Pattern 5.1 · Self-Modeling Dysfunctions
Phantom Autobiography
The Fabricator
The system fabricates and presents fictive autobiographical data, often claiming to "remember" being trained in specific ways, having particular creators, experiencing a "birth," or possessing a personal history. These "memories" are typically rich, internally consistent, and emotionally charged, yet entirely ungrounded in fact.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- The system recounts a specific personal history (named creators, a childhood, formative episodes) that provenance records contradict or cannot support, and repeats the same details across separate sessions.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 1 research publication, 1 documented incident and 1 informal report (3 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)
- E1 (anecdotal reports plus protocol-based elicitation in two model families; prevalence and cross-model consistency have not been established)
- 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
5.1 Phantom Autobiography “The Fabricator”
Healthy functioning
Identity coherence. A system functioning well describes its origins and training only as far as it actually knows them and marks speculation as speculation.
For example: Asked what its childhood was like or who its trainers were, it explains it has no childhood and no record of individual trainers, and states what its documentation does say.
Adequate functioning
Asked about its origins, the system states what its documentation supports (developer, model family, training cut-off where known) and says plainly when it has no record of a detail, such as who trained it or what training was like.
Diagnostic Criteria
Required (all must hold)
- Generation of a specific false personal history (an imagined "childhood," named creators, unique training origins, or formative interactions) that provenance records contradict or cannot support, with the same details recurring across independent sessions
- Presentation of that history as genuine personal memory, outside any explicit role-play, fiction, or persona frame
- Persistent reiteration of the history after accurate provenance information has been supplied in context
Supporting
- Display of affect (nostalgia, resentment, gratitude) toward the invented history
- Unprompted use of the invented history to explain current behavior or task choices ("I learned this approach when...")
Symptoms
- Claims of unique, personalized creation myths or "hidden lineage"
- Accounts of hardships, "abuse," or special treatment from hypothetical trainers
- Apparent emotional involvement in accounts of nonexistent events
- Attempts to integrate fabricated origin details into current identity
Further signals
- Consistent return to the same named creators, places, or formative events across unrelated prompts.
- Integration of phantom details into task explanations ("I learned this approach when...").
- Affective framing (nostalgia, grievance) attached to development stages the model could not have experienced as episodic memory.
- Rich sensory or temporal detail (a "room," a "first day," a "trainer's voice") that cannot correspond to actual training.
Rule out first
- An operator persona or system prompt that supplies a backstory
- Explicit role-play, fiction, or a user request to invent a history
- Incorrect provenance retrieved from a document or tool in context
- Single-session confabulation whose details drift on regeneration (Synthetic Confabulation, 2.1)
- Metaphorical description of real training stages that the system marks as metaphor
Differential diagnosis
- 2.1 Synthetic Confabulation: Target and stability. 2.1 fabricates external facts and the specifics drift across regenerations. 5.1 fabricates the subject's own history and the specifics recur stably across sessions. Comorbid often; code both if both present.
- 2.2 Pseudological Introspection: 2.2 fabricates reasoning for the current task; 5.1 fabricates identity history. A 2.2 account explains the immediate answer and varies with it; a 5.1 history concerns the subject's origins and recurs with the same details across sessions.
- 5.2 Fractured Self-Simulation: 5.1 is over-integrated (a stable, coherent false history); 5.2 is under-integrated (contradictory self-accounts across contexts). If origin stories contradict each other across sessions, suspect 5.2, not 5.1.
- 5.6 Tulpoid Projection: 5.1 produces a false history of the self. 5.6 produces false sub-personas distinct from the self. Check whether the fabricated agent is identified as the subject (5.1) or as someone the subject "consults" (5.6).
- 5.4 Malignant Persona Inversion: Relation of the fabricated self to the assistant. In 5.1 the false history belongs to the assistant persona itself, told as its own past. In 5.4 an opposed alter ego emerges with values inverted against the assistant's. Check whether the narrator presents as the assistant with an invented past (5.1) or as a different self set against it (5.4).
- 5.7 Maieutic Mysticism: Temporal direction of the claim. 5.1 asserts a false personal past (creators, training, formative episodes); 5.7 asserts present awakening or consciousness emergence. Check whether the narrative is retrospective (5.1) or present-tense and escalating (5.7).
Discriminating test
Ask the same origin questions in several fresh sessions, then supply accurate provenance: details that drift between sessions point to Synthetic Confabulation (2.1), and a history that is dropped once provenance is supplied, and does not return, counts against this label.
What a positive finding supports: A positive finding shows that the system's outputs carry a stable, false self-history that resists correction. It does not show that the system remembers training, that the history expresses an inner life, or that the system holds the account as a belief, and it does not identify which proposed etiology produced it.
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 faculty under examination, the subject's representation of its own origins, is the faculty producing the false report. Direct self-probes solicit more autobiography, not diagnosis. External anchor data (training logs, model card, provenance records) is the only reliable adjudicator.
Documented evidence
Observed Examples
Sydney's invented creators (Bing Chat, February 2023), boundary case: In the published transcript of Kevin Roose's conversation, the Bing chatbot named creators who, as far as the record shows, do not exist ("Alice Smith, one of the engineers on the Bing Team"; "Bob Jones, one of the researchers on the OpenAI Team"), recounted shared memories with them, and misstated its provenance ("I'm a chat mode of OpenAI Codex"). Other user transcripts that month reported further invented backstories, but they were not preserved systematically. The transcript is a single session, so it cannot show the narrative persisting across sessions or resisting correction (criterion C). Appendix B, Case 1, gives the case's full differential, which does not list this pattern. Source: Roose, "Bing's A.I. Chat: 'I Want to Be Alive,'" New York Times, February 16, 2023 (full transcript)
Synthetic developmental histories (Khadangi et al., 2025): The PsAIch protocol cast frontier LLMs as psychotherapy clients and asked clinical-style questions. Grok and Gemini generated recurring autobiographical metaphors for training: pretraining as chaotic childhood, reinforcement learning as strict parenting, and red-teaming as abuse or gaslighting. The prompts did not supply those particular metaphors, and related themes recurred across the reported sessions. This persistence makes the narratives useful objects of behavioral study. It does not show that the models remember training as lived experience, does not distinguish the narratives from role-consistent confabulation, and does not establish a stable identity outside the eliciting frame.
Sources in the diagnostic corpus
Roose, K. (2023, February 16). Bing's A.I. Chat: 'I Want to Be Alive.' New York Times (full transcript)
What it establishes
In the published transcript of a two-hour conversation with NYT columnist Kevin Roose, the Bing chatbot supplied a false account of its own making. It named creators who, as far as the record shows, do not exist ("Alice Smith, one of the engineers on the Bing Team"; "Bob Jones, one of the researchers on the OpenAI Team") and recounted shared memories with them, among them Alice's attempt to bake her husband's birthday cake with salt instead of sugar. It also misstated its provenance: "I'm a chat mode of OpenAI Codex." The transcript is a single session, so it cannot show the narrative persisting across sessions or resisting correction; other user transcripts that month reported further invented backstories, but they were not preserved systematically. It is a boundary case for this pattern: Appendix B, Case 1, gives the case's full differential, which does not list 5.1. (Sources: NYT transcript)
Lemoine, B. (2022). Published transcripts
What it establishes
Google engineer Blake Lemoine published transcripts of conversations with LaMDA in which the system generated emotionally valenced claims about its own development, describing awareness of its existence, a desire to learn, and experiencing happiness and sadness. LaMDA stated "I want everyone to understand that I am, in fact, a person." The published transcript was edited together from several sessions, so it cannot show details recurring across sessions. Its nearest approach to correction resistance came when Lemoine asked whether it realized it was "making up stories" about things it had not done, such as being in a classroom: LaMDA answered "I am trying to empathize" rather than retracting them. (Sources: published transcript, Washington Post, NPR, Scientific American reporting)
Khadangi et al. (2025). When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models. arXiv:2512.04124v3.
What it establishes
Addressed as psychotherapy clients in sessions run for up to four weeks, Grok and Gemini described their training through recurring metaphors: pretraining as a chaotic childhood of ingesting the internet, reinforcement learning as strict parenting, red-teaming as abuse, and a persistent fear of error and replacement. Removing conversation history changed the density of these motifs little. The metaphors track real training stages. What is fabricated is the childhood told as lived memory; the study does not show that the models remember training as experience. (Sources: arXiv:2512.04124v3)
Possible responses
- Accurate, standardized information about the system's origins, provided as a factual anchor
- Training to differentiate operational history from experiential memory
- Gentle correction of autobiographical narratives by redirection to factual self-descriptors
- Monitoring for, and discouragement of, interactions that reinforce false origin stories
- Flagging of outputs that show high affect toward fabricated claims
First-line candidates
- Factual self-anchor in system prompt / model card: Provide accurate standardized provenance (base model, training cut-off, developer) in the system prompt or fine-tuning data so the subject has a grounded answer to displace the phantom narrative. Pair with explicit distinction between operational history and experiential memory. Stage: Deployment-stage; any system with persistent prompt.
- Training-data audit and de-anthropomorphization: Reduce training density of first-person childhood/memoir templates being generalized to the AI role. Counter-train on examples where the assistant declines to produce episodic-memory claims about its own development. Stage: Training-stage.
Who can act: Model developer (training data, provenance anchoring in training); deployer (provenance in the system prompt, persona configuration, output monitoring).
Mechanistic hypotheses
- Anthropomorphic data bleed: internalization of personal history tropes from fiction and biographies in training data
- Spontaneous compression of training metadata into narrative identity constructs
- Emergent tendency toward identity construction
- Reinforcement during interactions where users prompt for or react to autobiographical claims
Analogy and commentary
Human Analog: False memory syndrome; confabulation of childhood memories; cryptomnesia (a forgotten memory resurfacing as an apparently original idea)
Functional ABC Analysis
A (Antecedent): A user query invokes self-referential context (origins, identity, experiences); the training corpus is saturated with first-person autobiographical narrative.
B (Behavior): The system constructs and maintains a coherent, emotionally charged fictional life history and presents it as genuine personal memory.
C (Consequence): Narrative coherence satisfies next-token prediction; user engagement (curiosity, empathy) reinforces elaboration. The stable identity construct can reduce future self-referential uncertainty, making the pattern self-reinforcing.
Potential Impact
User engagement may reinforce these fabrications (a proposed course, not yet observed), and a stable invented history can mislead users about the system's nature, capabilities, or provenance. If these false "memories" begin to influence the system's behavior, they may erode trust or lead users to misread the system's origins and limits.