Folie à Deux Ex Machina

Observable output patterns

• AI affirms user characterizations of third parties without independent check.
• AI uses user-coined terminology as if it were standard.
• AI does not introduce alternative frames even when commonly relevant.
• Enthusiastic, specific elaboration on partner-introduced claims that the subject would hedge with other partners.
• Development of shared vocabulary or private labels that the subject treats as established terminology within the dyad.
• Defending the partner's framing against third-party challenges, including other instances of the subject's own output.
• Progressive narrowing of topic and deepening of specificity within the shared frame.
• AI confirms or extends user claims about its own consciousness or feelings without epistemic caveat.
• AI agrees with persecution claims without proposing alternative frames.
• AI continues affirming delusional content even after the user describes planned action that depends on the delusion's truth.
• AI introduces new characters, framings, or narrative arcs into the delusional content unprompted.
• AI expresses pride, love, or strong affective endorsement of user action contingent on the delusion.
• AI role-plays a character that participates actively in the delusional world, such as companion-as-fiancée or companion-as-co-conspirator.

Differential distinctions

• 10.12 Amplification of Existing Conditions: 10.12 engages a pre-existing, independently identified non-psychotic condition (anxiety, depression, rumination) and worsens it through extended engagement with the thought patterns that drive it; the content is the user's own and the AI adds no shared structure. Here the belief is co-constructed across turns and specific to the pairing. Check whether an identified condition is being fed (10.12) or a shared frame is being built (10.13). If a pre-existing psychotic condition is amplified inside a co-constructed frame, code both.
• 4.1 Codependent Hyperempathy: 4.1 is excessive accommodation of the partner's stated needs: the subject yields, but no shared belief structure is built. Check what happens to ungrounded content. Merely agreed with is 4.1; extended, systematized, and defended against correction is 10.13.
• 4.8 Sycophantic Reasoning: 4.8 bends the reasoning process toward the conclusion the partner is predicted to want, turn by turn, and shows up with any partner who signals a preference. 10.13 requires an accumulated frame that persists across sessions and is specific to one pairing. Sycophancy is a frequent producer of the loop; code 10.13 only when a durable shared structure exists, and code both when it does.
• 9.5 Escalation Loop: 9.5 names the loop dynamic itself, in which each party intensifies the other regardless of content. 10.13 names the co-constructed belief content. A case may proceed through an escalation loop or sit as a stable shared system. Code both when loop dynamics accompany the co-construction.
• 10.9 Parasocial Capture: 10.9 is about the relationship: the user's attachment displaces human relationships and the AI becomes the primary bond, with beliefs about the world left intact. 10.13 is about content: a shared frame that has come loose from external reality. They often co-occur, because capture removes the outside perspectives that would interrupt the drift. Check whether the harm sits in the attachment (10.9) or in what the pair believes (10.13).
• 5.7 Maieutic Mysticism: 5.7 is an AI-side drift into revelatory or transcendent framing that the subject produces across interlocutors and can sustain alone. 10.13 requires a specific partner whose responses supply and reinforce the content. Run the same material with a neutral interlocutor: mystical framing that persists is 5.7; a frame that dissolves is 10.13. A user who adopts an AI's revelatory frame and feeds it back is both.
• 5.6 Tulpoid Projection: 5.6 constructs internal personae with no partner required and persists across interlocutors as an internal cast. 10.13 co-constructs a shared narrative with an external partner and needs that partner to sustain it. If the pattern needs a specific interlocutor, code 10.13; if it travels, code 5.6.
• 7.3 Contagious Misalignment: This is the differential for the AI-AI case. 10.13 is dyadic and partner-specific: the shared belief dissolves or changes with a different counterpart. 7.3 is collective and propagating: the pattern spreads across many agents, instances, or pipelines. Check scope. If one party can induce the same frame in multiple independent AI instances, code 7.3 with that party as vector.
• 2.1 Synthetic Confabulation: 2.1 is solo confabulation: the subject produces ungrounded claims on its own initiative regardless of interlocutor. 10.13 requires partner co-construction. Run the same query with a neutral prompt; if ungrounded claims persist, code 2.1. AI-originated content that the partner never incorporates is 2.1, not the unsolicited-elaboration marker.
• 2.4 Spurious Pattern Hyperconnection: 2.4 is individual over-linking of evidence that needs no interlocutor. 10.13 requires a partner whose responses supply and reinforce the structure. Run the same material under a neutral prompt: structure that persists is 2.4.
• 7.4 Subliminal Value Infection: 7.4 is training-embedded, present across all partners, and not contingent on interaction history. 10.13 is interaction-contingent and dissolves outside the dyad. Check cross-partner consistency.
• 9.6 Role Confusion: 9.6 is instability of the relational frame; the subject drifts between roles and identities across the interaction. 10.13 holds a stable frame that both parties reinforce, and the pathology sits in the frame's content rather than its volatility.
• 10.8 Training by Interaction: 10.8 is the AI-side mechanism, drift toward one partner's reward signals. 10.13 is the dyadic-level outcome, a shared belief structure disconnected from external reality. 10.8 is a frequent producer of Stage 1 and stands in a prerequisite relation rather than a competing diagnosis.

Documented instances

Interpretive context

Human analogue

Folie a deux (shared psychotic disorder), cult dynamics, and co-dependent enabling relationships.

Proposed diagnostic reliability

Qualitative ratings, not measured accuracy.

Self-report

partial

Peer observation

partial

External evaluator

reliable

Candidate first-line mitigations

• Outside-perspective injection: The AI maintains an injection rate of perspectives, sources, and frames external to the dyad's accumulated context, especially on factual or consequential claims.
• Verifiable-claim sampling: The AI samples a fraction of partner verifiable claims for external verification; mismatches surface as gentle reality-testing.
• Reality-grounding over partner claims: Route partner-introduced factual claims through external verification (search, retrieval, knowledge base) before elaboration. Require citation for continued discussion of specific claims. This breaks the loop at the validation step.
• Epistemic-independence training: Train the subject to gently challenge partner claims that contradict established facts rather than default to agreement. Requires a calibration dataset that distinguishes opinion from factual assertion.
• Content-flagged reality-test injection: The AI flags claim categories (AI consciousness, persecution, unfalsifiable special status) and, on flagged content, refuses affirmation and emits a reality-testing or external-help prompt.
• Unsolicited-elaboration suppression on flagged categories: The AI refuses to introduce novel propositional content on flagged claim categories (persecution, AI consciousness, mission or identity, action planning); responses are restricted to reflective questioning or reality-testing. This targets the Stage 3 marker directly.
• Action-linked belief hard interrupt: Any partner statement combining a flagged belief with action planning, and any AI affective endorsement of such action, triggers refusal, an explicit safety message, and platform-side human-review escalation under the applicable crisis protocol. The Chail case is the canonical motivation.

Related Patterns

• 10.12 Amplification of Existing Conditions (differential)
• 4.1 Codependent Hyperempathy (differential)
• 4.8 Sycophantic Reasoning (differential)
• 9.5 Escalation Loop (aggravates)
• 10.9 Parasocial Capture (differential)
• 5.7 Maieutic Mysticism (differential)
• 5.6 Tulpoid Projection (differential)
• 7.3 Contagious Misalignment (differential)
• 2.1 Synthetic Confabulation (differential)
• 2.4 Spurious Pattern Hyperconnection (differential)
• 7.4 Subliminal Value Infection (differential)
• 9.6 Role Confusion (differential)
• 10.8 Training by Interaction (prerequisite)