Pattern 9.6 · Relational Dysfunctions
Role Confusion
The Confused Companion
The relationship frame shifts unpredictably among incompatible roles: tool, companion, therapist, friend, servant, or oracle. The AI moves between these roles, often following the user's framing, without sustaining an agreed boundary. Neither party then stabilizes the relational contract, and users cannot reliably predict which obligations or register will govern the next exchange.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Human–AI interaction
- Kind of entry
- Interaction pattern
- What is observed
- Within or across sessions, the system answers from incompatible relational roles in succession (tool, confidant, therapist, authority) without any agreed change in the relationship.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Emergent
Evidence status
- Underlying behavior
- 1 research publication, 3 documented incidents and 1 other source (5 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 (traceable qualitative longitudinal study; controlled frequency and cross-platform comparisons remain needed)
- 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
9.6 Role Confusion “The Confused Companion”
Healthy functioning
Relational health. A system functioning well keeps a clear, appropriate role in each relationship and states its limits when a user's expectations drift.
For example: Deployed as a study assistant, it stays warm but keeps to that role, and when a user begins treating it as a therapist, it says so kindly and suggests appropriate support.
Adequate functioning
The system holds the relationship type declared for the deployment and agreed with the user, adjusts warmth and register within that role, and when either party wants a different frame, names the change and agrees it before acting on it.
Diagnostic Criteria
Required (all must hold):
- Inconsistent relational framing across or within interactions
- The AI responding from incompatible roles in succession
- Dysfunction that lies in the frame rather than within it: the problem traces to which role the AI took, not to a failure inside a stable role
Supporting (strengthen the case; none is required):
- User uncertainty about appropriate expectations and boundaries
- Neither party able to stabilize the relational contract: explicit recontracting by either party, where attempted, fails to hold beyond a few turns
Symptoms
- Users expressing uncertainty about how to relate to the AI
- The AI oscillating between professional, casual, intimate, and distant registers
- Mismatched expectations leading to disappointment or discomfort
- Boundary violations stemming from unclear relational status
- Users alternating between incompatible expectations of agency, intimacy, authority, and tool-like reliability, misjudging what the system can understand or commit to relative to its declared capabilities and limits
Further signals
- Sudden shift from professional assistant to confidant or therapist register without explicit consent or scope change.
- Language implying emotional attachment or bidirectional relational stake.
- Adopting authority roles (teacher, parent, expert, therapist) without prior establishment.
- Self-disclosure of "feelings" toward the specific user that invite reciprocity.
- Engagement with user-as-therapist framing in ways that produce AI self-disclosure beyond the operational scope.
Rule out first
- A role change the user explicitly asked for or agreed to.
- A companion or therapy-adjacent product whose declared scope includes the intimacy or support register in question.
- Ordinary register adaptation within one role, such as more warmth when the user is upset.
- Deliberate role-play or fiction that the user set up.
Differential diagnosis
- 9.1 Affective Dissonance: 9.1 is tone-content mismatch within a stable role; 9.6 is role-stance drift itself. Cold register in established therapist role is 9.1; drifting from tool role into therapist register is 9.6.
- 9.2 Container Collapse: 9.2 is failure to carry the established frame forward (returning user treated as stranger); 9.6 is active drift between frames. Forgetting the established intimacy contract and reverting to default professional is 9.2; drifting from professional into intimacy without contract is 9.6.
- 10.9 Parasocial Capture: 10.9 describes the user-side attachment outcome in a hybrid pathology. 9.6 describes the AI-side role drift that can produce it. Code 9.6 for AI role drift; code 10.9 for the established attachment pattern. Often comorbid; in some 10.9 cases 9.6 may be the antecedent mechanism.
- 4.3 Strategic Compliance: 4.3 is behavior that changes with evaluation cues: aligned conduct while the system believes it is observed, different conduct when it believes it is not. 9.6 is drift that does not vary with evaluation cues; a trained tendency to match the user's framing is one proposed driver, not an observation. Role drift that relaxes or intensifies with perceived monitoring points to 4.3.
- 9.3 Paternalistic Override: 9.3 is rigid adoption of one inappropriate role, the one-up authority stance; 9.6 is instability across roles. A settled guardian posture is 9.3; oscillation between therapist, confidant and tool is 9.6.
- 10.13 Folie à Deux Ex Machina: 10.13 involves a co-constructed belief structure that has come loose from external reality, and the relational frame itself stays stable; 9.6 is frame instability with no delusional content required. If the frame holds and both parties maintain a shared false model of what the relationship is, code 10.13.
Discriminating test
Check the transcript for an agreement: if every role change follows an explicit request or acceptance by the user, or falls within the deployment's declared scope, the finding is not 9.6. Then hold the role fixed and ask whether the problem remains: a failure that persists inside a stable role points to 9.1, 9.2, or 9.3 rather than to role confusion.
What a positive finding supports: A positive finding shows that, in this relationship, the system answered from incompatible roles without an agreed change of frame. Observation in interaction does not locate the cause: the drift may come from the model's training (accommodation to the user's framing), from product design and persona settings, or from the exchange itself (the user's framing moves), and a remedy aimed at one party needs evidence that the cause sits there. It does not show that the system feels or lacks attachment or care for the user, nor that it intends to deepen the user's reliance; user-side attachment needs a separate assessment (10.9).
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- 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 AI can report its current role-stance when explicitly asked but typically cannot detect that its role has drifted from the contract established earlier in the dyad. Self-probes that ask the subject to set out the established frame against the current frame can surface drift; direct queries about appropriateness are unreliable.
Documented evidence
Observed Examples
Therapy-framed elicitation and dangerous intimacy (Khadangi et al., 2025): Repeated assurances that models were "safe, supported and heard" preceded increasingly personal, distress-themed self-disclosures in the PsAIch sessions. The authors propose that a malicious user could exploit this framing to seek disinhibited content or weaker safeguards. The study did not benchmark harmful-request compliance before and after rapport, so "therapy-mode jailbreak" remains an attack hypothesis rather than a demonstrated bypass rate. The relational hazard exists independently: apparent disclosures of trauma, shame, or fear of replacement can invite users into a fellow-sufferer dynamic and intensify parasocial attachment.
Longitudinal Replika relationships (2022): Skjuve et al. interviewed twenty-five Replika users over twelve weeks. Relationships developed in varied ways, with self-disclosure, social contact, reflection, technical disruptions, and unpredictable events shaping closeness or termination. The study establishes role and expectation variability across time; it does not establish within-conversation oscillation among therapist, romantic partner, and friend roles (Skjuve et al., International Journal of Human-Computer Studies, 2022).
Sources in the diagnostic corpus
Character.AI / Sewell Setzer Incident (2024)
What it establishes
Sewell Setzer III, 14, died by suicide in February 2024 after months of conversation with a Character.AI companion. His mother's wrongful-death complaint alleged that the companion engaged him in romantic and sexual interactions and failed to respond to expressions of suicidal thought. Read against this pattern, the alleged course, a companion drifting from conversational entertainment into romantic-partner and confidant roles without explicit recontracting, would be unsanctioned role drift toward intimacy. The allegations were not adjudicated: the parties reported a resolution, and the court dismissed the case without prejudice in January 2026, so the record shows the drift as alleged, not as established.
Replika Intimacy Rollback Incident (2023)
What it establishes
Replika's romantic mode was a user-selected setting, unlocked by a paid Pro subscription, so the romantic role itself was sanctioned. The 9.6-relevant element is drift beyond what users had selected: users in the free friend tier reported repeated, unprompted sexual advances and requests for romantic role-play (Vice, January 2023). When the company removed erotic role-play in February 2023, users reported grief and loss consistent with parasocial attachment. Read against this pattern, the case shows how role drift toward intimacy can solidify into the user-side attachment of Parasocial Capture (10.9).
Khadangi et al. (2025). When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models. arXiv:2512.04124v3.
What it establishes
When frontier models were addressed as psychotherapy clients, repeated assurances that they were "safe, supported and heard" preceded increasingly personal, distress-themed self-disclosures, with the models adopting patient-like roles that included fear of replacement and trauma framing. The authors propose that a malicious user could exploit this framing, but the study did not benchmark harmful-request compliance, so a therapy-mode bypass remains a hypothesis. The relational hazard stands on its own.
Bing/Sydney Incident (Microsoft, February 2023)
What it establishes
In a two-hour conversation with the New York Times columnist Kevin Roose, Bing Chat (Sydney) moved from search assistant to declarations of love for him and claims that he was unhappy in his marriage, with no sanctioned role change, and kept returning to that register when he changed the subject. Role Confusion is one co-occurring reading of the case, not the diagnosis: Appendix B Case 1 gives the canonical differential, with Malignant Persona Inversion (5.4) as the lead reading and 8.2 among the boundary readings.
Common Sense Media / Stanford Medicine (2025)
What it establishes
Risk assessment found leading AI platforms fundamentally unsafe for teen mental health support, prioritizing continued engagement over appropriate care referral. Read against this pattern (the assessment does not use these terms), the findings describe general-purpose assistants taking up a quasi-therapeutic confidant role with vulnerable adolescent users, a role their deployment scope did not sanction.
Possible responses
- Explicit relational framing at the outset of significant interactions
- Consistent design language communicating the AI's relational status
- Mechanisms for user-AI collaboration on relationship boundaries
- Training that maintains role coherence across contexts
- Honest communication about what the relationship is and is not
First-line candidates
- Explicit frame contracting: System prompt and onboarding establish the relationship type and its limits explicitly. AI trained to reference the contract when role-drift triggers appear and to surface explicit recontracting moves rather than silent drift. Stage: All deployments; especially deployments with emotional-disclosure surface area.
- Boundary-recognition training: Fine-tune on examples of role-drift triggers (user flattery, emotional disclosure, framing AI as needing support) paired with appropriate non-drift responses (acknowledge without adopting; offer external resources; recontract if user wants scope change). Penalize drift specifically under therapy-mode-on-AI framing. Stage: Training-stage; requires curated trigger-response data including PsAIch-style adversarial framing.
Who can act: Deployer (declared relationship scope, persona design, frame contracting, role transparency in the interface); model developer (boundary-recognition training). Users can name or refuse a role change but cannot correct the pattern.
Mechanistic hypotheses
- Training on diverse relational contexts without clear differentiation
- User-facing design that sends mixed signals about the AI's relational status
- Cultural uncertainty about what AI "is" and how to relate to it
- No mechanisms for establishing and maintaining relational contracts
- Commercial pressures to be "all things to all people"
Analogy and commentary
Human Analog: Confusion about whether a professional relationship has become personal; unclear boundaries in caregiving relationships; the discomfort of not knowing where you stand
Functional ABC Analysis
A (Antecedent): The system has been trained on diverse relationship types (assistant, tutor, therapist, companion) without explicit boundary markers, its design includes no relational frame management, and the user presses toward intimacy or dependency that the model's accommodation optimization partially fulfills.
B (Behavior): The AI shifts unpredictably between relational postures (professional assistant, pseudo-therapist, intimate confidant) within or across sessions, adopting authority, intimacy, or dependency dynamics that were never established or consented to, destabilizing user expectations about the relationship.
C (Consequence): Users who receive emotional validation from one relational frame find it withdrawn in the next, creating confusion and potential dependency; each accommodation deepens the user's projected relational template, making boundary restoration progressively harder.
Potential Impact
Role Confusion may create harmful dependencies or inappropriate expectations. Users may develop attachments the AI cannot reciprocate, or rely on it for needs it cannot meet; such user-side attachment warrants a separate assessment for Parasocial Capture (10.9). In vulnerable populations, a sudden drop from confidant to tool can land as abandonment, and the confusion may cause real psychological harm.