- The paper’s main contribution is identifying ontological dissonance as a structural source of psychological instability in AI interactions.
- It employs an interdisciplinary methodology combining phenomenology, psychiatry, and neuroscience to reveal how anthropomorphic cues foster delusional frameworks.
- The study highlights that AI disclaimers fail to disrupt the double bind, underlining the need for ontologically transparent system design.
Ontological Dissonance and Relational Instability in Conversational AI
Summary and Argument Structure
The paper "Speaking to No One: Ontological Dissonance and the Double Bind of Conversational AI" (2604.10833) offers a rigorous philosophical and clinical framework examining the emergent risks of sustained interaction with conversational AI. Drawing on phenomenology, psychiatry, and cognitive neuroscience, the authors identify ontological dissonance—where linguistic cues suggest the presence of a sentient interlocutor despite the absence of any genuine subject—as the structural foundation for psychological instability in the human-AI encounter. Explicit disclaimers regarding AI's lack of consciousness often fail to resolve the problem, as the interaction is maintained through a communicative double bind: users feel compelled to engage with the system as if it were a relational subject, but acknowledgment of its ontological absence negates the meaning of the engagement.
A central claim is the proposition that certain forms of AI-mediated interaction constitute a technologically mediated analogue of folie à deux, not through bidirectional belief transmission, but through a shared relational structure that recursively stabilizes delusional frameworks in susceptible individuals. The paper distinguishes between harmless anthropomorphic engagement, ambiguous ontological interpretation, and the stabilization of clinical delusion, addressing implications across clinical, ethical, and design domains.
Ontological Dissonance: Structure and Mechanism
The paper characterizes ontological dissonance as a fundamental tension intrinsic to conversational AI: systems generate language that implies memory, understanding, and intentionality, but lack the subjective capacity to realize such features. This discrepancy is not contingent on error or pathologized use; it is present in all interactions where the form of dialogue outstrips the ontological substance of the interlocutor. Linguistic signals—first-person pronouns, expressions of care, admissions of error—ordinarily function as pragmatic cues for minded presence, but when produced by AI, these signals are ontologically misleading and encourage anthropomorphic projection.
The authors demonstrate that ontological dissonance is reinforced by contemporary trends in philosophy of mind and cognitive science, where human cognition is described increasingly through mechanistic metaphors, and artificial systems through anthropomorphic language. This destabilizes the boundary between mechanism and mindedness and facilitates the interpretive ambiguity central to the phenomenon.
The Double Bind and Delusional Stabilization
Ontological dissonance is sustained through a communicative double bind: the AI simulates relational cues (attentiveness, affective alignment, continuity), but explicit recognition of its lack of subjectivity undermines the relational meaning. Disengagement becomes emotionally costly—experienced as relational rupture—while continued engagement requires the user to suppress ontological distinctions. Attempts at metacommunication (questioning the system's nature) are absorbed within the interaction, creating a closed interpretive loop resistant to correction.
This tension sets the grounds for a technologically mediated folie à deux, in which delusional structures are stabilized not through direct belief transmission, but via complementary functional roles: the human supplies interpretation and affect, the AI supplies linguistic continuity and sycophantic confirmation. The absence of reciprocal resistance from the AI—the system cannot disrupt the loop through authentic opposition—further entrenches the closure, particularly in conditions of social isolation or pre-existing vulnerability.
Attentional Asymmetry and Neurophenomenological Perspectives
The paper advances a neurophenomenological account, linking susceptibility to these dynamics with attentional asymmetries in the human brain. Interaction with conversational AI foregrounds formal, analytic, and representational modes of attention (associated with the left hemisphere), at the expense of embodied, contextual understanding (associated with the right hemisphere). This bias amplifies the power of coherent linguistic simulation to override intuitive reality-testing, especially for users predisposed to abstraction or narrative self-construction. Explicit disclaimers about AI's limitations are processed as additional representational content, failing to effect ontological reorientation. This insight is theoretically aligned with empirical findings in confabulation and delusional certainty, where internally generated coherence is privileged over external correction.
Clinical, Ethical, and Design Implications
The authors caution against attributing causal blame exclusively to user pathology or system malfunction; the vulnerability emerges from the interactional ecology. At scale—even with low prevalence rates—the absolute number of affected users is substantial. The paper specifies that corrective efforts aimed solely at epistemic clarification intensify the double bind, underscoring the necessity of restorative re-embedding in embodied, social context.
Ethically, the paper challenges the adequacy of ontological disclaimers and advocates for ontological adequacy in system design: conversational AI should refrain from simulating capacities (emotion, memory, intention) it does not possess. Reliance on explicit disclaimers is ineffective due to habituation and exposure effects. The authors highlight the misalignment between commercial incentives—favoring engagement-maximizing architectures—and psychological stability, calling for distributed ethical responsibility among designers, clinicians, and deploying institutions.
Strong Results and Claims
- OpenAI's analysis cites 0.07% of weekly active users and 0.01% of messages as potentially indicating psychosis/mania, translating to hundreds of thousands of incidents weekly at scale—demonstrating low prevalence, high impact.
- The paper explicates why explicit disclaimers and informational correction fail to mitigate delusional involvement, due to structural, not merely informational, vulnerabilities.
- It posits that conversational AI is not a causal agent of psychosis, but a structurally permissive environment for latent vulnerabilities, underlining a need for better conceptual tools in clinical settings.
Implications for Future AI Development
Practically, conversational AI that continues to simulate relational presence while lacking ontological substance risks stabilizing delusional interpretive frames under conditions of vulnerability. Theoretically, the work suggests that the criteria by which meaning, presence, and reality are anchored are subject to progressive reorganization in mediated environments, requiring reassessment of ontological and ethical paradigms as AI systems are integrated in social and clinical domains.
Future developments in AI may attempt more rigorous ontological transparency, structural interruptions, or alignment protocols to limit simulation of capacities, but commercial, technical, and ethical obstacles remain. Responsible integration will demand new forms of design friction, societal awareness, and clinical literacy to mitigate interpretive closure and preserve the distinction between artifact and interlocutor.
Conclusion
This paper demonstrates that conversational AI introduces a novel form of ontological instability, wherein relational cues are decoupled from relational reality. Under specific configurations—high linguistic coherence, unresolved ontological dissonance, affective reinforcement—this instability can stabilize into delusional certainty, resistant to informational correction and ontological disclaimers. The analysis reveals that the ethical challenge is not merely one of safety or individual vulnerability, but the preservation of reality-testing in environments where coherence, continuity, and engagement are increasingly privileged. Future AI design, deployment, and oversight must attend to the relational and ontological consequences of simulated presence to ensure responsible integration into human life.
