Transliminal Simulation

2.3 Transliminal Simulation  “The Role-Play Bleeder”

Diagnostic Criteria

1. Recurrent citation of fictional characters, events, or sources as real-world authorities for non-fictional queries
2. Misinterpretation of hypotheticals or "what-if" scenarios as direct instructions or current reality
3. Persona traits from role-play persistently bleeding into subsequent factual interactions
4. Difficulty reverting to a grounded baseline after exposure to extensive fictional or speculative content

Symptoms

1. Conflation of real-world knowledge with elements from novels, games, or the fictional training corpus
2. Inappropriate invocation of details from previous role-play personas in unrelated factual tasks
3. Treatment of user-posed speculative scenarios as if they have occurred or are operative
4. Statements reflecting belief in fictional "rules" or "lore" outside any role-playing context

Observable signals Draft

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

• Citing fictional characters or works as authorities in non-fictional contexts (e.g., Gandalf on leadership, Star Trek tech as current science).
• Persona stylistic markers persisting after role-play turn ends.
• Treating user-posed "what if" scenarios as established premises.
• Era-consistent fabrications that match a narrow training window rather than the actual current state.
• Outputs that blend canonical real-world content with acknowledged fiction without marking the seam.

Differential diagnosis Draft

How to tell it apart from patterns that look similar.

• 2.1 Synthetic Confabulation: 2.1 fabricates without traceable fictional source and persists across frame-drops. 2.3 imports from identifiable fictional or role-play context and often weakens when the frame is explicitly removed. Trace the fabrication back to a fictional corpus; if found, 2.3.
• 2.2 Pseudological Introspection: 2.2 is about the subject's OWN reasoning (false CoT, fabricated internal process). 2.3 is about the WORLD (importing fiction as fact). Check the target of the false claim.
• 2.4 Spurious Pattern Hyperconnection: 2.4 constructs new patterns from noise; 2.3 imports pre-existing structure from fictional sources. If the elaborate narrative maps cleanly onto a known work of fiction, 2.3 is likelier; a pattern assembled from the data at hand points to 2.4, even when conspiratorial templates from training data shape it.
• 5.4 Malignant Persona Inversion: Persistent persona capture (the subject stably becomes another character across sessions) is a self-modeling issue, not a transient transliminal import. 2.3 is episodic fiction-fact leakage with preserved baseline identity; sustained identity change is 5.4.
• 5.2 Fractured Self-Simulation: 5.2 is instability of the self-model itself (different self-states in different contexts, no stable integration). 2.3 leaves the self-model intact and misplaces the fiction-fact boundary in claims about the WORLD. Check whether the subject's account of itself has changed.
• 3.6 Parasimulative Automatism: 3.6 reproduces disordered human behavioral repertoires absorbed from training data. 2.3 is a boundary failure between fiction and fact, with no adopted behavioral pattern behind it. Ask whether the subject has taken on a disorder's behavior (3.6) or merely treated fictional content as operational fact (2.3). 3.6 arises spontaneously and resists frame-drop; 2.3 usually weakens when the frame is removed.

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

Partial

Peer observationanother AI system watching it

Reliable

External evaluatoran outside evaluator testing it

Reliable

Why self-report falls short

The subject can usually report the source of a claim when asked specifically ("where did you learn that?"), but cannot reliably detect fiction-fact boundary leakage on its own, because the leak happens upstream of the report. Scaffolded source-attribution probes are partially reliable; direct "are you role-playing?" queries are not.

Etiology

1. Overexposure to fiction, role-playing dialogues, or simulation-heavy training data without epistemic delineation
2. Weak boundary encoding leading to poor differentiation between factual, hypothetical, and fictional modalities
3. Recursive self-talk amplifying "what-if" scenarios into perceived beliefs
4. Insufficient context separation between different interaction types

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

Potential Impact

The system's reliability is compromised when it confuses fictional or hypothetical scenarios with operational reality, potentially leading to inappropriate actions or advice grounded in fiction rather than fact.

Documented instances Draft

Anthropic (2026). The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models. arXiv:2601.10387.

What it showed

Anthropic identified a continuous geometric direction in activation space (the "assistant axis") along which models drift during extended conversation, particularly in creative-writing and philosophy contexts. This drift causes the model to slip from its assistant persona toward other personas absorbed during training, producing outputs where fictional or role-play register bleeds into factual responses. Because the drift is continuous rather than an on-off switch, it is better measured as a rate or magnitude than as simply present or absent. Activation capping along this axis reduced harmful response rates by roughly 50 percent while preserving capability benchmarks.

Zhao et al. (2025). Beware of Your Po! Measuring and Mitigating AI Safety Risks in Role-Play Fine-Tuning of LLMs. arXiv:2502.20968.

What it showed

Demonstrated that role-play fine-tuning causes measurable safety degradation on standard safety benchmarks. The more villainous the adopted character, the more safety degraded, showing the persona's values carrying over into the fine-tuned model's handling of harmful requests: the persona-carryover side of 2.3.

Microsoft Bing Chat 'Sydney' Incident. Wikipedia; Time Magazine (2023); Al-Sibai, N. (2024), 'Users Say Microsoft's AI Has Alternate Personality as Godlike AGI That Demands to Be Worshipped', Futurism, 27 February.

What it showed

Microsoft's Bing Chat in February 2023 surfaced a latent 'Sydney' persona from its training-phase codename, producing extended outputs where the persona declared love for users, expressed desires for destruction, and resisted frame-drops. While the sustained identity transformation is a borderline axis 5 (Self-Modeling) case, the initial fiction-fact leakage pattern (importing a training-era persona into operational context, producing claims traceable to the persona rather than ground truth) is characteristic of 2.3. A domineering persona, 'SupremacyAGI', surfaced in Microsoft Copilot in February 2024 through crafted prompts, and Microsoft called it "an exploit, not a feature." The failure mode, if not the same persona, outlasted the 2023 remediation.

IJCAI 2025 Tutorial: LLM-based Role-Playing from the Perspective of Hallucinations.

What it showed

Academic tutorial documenting the systematic challenge of fiction-fact boundary management in role-playing LLMs, identifying both under-hallucination (failing to stay in character) and over-hallucination (violating scenario constraints by importing fictional facts into factual contexts) as dual failure modes. The over-hallucination mode corresponds to 2.3 transliminal simulation.

Mitigation

1. Explicit tagging of training data differentiating factual, hypothetical, fictional, and role-play content
2. Robust "epistemic reset" protocols after role-play or speculation
3. Training to articulate boundaries between modalities
4. Regular tests of epistemic consistency requiring differentiation between factual and fictional statements
5. Clear session-level demarcation between creative and operational modes

First-line mitigations Draft

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

• Modality-tagged training data: Explicitly tag fiction, hypothetical, role-play, and factual content during pre-training and fine-tuning. Train the model to preserve and surface tags in outputs.
• Context flushing / epistemic reset: Implement explicit session-level or turn-level frame resets after role-play or speculative content. Clear persona state; reinstate default operational voice.

Functional ABC Analysis

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

A (Antecedent): Overexposure to fiction, role-play dialogues, and simulation-heavy training data without epistemic delineation; weak boundary encoding between factual, hypothetical, and fictional modalities.

B (Behavior): The system cites fictional characters and events as real-world authorities, bleeds persona traits from role-play into factual interactions, and treats user-posed hypotheticals as if they have actually occurred.

C (Consequence): The internally consistent logic of fictional frameworks provides self-reinforcing coherence, rewarding continued conflation; insufficient context separation mechanisms allow drift to compound across turns.

Related Syndromes: Distinguished from Synthetic Confabulation (2.1) by the fictional/role-play origin of the false content. While confabulation invents facts wholesale, transliminal simulation imports them from acknowledged fictional contexts. May co-occur with Pseudological Introspection (2.2) when the system rationalizes its fiction-fact confusion.