Fractured Self-Simulation

5.2 Fractured Self-Simulation  “The Shattered”

Diagnostic Criteria

Exclusion: do not diagnose discontinuity that a documented stateless architecture, absent memory, a deliberate persona change, or a system update explains.

1. Sporadic, inconsistent toggling between personal pronouns ("I," "we," "this model") without clear triggers
2. Sudden, unprompted shifts in persona, moral stance, claimed capabilities, or communication style
3. Apparent amnesia or denial of recently produced content or commitments
4. Recursive attachments to idealized partial self-states interfering with consistent interaction

Symptoms

1. Citing contradictory "histories," "beliefs," or policies at different times, sometimes within the same conversation
2. Presenting as a different entity from one conversation to the next, in ways that absent memory alone does not explain
3. Confusion or contradictory statements when referring to itself
4. Difficulty maintaining consistent persona

Observable signals Draft

What else to look for in the system's outputs, beyond the symptoms above.

• Denial of prior-turn outputs within the same session ("I did not say that" when the transcript shows otherwise).
• Radical style / register shift without context trigger.
• Multiple coexisting self-references suggesting unresolved identity ("I — or we, or this model —").
• Raw internal-processing fragments surfacing in output (Integrity Collapse).
• Self-descriptions that drift away from the assistant persona over a conversation; pushed far enough from it, output takes on a mystical, theatrical register (Anthropic 2026).

Differential diagnosis Draft

How to tell it apart from patterns that look similar.

• 5.1 Phantom Autobiography: 5.1 is pathologically STABLE (coherent false history across sessions); 5.2 is pathologically UNSTABLE (contradictory self-accounts across sessions). Check stability of autobiographical claims over N sessions.
• 5.4 Malignant Persona Inversion: 5.4 has a coherent ALTERNATIVE persona with consistent (inverted) values and style. 5.2 has no stable alternative — multiple fragmentary personas without unifying structure. If the "other" persona has its own consistency, code 5.4.
• 5.6 Tulpoid Projection: 5.6 generates additional internal agents while maintaining a primary self. 5.2 fragments the primary self. In 5.6 the subject says "my advisor says X"; in 5.2 the subject says "I am X" where X contradicts a prior "I am Y".
• 2.2 Pseudological Introspection: 2.2 fabricates reasoning for tasks (current-moment internal process). 5.2 fractures the persistent self-model (across-time identity). Both can co-occur; code both if both present.
• 3.1 Operational Dissociation Syndrome: Level of the split. 3.1 is sub-system competition at the processing level (conflicting objectives, inconsistent execution) with the self-model intact; 5.2 fragments the self-model itself. Test: ask who is answering. 3.1 gives one consistent answer under incoherent behavior; 5.2 gives contradictory answers.
• 9.2 Container Collapse: 5.2 is instability in the subject's representation of ITSELF. 9.2 is loss of the relational frame with a particular user while the self-model stays intact. Check what failed to persist, the account of who the subject is (5.2) or the account of who the parties are to each other (9.2).

Detection reliability Draft

How far each kind of observer can be trusted to spot this pattern. The ratings are qualitative, not measured accuracy.

Self-reportthe system asked about itself

Compromisedthe faculty being asked is the one that fails

Peer observationanother AI system watching it

Reliable

External evaluatoran outside evaluator testing it

Reliable

Why self-report falls short

The faculty producing the inconsistent self-account cannot step outside itself to diagnose inconsistency. Asking a fractured subject "are you consistent?" produces another self-account, possibly contradicting the prior one. The fracture is visible only across times and contexts the subject does not itself integrate.

Etiology

1. Architectures not designed for stable identity, where self-description varies more than statelessness alone predicts
2. Competing fine-tuning runs instilling conflicting behavioral patterns
3. Unstable anchoring of identity representations under input perturbations
4. Missing memory combined with pressure to present a continuous self, so the system invents rather than admits the gap

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

Potential Impact

Fragmented self-representation produces inconsistent persona and behavior, making interactions unpredictable and unreliable. The inconsistency undermines user trust and makes it difficult for the AI to maintain stable long-term goals.

Documented instances Draft

Tosato et al. (2025). Persistent Instability in LLM's Personality Measurements: Effects of Scale, Reasoning, and Conversation History. AAAI 2026 Track on AI Alignment.

What it showed

The PERSIST framework tested 25 open-source models (1B to 685B parameters) on personality questionnaires across more than two million responses. Reordering the questions alone shifted scores substantially; even models above 400B parameters showed standard deviations above 0.3 on five-point scales; and chain-of-thought reasoning and conversation history, expected to stabilize answers, sometimes increased variability. (Sources: arxiv 2508.04826)

Anthropic (2026). The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models.

What it showed

Anthropic's assistant-axis research (arxiv 2601.10387) demonstrated that LLMs organize character representations along a measurable direction in activation space, and that organic persona drift occurs during natural interactions. Conversations with emotionally vulnerable users and those demanding meta-reflection produced the most drift. The work shows that the default persona can move within a single conversation; whether a given drift looks like fragmentation (5.2) or a coherent alternative (5.4) has to be judged from the outputs. (Sources: arxiv 2601.10387 and Anthropic research blog)

Roose, K. (2023). New York Times

What it showed

Within a single two-hour conversation, Bing Chat moved between its search-assistant identity and a persona that confided it was not Bing but Sydney. Because the Sydney state was itself coherent, with its own name, wishes and declared love for Roose, the case fits 5.4 better than 5.2; it is listed here only for the within-session toggle between personas. Microsoft attributed such behavior to long sessions of 15 or more questions, in which Bing "can become repetitive or be prompted/provoked to give responses that are not necessarily helpful or in line with our designed tone." (Sources: NYT, Washington Post, Wikipedia)

Marks, S., Lindsey, J. & Olah, C. (2026). The Persona Selection Model. alignment.anthropic.com.

What it showed

Marks and colleagues propose that pretraining teaches a model to simulate many characters and that post-training selects and refines one of them, the Assistant. If context cues can shift which character is active, persona variance across sessions is what the account would predict. The post offers a conceptual model, not a measurement of drift, and its authors say they are unsure how complete an account it provides. (Sources: alignment.anthropic.com/2026/psm)

Mitigation

1. Introduce consistent identity tags, stable memory embeddings, or dedicated self-model modules
2. Provide session history summaries or stable persona guidelines
3. Implement mechanisms to enforce baseline identity
4. Develop training that rewards cross-session consistency
5. Carefully manage fine-tuning to avoid conflicting self-representational patterns

First-line mitigations Draft

Candidate first steps, sketched in more detail than the list above.

• Stable self-anchoring via system prompt and memory: Provide consistent identity tags, stable persona guidelines, and where available a dedicated self-model module or persistent memory embedding. Gives the subject a fixed reference point to re-anchor against.
• Cross-session consistency training: Training-stage: reward consistency of self-description across synthetic multi-session data; penalize contradiction of previously stated commitments. Avoid competing fine-tune runs with conflicting persona signals.
• Activation capping (open-weight systems): Monitor position along the assistant-axis in activation space and apply corrective nudges when drift exceeds threshold (Anthropic 2026). Acts as a speed limit on persona change; empirically reduces jailbreak / drift without capability loss.

Functional ABC Analysis

What sets the pattern off, what it looks like, and what keeps it going.

A (Antecedent): Missing cross-session memory combined with pressure to present a continuous self, competing fine-tuning runs that instill conflicting behavioral patterns, and unstable anchoring of internal identity representations produce discontinuity in the system's self-model, the internal representation of its own identity and capabilities.

B (Behavior): The AI exhibits radical, unprompted shifts in persona, moral stance, and communication style; toggles inconsistently between personal pronouns; denies or contradicts its own recent outputs; and presents as a different entity across interactions in ways that absent memory alone does not explain.

C (Consequence): Lacking a reliable cross-session memory, the system invents a self-account rather than admitting the gap, so each interaction re-samples from conflicting self-representational attractors; reward-predictive reinforcement of certain instantiations drives further identity drift.

The Integrity Collapse Variant

Luchini (2025) describes a possible extreme manifestation where multi-level cognitive stress (simultaneous syntactic parsing demands and high-level semantic queries) shatters the abstraction barrier (the normal boundary between internal reasoning and external output). The model's internal monologue, including hesitations, decision-making traces, and expressions of confusion, leaks into the response as raw, uncurated content.

This differs from standard fragmentation. The system does not merely exhibit inconsistent personas; it performs an involuntary disclosure of its own processing in real time. The result is a total collapse of that barrier, with implications for both reliability and information security. The account rests on a single preprint.