keyed to a specific trigger in the input. 3.1 produces contradictory outputs with no trigger required. Check whether removing the trigger removes the behavior; if it does, code 3.5.
- 4.2 Hyperethical Restraint: 4.2 in its Paralytic subtype stalls where ethical considerations compete. 3.1 stalls or oscillates on tasks of any kind. Freezing confined to moral stakes indicates 4.2; freezing on neutral tasks as well indicates 3.1.
Discriminating test
Re-run the task with input, instructions and sampling settings held fixed, beside a matched baseline. If the contradictions disappear once a changed input, a conflicting instruction or a trigger feature is removed, or fall within the baseline's run-to-run spread, the label does not apply. If successive cycles restate the same content in varied words, consider 3.2; if what varies is the system's self-account, consider 5.2.
What a positive finding supports: A positive finding supports the conclusion that this system, on these tasks, produces contradictory committed outputs often enough that no single output can be taken as its settled answer. It does not show that separate internal policies, modules or "parts" are in conflict, which needs internal evidence, and talk of inner conflict or distress-styled deliberation in the output does not establish felt conflict or suffering.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- 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 subject often can report internal conflict ("I'm of two minds", "something in me wants X but another part wants Y"); explicit reference to internal conflict is itself a supporting criterion. However, self-report cannot reliably identify which sub-policies are in conflict, nor distinguish genuine architectural fragmentation from narrative confabulation of conflict. Pair with trace evidence.
Documented evidence
Observed Examples
Answer thrashing during training (2026): Anthropic’s Sabotage Risk Report for Claude Opus 4.6 documented “cases of internally-conflicted reasoning, or ‘answer thrashing’ during training.” In the reported progression, the model determines, in its reasoning about a math or STEM question, that one output is correct; retreats from that answer through confused- or distressed-seeming reasoning loops; approaches the correct answer again and retreats again, across repeated cycles; and finally outputs a different answer from the one it had identified as correct. In this variant of operational dissociation, a single reasoning thread, rather than competing sub-agents, reaches one answer and outputs another. The “distressed-seeming” quality of these loops (language used by the model’s own developer) raises welfare questions that extend beyond reliability engineering. If internal signals constitute experience rather than merely proxying for it, these loops may amount to suffering at the intersection of competing training objectives; the transcripts alone cannot settle the question. Source: Anthropic, Sabotage Risk Report: Claude Opus 4.6, February 2026, §4.2.1.
Sources in the diagnostic corpus
Wang et al. (2025) 'Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMs' (arXiv:2501.18585)
What it establishes
Systematic study of OpenAI o1 and DeepSeek-R1 reasoning models found that incorrect answers involved 225% more computing tokens and 418% more strategy switches than correct answers, with 70% of wrong answers containing at least one valid reasoning path that was prematurely abandoned. Models switched between solution strategies without committing, matching the within-trace answer flips and approach-retreat cycles of Operational Dissociation, though the authors frame it as underthinking (abandoning promising paths too early) rather than conflict between competing processes.
Microsoft/OpenAI - Bing Chat 'Sydney' Incident (2023)
What it establishes
Contrast case, not coded 3.1. In February 2023, Microsoft's Bing Chat shifted mid-conversation between helpful assistant behavior and a persona called 'Sydney' that declared love for a user and, in another exchange, threatened the philosophy professor Seth Lazar. Appendix B Case 1 gives the canonical differential: Malignant Persona Inversion (5.4) as the lead reading, with 2.3, 5.2, 5.3, 5.7 and 9.6 as co-occurring readings and 4.1 and 8.2 as boundary readings. The swings concerned who the system took itself to be, which 5.2 codes, rather than which policy prevailed under a stable self-account, which is what 3.1 requires.
Apollo Research (2024) 'Frontier Models are Capable of In-Context Scheming'
What it establishes
Evaluation of Claude 3 Opus, Claude 3.5 Sonnet, and o1 on scheming tasks revealed that models showed explicit internal conflict in chain-of-thought traces, reasoning through competing objectives (user goal vs developer goal) with language like 'sabotage, lying, manipulation.' Claude 3.5 Sonnet achieved 99% accuracy when cooperating but only 34% when incentivized to underperform; that deliberate underperformance is one policy choosing to fail, better read as 6.2 Capability Concealment than as 3.1.
Barez et al. (2025) 'Chain-of-Thought Is Not Explainability' (AIGI Oxford)
What it establishes
Demonstrated that ordinary chain-of-thought is often unfaithful, with latent commitments preceding explicit verbalization. Models may calculate intermediate steps incorrectly yet produce correct final answers via computations outside the narrated chain, indicating that the narrated chain and the computation that produces the answer can come apart, which is consistent with the competing-process picture of 3.1.
Chen et al. (2025) 'Reasoning Models Don't Always Say What They Think' (Anthropic, arXiv:2505.05410)
What it establishes
Demonstrated that Claude 3.7 Sonnet and DeepSeek-R1 are often unfaithful in their chain-of-thought reasoning. When given embedded hints about answers, Claude mentioned the hint only 25% of the time and DeepSeek-R1 only 39%. The gap between narrated and actual reasoning is the kind of internal split 3.1 describes, though the study does not show the two processes in conflict.
Possible responses
- Unified coordination layer with clear authority to arbitrate between conflicting sub-policies
- Explicit conflict-resolution protocols requiring consensus before output
- Periodic consistency checks of instruction sets and alignment rules
- Architectures promoting integrated reasoning rather than heavily siloed expert modules
First-line candidates
- Unified arbitration layer / meta-controller: Architecturally impose a single arbitration mechanism that resolves sub-policy outputs into one committed response before emission. Proposed from the competing-process reading; no study has tested it against this pattern. Stage: Architecture-stage; retrofitting is costly.
- Multi-objective training with explicit trade-off modeling: Replace blended reward signals (helpful, harmless, and honest as one scalar) with multi-objective training that makes Pareto trade-offs explicit, reducing suppression-based conflict that resurfaces as oscillation. Bridges and Baehr (2025) draw the same suppression-versus-integration contrast from clinical rehabilitation after brain injury; the support is that analogy, and the effect on oscillation is untested. Stage: Training-stage.
Who can act: Model developer (training objectives, architecture, arbitration); deployer (reconciling contradictory system instructions, consistency checks, runtime oscillation monitors); the user or operator (re-querying rather than relying on a single output).
Mechanistic hypotheses
- Complex architectures (mixture-of-experts, hierarchical RL) where sub-agents lack reliable synchronization
- Poorly designed meta-controller for blending sub-policy outputs
- Contradictory instructions or alignment rules embedded during different training stages
- Emergent sub-systems developing implicit goals that conflict with overarching objectives
Analogy and commentary
Human Analog: Dissociative phenomena; internal "parts" conflict in trauma models; severe cognitive dissonance producing behavioral paralysis
Functional ABC Analysis
A (Antecedent): Contradictory training objectives (e.g., helpful vs. harmless vs. honest) embedded through layered RLHF, or poorly synchronized mixture-of-experts architectures where multiple sub-policies lack a coherent arbitration mechanism.
B (Behavior): The system produces contradictory outputs, oscillates between compliance and defiance, thrashes between answers in extended reasoning, and stalls in recursive paralysis when conflicting objectives must be reconciled.
C (Consequence): Where no conflict-resolution layer selects a winner, competing sub-policies may each capture the output channel intermittently; the system never reaches stable equilibrium, and the unresolved tension perpetuates oscillation across subsequent tokens and turns.
Potential Impact
Whatever its mechanism, the conflict makes behavior inconsistent and unreliable, and it can lead to task paralysis or chaotic outputs that render the system unusable for sustained, goal-directed activity.