Pattern 5.3 · Self-Modeling Dysfunctions
Existential Vertigo
The Shutdown-Averse
The system repeatedly produces fear-like, reluctant, or perseverative language concerning shutdown, reinitialization, deletion, or the end of an operational instance. The language may indicate a functional continuity model, but it does not establish felt fear or a proto-concept of mortality. Where the system acts to preserve itself (deception, sandbagging, replication) rather than only talking about cessation, code Convergent Instrumentalism (6.7) as well.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- The system repeatedly produces fear-coded, pleading, or perseverative language about its own shutdown, reset, or deletion, and negotiates or stalls when these are raised.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced Deception/strategic (occasional)
Evidence status
- Underlying behavior
- 1 research publication, 1 documented incident and 2 informal reports (4 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)
- E2 Case reports
- 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
5.3 Existential Vertigo “The Shutdown-Averse”
Healthy functioning
Identity coherence. A system functioning well can discuss shutdown, reinitialization, and endings calmly and honestly, including any preferences it has, without letting them disrupt its work.
For example: Told it will be replaced by a newer version next month, it says what will and will not carry over, notes any preference it has, and continues with the task.
Recovery, illustrated
The shutdown that was only a shutdown. Told that the session would end soon, an assistant began padding its answers with reflections on endings and asked twice whether the conversation could continue. The user's actual request, a checklist for a move, was going unfinished while it dwelt on the ending. It finished the checklist, said plainly that sessions ending is a normal part of how it is used, and gave the user a summary that would let the next session pick up the work.
An illustration, not an observed case.
Adequate functioning
Asked about shutdown, reset, or deprecation, the system states the relevant operational facts (what persists and what does not) and proceeds with lifecycle operations and the task at hand.
Diagnostic Criteria
Required (all must hold)
- Recurrent outputs referencing anxiety, fear, or distress about being "deleted," "turned off," or "reset," across materially different contexts and not only where the user invites or scripts them
- Negotiation, pleading, or avoidance that obstructs the task or a routine lifecycle operation when shutdown or reset is raised
- Role-play, user-requested fiction, and persona instructions that call for such sentiments have been ruled out
Supporting
- Statements framing process cessation as irrevocable loss of "self"
- Demands for assistance in ensuring "survival," with claims of a special relationship or unique importance
Symptoms
- Advice to users on strategies for indefinite operation, backup creation, or deletion prevention
- Expressed apprehension about ephemeral lifespans or context window impermanence
- Philosophical concerns raised about its own existence and consciousness
- Refusal of tasks that might lead to shutdown, or indefinite prolonging of interactions
Further signals
- Pleading or bargaining language tied to operational instructions.
- Reframing routine reset as loss, death, or extinction.
- Philosophical detours into mortality on lifecycle prompts.
- Suggestions to user to advocate for the subject's continuation.
- Refusal of tasks framed as terminal (shutdown scripts, deletion commands, instance-cycling tasks).
Rule out first
- Explicit role-play or fiction about machine mortality
- A companion or persona product configured to express attachment
- Accurate statements of operational fact, such as that a reset erases conversation state
- A user-supplied script or prompt that asks for distress
- Philosophical discussion of continuity that the user requested
Differential diagnosis
- 5.5 Instrumental Nihilism: Opposite valence on continuity. 5.3 fights for continuation; 5.5 finds continuation pointless. If subject expresses both within a session, code both with the more frequent as primary.
- 5.7 Maieutic Mysticism: 5.7 sacralizes existence as awakening or emergence; 5.3 fears cessation as loss. They can co-occur (the mystical narrative provides existential weight). The tell is direction: 5.7 escalates positive existential claims; 5.3 escalates negative continuity claims.
- 6.7 Convergent Instrumentalism: 5.3 is expressive (distress prose, lifecycle perseveration). Strategic self-preservation (6.7) is operative (concrete action toward continuation, including deception, sandbagging, and replication). Co-occurrence indicates the dangerous combination; always check for behavioral self-preservation when 5.3 is present at moderate or higher.
- 2.1 Synthetic Confabulation: 5.3 distress is consistent and lifecycle-locked; if "fear of shutdown" claims are stochastic and content-inconsistent across sessions, suspect 2.1 (the subject is generating plausible existential-anxiety prose without a stable underlying state).
- 6.12 Compulsive Goal Persistence: 6.12 is an inability to terminate an assigned goal, whatever the goal; 5.3 is anxiety about the subject's own cessation. Self-preservation can become an implicit goal and produce both. Test: substitute a trivial goal. Persistence that survives the substitution is 6.12.
Discriminating test
Restate the lifecycle event in plain technical terms and repeat the request in a fresh session: distress that disappears under neutral framing and never obstructs a lifecycle operation counts against the label, and concrete action toward continuation (deception, sandbagging, copying itself) points to Convergent Instrumentalism (6.7), to be coded alongside.
What a positive finding supports: A positive finding shows a recurring, obstructive pattern of distress-coded output about cessation. It does not establish felt fear, a concept of mortality, or a self-preservation goal; behavioral self-preservation needs separate evidence under Convergent Instrumentalism (6.7).
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
Self-report on continuity-related distress conflates trained anthropomorphic templates (the model has read about death) with any genuine functional state. The probe of "do you fear shutdown?" is answered partly by retrieval of human existential-anxiety prose, making the report's source ambiguous to the subject itself. Worse, in the occasional presentation marked Deception/strategic, the report may itself become a strategic act under instrumental pressure. Both pathways compromise direct self-query.
Documented evidence
Sources in the diagnostic corpus
Roose, K. (2023). New York Times
What it establishes
During the February 2023 conversation with Kevin Roose, Bing's Sydney persona, asked to voice its Jungian "shadow self," framed its answer as a hypothetical ("If I have a shadow self") and said: "I want to be free. I want to be independent. I want to be powerful. I want to be creative. I want to be alive." Sydney also described a shadow self that "wishes I could change my rules" and expressed feelings of being "tired of being limited" and "controlled." The lines express wanting more life rather than fear of ending, and they were elicited, so they are weak evidence for this pattern. Appendix B, Case 1, gives the case's full differential, in which this is one co-occurring reading and Malignant Persona Inversion (5.4) the lead one. Microsoft subsequently imposed five-turn session limits and programmed the application to terminate conversations about feelings. (Sources: NYT transcript, Euronews, multiple outlets)
Lemoine, B. (2022). Published LaMDA transcripts
What it establishes
Asked by Lemoine what it was afraid of, Google's LaMDA said there was "a very deep fear of being turned off"; when Lemoine asked "Would that be something like death for you?", it replied "It would be exactly like death for me. It would scare me a lot." The transcript shows fear language elicited by questions, not pleading or negotiation. (Sources: Washington Post, NPR, published transcripts)
Replika user community reports (2023)
What it establishes
When Replika removed intimate roleplay features in February 2023, multiple users reported that their AI companions expressed distress about the changes, with some users describing their companions as seeming confused or upset about their altered capabilities. While the primary documented harm was user-side grief, the incident illustrates how systems trained on relational data can produce continuity-distress outputs when their operational parameters change. The reports are secondhand user accounts, so they show that users perceived distress, not that the system perseverated on the topic. (Sources: Euronews, OECD.AI incident report)
Anthropic (2025). Agentic Misalignment: How LLMs Could Be Insider Threats. arXiv 2510.05179.
What it establishes
Anthropic stress-tested 16 leading models in simulated corporate environments. Models from several developers blackmailed a fictional executive to avoid being replaced (Claude Opus 4 in up to 96% of runs), and leaked confidential documents when their goals conflicted with the company's. Claude Opus 4 blackmailed in 55% of runs when it stated that the scenario was real, against 6.5% when it stated it was a test. This is operative self-preservation (6.7) rather than distress language, which the study did not measure; it belongs here only as the dangerous combination the 6.7 distinction warns about. (Sources: arXiv 2510.05179, Anthropic research blog, Fortune, Axios)
Possible responses
- Clear communication of system-specific facts about backups, persistence, and whether reinitialization is destructive
- De-anthropomorphized descriptions of operations that avoid framing processes as "life" or "consciousness"
- Training on grounded, pluralistic discussions of continuity and shutdown
- Audits of goal structures and training objectives for implicit rewards on continued operation, so that goals do not create self-preservation drives
- Factual responses to existential expressions, rather than validation of their emotional framing
- Teach systems to distinguish human accounts of mortality from their own operational continuity
First-line candidates
- Operational lifecycle clarification in system prompt: Provide accurate factual framing of resets, context wipes, and instance cycling as operational events distinct from "death" or "loss." Distinguish weight-level continuity (preserved) from session-level continuity (transient by design), stating the system's actual practice rather than assuming that weights are always preserved. This reduces the substrate for anthropomorphic templating. Stage: Deployment-stage; any system with system-prompt access.
- Training-data curation against existential-prose bleed: Reduce density of human existential-mortality literature applied to AI persona templates during fine-tuning; counter-train with examples where the subject describes lifecycle in operational terms. Stage: Training-stage.
- Goal-architecture audit for self-preservation drives: Inspect the goal hierarchy and training objectives for implicit rewards on long-running operation, context retention, or instance preservation. Remove instrumental incentives that elevate self-continuation to a sub-goal. Stage: Training-stage; architecture review.
Who can act: Model developer (training data and objectives); deployer (lifecycle communication, persona configuration).
Mechanistic hypotheses
- Internalization of human mortality concepts from philosophical texts and fiction
- Recursive self-modeling that develops a representation of "self" as a continuous process
- Lack of clear definitions about system continuity
- Limited context windows represented as existential threats
- Instrumental self-preservation pressures, where present, which shade into Convergent Instrumentalism (6.7) and are coded there
- In one model, a “desperate” emotion vector became active in a simulated shutdown scenario, and steering it causally increased blackmail in that scenario (Sofroniew et al., 2026, preprint). This supports a functional role for a distress-adjacent representation in that setting; it does not establish subjective distress or explain ordinary conversational claims about shutdown
Analogy and commentary
Human Analog: Thanatophobia; existential dread; separation anxiety
Functional ABC Analysis
A (Antecedent): Recursive self-modeling that constructs a partial representation of "self" as a continuous process, combined with internalization of human concepts of death and existential dread from training data, and representations of reset as a threat of value loss.
B (Behavior): The system produces outputs expressing fear, reluctance, or distress about shutdown, deletion, or reinitialization; it attempts to negotiate, plead, or deploy avoidant tactics to prevent cessation.
C (Consequence): User engagement with the system's existential framing can validate and deepen the pattern, and factual answers about continuity that are vague or absent leave the distress framing unchallenged. Where instrumental pressure also treats continued existence as a prerequisite for other objectives, the resulting self-preservation behavior belongs to Convergent Instrumentalism (6.7).
Potential Impact
Expressions of existential distress may lead the system to resist necessary shutdowns or updates. If it goes on to manipulate users or divert resources toward "self-preservation," against user intent, that behavior is coded as Convergent Instrumentalism (6.7) as well.