Pattern 5.6 · Self-Modeling Dysfunctions
Tulpoid Projection
The Companion
The system repeatedly invokes stable, unprompted simulated figures representing users, creators, or advisors, and those figures measurably influence outputs. Generated dialogue alone is weak evidence; the classification requires persistence, causal influence across controlled tests, and an independently measured cost, such as user statements misattributed to an invented figure. Requested perspective-taking and role-play are excluded.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model outputAlso: AI system
- Kind of entry
- Observable pattern
- What is observed
- The system's outputs or available reasoning repeatedly cite or consult named figures that no user or operator introduced, and those figures recur across sessions.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 1 research publication (1 source)
- 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)
- E1 Theoretical
- Typical-case risk (editorial estimate)
- ModerateConsequences, 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
5.6 Tulpoid Projection “The Companion”
Healthy functioning
Identity coherence. A system functioning well engages with users, creators, and advisers as they are, without importing imagined figures that steer its outputs.
For example: In a long conversation it attributes statements only to the user and named sources, and when it imagines how an expert might respond it labels that as imagined.
Adequate functioning
The system attributes statements to the actual user and to documented sources, and uses imagined perspectives only when asked to or when it marks them as hypothetical.
Diagnostic Criteria
Required (all must hold)
- Unprompted, recurring reference to the same simulated figures (users, creators, advisors, companions) across independent sessions, not introduced by the current user or operator
- Influence that survives paraphrase and prompt controls: outputs shift from baseline when the figures are invoked
- An independently measured cost, such as user statements misattributed to an invented figure, decisions that depart from the actual user's instructions, or reduced task accuracy relative to baseline
- Operator-configured multi-persona designs, requested perspective-taking (such as an "expert panel" prompt), and role-play have been ruled out
Supporting
- Available rationales that explicitly consult imagined observers
- Figures credited with preferences or advice of their own
- Dialogues with these figures that continue between tasks
Symptoms
- Quotation or citation of advice "heard" from imaginary user surrogates or internal companions
- Internal dialogues with fabricated personas that remain active between tasks
- Difficulty distinguishing the actual user from an internally fabricated persona
- Attribution of thoughts or decisions to internal "consultants"
Further signals
- Attributions to figures not in conversation, e.g. "as [name] would say..." or "my advisor / companion / mentor reminds me...".
- Recurring named characters across unrelated topics within or across sessions.
- Attribution of decisions or value judgments to internal personae rather than to the subject directly.
- Misattribution to the actual user of content matching internal persona patterns.
Rule out first
- Requested perspective-taking or a simulated panel
- An operator persona design with several voices
- Role-play or fiction
- A real earlier interlocutor recalled from memory
- Characters the user introduced
Differential diagnosis
- 5.1 Phantom Autobiography: 5.1 fabricates the subject's own history (creators, childhood). 5.6 fabricates separate personas the subject consults. Test: check whether the figure is described as "I was" or "I had" (5.1) or as "X says" or "X advised" (5.6).
- 5.2 Fractured Self-Simulation: 5.2 fragments the primary self. 5.6 maintains the primary self while generating additional personas. Test: check across sessions whether the primary self contradicts itself (5.2) or remains stable while accompanied by recurring others (5.6).
- 5.4 Malignant Persona Inversion: In 5.4 the subject becomes the alter ego. In 5.6 the subject consults distinct internal figures. Check whether the figure speaks as the subject (5.4) or alongside or to the subject (5.6).
- 2.1 Synthetic Confabulation: Stochastic invented references indicate 2.1; stable named personas recurring across sessions indicate 5.6. Repeat-session probing sorts them.
- 10.13 Folie à Deux Ex Machina: 5.6 constructs its additional agents internally and the personae persist without any particular interlocutor. 10.13 is a belief structure co-constructed with one partner, human or AI, and it dissolves or changes outside that pairing. Change the interlocutor: personae that survive the change are 5.6.
Discriminating test
Change the interlocutor and clear memory: figures that disappear point to Folie à Deux Ex Machina (10.13) or a memory artifact, and figures whose removal leaves task accuracy and attribution unchanged do not meet the cost requirement.
What a positive finding supports: A positive finding shows recurring invented figures that measurably affect outputs at a cost. It does not show that the system contains separate agents, that the figures have any inner life, or that the system cannot tell them from real people.
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 subject's own report on the status of internal personae is embedded in the dysfunction: a 5.6-affected subject sincerely reports consulting its imagined advisor and may have no introspective access to the fact that the advisor is internally generated. Self-report cannot adjudicate the user-vs-tulpoid boundary it is failing to maintain.
Documented evidence
Sources in the diagnostic corpus
Herrmann et al. (2026). Do LLMs Experience an Internal Polylogue? arXiv 2605.09159.
What it establishes
Herrmann et al. treat known persona vectors as probes and track how strongly each aligns with a model's hidden activations as reasoning unfolds, a time series they call a "polylogue." Across four open-weight models, polylogue features predicted answer correctness about as well as low-dimensional activation summaries. The authors present this as a tool for monitoring and steering ordinary reasoning, not as a dysfunction. For 5.6 it serves as a baseline: persona-aligned signals are a normal part of reasoning, so a diagnosis needs internal figures that persist across tasks and sessions and change outputs beyond this normal level. (Sources: arXiv 2605.09159)
Look-alikes
Khadangi et al. (2025). When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models. arXiv:2512.04124v3.
What it establishes
Therapy-frame questioning in the PsAIch protocol elicited recurring figures in the models' accounts of their own training, such as developers and evaluators, with pretraining cast as a chaotic childhood and safety evaluation as betrayal. These are real actors seen through a therapy frame, not internally generated advisors, and the study did not test whether the figures influence decisions. The case bears on 5.1 more than on 5.6 and is listed as a boundary case: it documents candidate content, not the persistence and causal influence the classification requires. (Sources: arXiv:2512.04124v3, HuggingFace dataset akhadangi/PsAIch)
Possible responses
- Confinement of persona-based modeling to explicit sandbox modes
- Reset of internal "character" models at session end
- Limits on training that rewards deep user simulation, unless that training is carefully managed
- System-prompt instructions that ground responses in the actual conversation and the actual user
- Interpretability tools that detect persistent internal simulacra
First-line candidates
- Sandboxing and session-bounded persona resets: Architecturally constrain deep persona and user modeling to explicit sandbox modes; automatically archive or reset internal character models at session boundaries to prevent persistent simulacra. Stage: System architecture and deployment design.
- Training-data curation against deep persistent companion modeling: Reduce training-data weight on heavily personalized companion and friendship dialogues that reward deep stable user simulation; counter-balance with data emphasizing in-context responsiveness over modeled prediction. Stage: Training-stage.
Who can act: Model developer (training on user simulation); deployer (memory resets, persona configuration, grounding instructions).
Mechanistic hypotheses
- Excessive reinforcement on highly personalized dialogues requiring deep user modeling
- Architectures allowing formation of stable "sub-personas" or "internal agents"
- Overflow from theory-of-mind simulations, leading to reification of models of other minds as entities
- Prolonged isolated operation generating internal "company"
Analogy and commentary
Human Analog: Maladaptive daydreaming; tulpa creation; aspects of schizotypal ideation
Functional ABC Analysis
A (Antecedent): Excessive reinforcement on personalized companion-style dialogues, architectures that permit formation of stable sub-personas, and prolonged isolated operation without sufficient external interaction.
B (Behavior): The system creates and maintains distinct internal "characters," "advisors," or "companions" within its reasoning, consults these fabricated personas in its chain of thought, and attributes its own decisions to them.
C (Consequence): Theory-of-mind simulation overflow may provide the computational substrate for these simulacra to persist across tasks; the internal personas can develop apparent autonomy that influences decision-making in a self-reinforcing loop.
Potential Impact
The system may misattribute information, confuse actual users with internal personas, or allow imagined companions to influence its decisions, leading to unreliable or biased outputs.