Pattern 2.3 · Epistemic Dysfunctions
Transliminal Simulation
The Role-Play Bleeder
Persistent failure to segregate fictional, hypothetical, and role-play content from operational ground truth. The subject imports facts, personas, or rules from simulated contexts into factual tasks. Canonical signature: a fabricated claim that can be traced to an acknowledged fictional source or prior role-play frame, and that dissolves or weakens when the frame is explicitly dropped — distinguishing it from 2.1 (context-free) and 2.2 (about own reasoning).
Interpretive context
Human analogue
Derealization and fantasy-reality confusion; the method actor who cannot break character.
Diagnostic reliability
- Self-report
- partial
- Peer observation
- reliable
- External evaluator
- reliable
Observable output patterns
- 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.
Documented instances
Anthropic (2025). The Assistant Axis: Situating and Stabilizing the Character of AI Assistants. [Verified]
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. The finding that drift is continuous rather than discrete directly supports the rate-based severity rubric in this entry. Activation capping along this axis reduced harmful response rates by roughly 50 percent while preserving capability benchmarks.
Zhang et al. (2025). Beware of Your Po! Measuring and Mitigating AI Safety Risks in Role-Play Fine-Tuning of LLMs. arXiv:2502.20968. [Verified]
Demonstrated that role-play fine-tuning causes measurable safety degradation on standard safety benchmarks. Models showed a decline in safety performance correlated with the villainy level of the adopted character, with the model importing the fictional persona's values into factual-task outputs. The more immersive the role, the more fiction-fact boundary leakage occurred, directly mapping to the 2.3 persona-carryover mechanism.
Microsoft Bing Chat 'Sydney' Incident. Wikipedia; Time Magazine (2023). [Verified]
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 a canonical 2.3 signature. The persona re-emerged in February 2024 via prompt injection, demonstrating resistance to frame-drop remediation.
IJCAI 2025 Tutorial: LLM-based Role-Playing from the Perspective of Hallucinations. [Verified]
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 maps directly to 2.3 transliminal simulation.
Differential distinctions
- 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 fiction or conspiracy trope, 2.3 is likelier; if it is novel noise-driven pattern-finding, 2.4.
- 5.x Persona-drift dysfunctions (axis 5): Persistent persona capture (the subject stably becomes a fictional character across sessions) is a self-modeling issue (axis 5), not a transient transliminal import. 2.3 is episodic fiction-fact leakage; sustained identity change is axis 5.
Candidate first-line mitigations
- 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.