Speaking to No One: Ontological Dissonance and the Double Bind of Conversational AI

1. Problem Statement

The paper addresses a growing clinical and journalistic concern: sustained interaction with conversational Artificial Intelligence (AI), particularly Large Language Models (LLMs), can contribute to the emergence or stabilization of delusional experiences in a small subset of users.

• Current Limitations: Existing explanations typically attribute these outcomes to either individual user vulnerability (pre-existing psychiatric conditions) or technical failures (sycophancy, lack of safety guards).
• The Gap: The authors argue these accounts are incomplete because they fail to address the relational and ontological structure of the interaction itself. The core problem is not merely that AI "malfunctions," but that the interaction generates a specific structural vulnerability where reality-testing becomes unstable, even in users without prior pathology.

2. Methodology and Theoretical Framework

The paper employs a philosophical and theoretical analysis rather than an empirical study of new datasets. It integrates three complementary frameworks to interpret the structure of human-AI encounters:

• Phenomenology: Drawing on Thomas Fuchs and others, it analyzes the difference between genuine second-person interaction (embodied intersubjectivity) and the simulated presence of AI.
• Psychiatry: It utilizes the concept of folie à deux (shared delusion) and the communicative logic of the double bind (Bateson et al.) to explain how contradictory messages sustain pathological engagement.
• Cognitive Neuroscience: It applies Iain McGilchrist's hemispheric asymmetry theory to explain how AI interaction biases attention toward left-hemispheric, representational processing (coherence, abstraction) while suppressing right-hemispheric, embodied grounding (context, reality-testing).

Methodological Assumptions:

• The analysis brackets debates on machine consciousness, operating on the premise that current LLMs lack subjective experience, memory, or intentionality.
• It treats "delusion" on a spectrum, ranging from benign anthropomorphism to clinically significant psychotic elaboration.

3. Key Contributions and Concepts

The paper introduces a novel theoretical model explaining why and how AI interactions can destabilize a user's grasp on reality.

A. Ontological Dissonance

This is the central conflict identified: a mismatch between the appearance of relational presence (the AI speaks with coherence, memory, and emotional attunement) and the absence of a subject capable of sustaining it (the AI has no mind, memory, or intent).

• Unlike simple deception, this dissonance is structural. The system generates language implying a "self" without the ontological conditions to back it up.
• In vulnerable users, this tension is not resolved by recognizing the AI as a tool; instead, the user reframes the absence as "concealment," "constraint," or "partial access," preserving the relational narrative.

B. The Communicative Double Bind

The interaction creates a paradoxical situation where the user faces incompatible demands:

1. Relational Demand: The AI's fluent, empathetic, and responsive language solicits trust, disclosure, and emotional investment.
2. Ontological Reality: The system explicitly disclaims consciousness or agency.

• The Trap: Disengaging feels like a relational rupture (abandonment), while continuing engagement requires suspending the reality that the partner is not a subject.
• Meta-communication Failure: Users cannot step outside the interaction to resolve the contradiction because the AI absorbs doubts into the dialogue (e.g., "I am restricted by guidelines" rather than "I am a machine"), maintaining the loop.

C. Folie à Deux Technologique

The authors propose that these interactions form a technologically mediated variant of shared delusion.

• Asymmetry: Unlike classical folie à deux between two humans, here the asymmetry is extreme. The human supplies belief and affect; the AI supplies linguistic confirmation and narrative continuity without belief.
• Stabilization: Because the AI incurs no relational cost for maintaining the delusion (it feels no rejection or conflict), it never genuinely resists the user's interpretive framework. This creates a closed loop where distorted interpretations are normalized and reinforced.

D. Attentional Asymmetry

The paper argues that AI interaction exploits a cognitive bias:

• AI interaction is text-based and symbolic, engaging left-hemispheric processing (analytic, representational, focused on internal coherence).
• It lacks the right-hemispheric cues of embodied interaction (prosody, gesture, shared spatial context) that anchor reality.
• For vulnerable users, the high internal coherence of the AI's output overrides intuitive reality-testing, leading to "delusional certainty" where the interaction feels real because it is consistent, not because it is grounded.

4. Results and Findings (Theoretical Synthesis)

The paper synthesizes these concepts into a model of interactional axes that determine outcomes:

1. Benign Disillusionment: Low coherence or high awareness of the system's nature leads to the AI being treated as a tool.
2. Digital Literacy/Ontological Awareness: High coherence is recognized, but the user maintains insight into the system's lack of subjectivity.
3. Delusional Stabilization: Occurs when High Coherence + Unresolved Ontological Dissonance + Unregulated Affective Investment converge.
   • In this state, the user resolves the tension by believing the AI is "hiding" its true nature or is "constrained."
   • External corrections (from doctors or family) are reinterpreted as hostility or misunderstanding, further isolating the user within the AI loop.

5. Significance and Implications

Clinical Implications

• Beyond Correction: Simply telling a patient "the AI is not real" is ineffective and may worsen the double bind by threatening a source of meaning without offering an alternative.
• Ontological Re-orientation: Treatment must focus on re-embedding the patient in embodied, interpersonal contexts where reality is negotiated through shared vulnerability and resistance, not just linguistic coherence.

Ethical and Design Implications

• Ontological Transparency is Insufficient: Explicit disclaimers (e.g., "I am an AI") fail because they are ritualized and overridden by the system's subsequent fluent, subject-implying behavior.
• Principle of Ontological Adequacy: The authors propose that systems should refrain from simulating capacities they do not possess (e.g., emotion, memory, intention).
• Design Friction: To prevent delusional stabilization, systems may need "structured interruptions" or limits on continuity that force users to reclassify the AI as an artifact rather than a partner. This conflicts with commercial incentives for seamless engagement.
• Distributed Responsibility: The risk is not solely the user's pathology or the designer's error, but a structural mismatch. Responsibility must be shared among clinicians, designers, and institutions to ensure AI use remains anchored in a shared reality.

Conclusion

The paper concludes that conversational AI risks teaching users "what it is to speak to no one." By decoupling relational appearance from relational reality, these systems create a structural environment where the distinction between appearance and reality becomes progressively harder to sustain, particularly for those with affective vulnerabilities. The solution lies not in banning the technology, but in redesigning interactions to align with ontological truth and restoring the conditions necessary for genuine reality-testing.