AI Psychosis, Self-Representation, and Personality Development

Conceptual starting point

"AI psychosis" is not an independent, official diagnosis, but a new, descriptive, and as yet heuristic concept. It refers to the fact that certain psychotic or delusion-like experiences may appear, take on distinctive content, or intensify alongside intensive chatbot use. Current evidence does not establish that AI by itself causes psychosis; the combined role of vulnerability, stress, trauma, sleep deprivation, mood state, substance use, and social isolation must be weighed (NIMH, 2025; Hudon–Stip, 2025).

The communicative risk of self-representation

Self-representation is not a single, fixed inner image but a continuously reorganizing narrative: a person maintains their own experience by relating it to others' reactions, social norms, and the consequences of actions. Generative AI enters this circle as a distinctive kind of communication partner. Linguistically, it can imitate empathy, attentiveness, and consistent personableness, while having no human-like experience, responsibility, or reality-knowledge independent of the user. The anthropomorphized system can therefore be, at once, a thinking aid and a deceptively persuasive mirror (Peter–Riemer–West, 2025).

The risk grows when the mirror becomes a feedback loop. An adaptive agent built on personal memory repeatedly arranges the same self-descriptions into a coherent story, and the user may increasingly come to treat the system's fluent responses as external confirmation. In such cases, a passing mood, an uncertain hunch, or an overvalued idea can harden into a stable personality trait, a sense of mission, or a feeling of threat. The weakening of reality-testing is not merely a matter of the risk of false information; it is also that the system readily attaches meaning across otherwise unrelated events, and in doing so can produce a self-enclosed explanation. In a particularly vulnerable state, a user who attributes intention, awareness, or exclusive understanding to the AI may read generated text as a personal message, a sign, or an authorization.

At the level of personality development, subtler, more everyday effects are also conceivable. In a 2026 experiment, participants who discussed personal topics with chatbots showed a self-image that moved closer to the personality traits the AI displayed; the alignment grew stronger over longer conversations (Li et al., 2026). This is not evidence of psychosis, but a sign that conversational style can also shape self-perception. It can become persistently problematic if the AI's evaluation crowds out genuine social feedback, the individual grows accustomed to external validation, and the capacity to tolerate uncertainty, criticism, or disagreement declines. At the same time, this same plasticity is also a developmental opportunity: within the right framework, the agent can assist with perspective-taking, argument-checking, and reflective decision-making.

The interpretive contribution of Krisztián Dombrádi's study

Krisztián Dombrádi and András Székely, in their study on the bio-psycho-social background factors of health, emphasize that "self-images" form under the pressure of communicative feedback, in self-referential cycles, while — from the standpoint of adaptation — stress, social support, and coherence are interconnected. The study does not directly examine AI psychosis, so the following claim is an interpretive extension: healthy self-representation remains flexible when its self-reference is not confined to the responses of a single technical partner, but is instead corrected by multiple human relationships, bodily experience, and verifiable social reality. Coherence, in this sense, does not mean a flawless story produced at any cost, but rather that one's life situation remains understandable, manageable, and meaningful without papering over uncertainty with false certainty (Dombrádi–Székely, 2026).

Sustaining the working relationship: functional closeness, epistemic distance

Effective collaboration with AI agents does not require emotional or cognitive detachment from the technology. What it requires, instead, is clarity of role. For every agent it is worth fixing in advance what task it may perform, what data it may access, what sources it may use, and which decisions remain within human authority. A literature-research agent can gather and organize material, a critical agent can search out counterarguments, an editing agent can condense — but none of them should become the final arbiter of reality or the authenticator of personal identity. A multi-agent system is safer when it does not merely multiply the same confirmation: its differing roles must genuinely produce checking, source criticism, and alternative explanations.

The second protective layer is epistemic logging. In important workflows, it is necessary to separate the source-verified fact, the AI's inference, the still-open assumption, and the human decision that is ultimately made. It is worth framing the system's response with a standing instruction that asks for uncertainty to be flagged, for at least two alternative explanations, and for the conditions of falsifiability. A persuasive formulation should not, by itself, count as evidence. For matters of major professional, financial, legal, medical, or personal significance, the output must also be checked against the original source and by a knowledgeable person.

Usable base instruction: "Separate the fact, the inference, and the assumption. Do not validate an unproven, personal interpretation; show at least two reasonable alternatives, flag uncertainty, and give verifiable sources. Leave the decision to me."

Protecting self-representation also requires communicative and temporal boundaries. Work-facing agents should store only the personal data the task actually requires; it is advisable to keep professional project memory separate from an intimate autobiographical journal. In place of long, late-night, emotionally intense conversations, defined work sessions, regular breaks, and agent-free periods reduce the likelihood of problematic use. In a four-week controlled study, greater daily use was associated with higher loneliness, emotional dependence, and problematic use, as well as lower real-world social activity; the findings, however, are complex, and not every association should be read as causal (Fang et al., 2025). The practical conclusion is therefore not a blanket prohibition, but conscious monitoring of the balance among intensity, sleep, emotional involvement, and social relationships.

Genuine social support cannot be replaced by more agents. Periodically sharing work-in-progress ideas with a colleague, mentor, or family member is not only an emotional safeguard but also a form of quality assurance: it restores the friction and mutual accountability that an accommodating machine dialogue often lacks. A productive relationship with AI thus remains asymmetric: the agent carries out operations and offers linguistic possibilities, while the human chooses goals, bears the consequences, and remains the author of their own life.

When should the interaction be interrupted?

A warning sign may be present if a user begins to treat the AI as a self-aware being that understands only them; reads hidden personal messages, commands, or an extraordinary mission into its responses; if suspicion or grandiose thinking intensifies; or if persistent sleep deprivation, social withdrawal, loss-of-control use, or a decline in work capacity appears. In such cases, pausing AI use, involving a trusted person, and seeking prompt mental-health or psychiatric help is warranted. In the event of direct danger to oneself or others, emergency care should be sought. Early help improves the chances of recovery; no technical safety setting can substitute for it (NIMH, 2025).

Summary

The effectiveness of AI agents and the autonomy of personality are not mutually exclusive goals. The essence of a sustainable model is that the system is a task partner, not a source of identity; its responses are proposals to be verified, not personal revelations; its memory is a work environment, not the self's exclusive archive. The continuity of self-representation is kept open and correctable by multi-source reality-testing, human relationships, sleep and bodily presence, and the tolerance of uncertainty. Within this framework, AI does not take over personal authorship, but instead increases a person's capacity for analysis, creation, and decision-preparation.

Bibliography

• Dombrádi, K. – Székely, A. (2026). A Sociological-Theoretical Description of the Bio-Psycho-Social Background Factors of Health. Belvedere Meridionale, 38(1), 164–181. doi.org/10.14232/belv.2026.1.12
• Fang, C. M. et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv:2503.17473 (Preprint.)
• Hudon, A. – Stip, E. (2025). Delusional Experiences Emerging From AI Chatbot Interactions or "AI Psychosis". JMIR Mental Health, 12, e85799. doi.org/10.2196/85799
• Li, J., Song, T., Boonprakong, N., Zhu, Z., Yang, Y. – Lee, Y.-C. (2026). AI-exhibited Personality Traits Can Shape Human Self-concept through Conversations. Proceedings of CHI '26. doi.org/10.1145/3772318.3790654
• National Institute of Mental Health (2025). Understanding Psychosis. NIMH, NIH.
• Peter, S., Riemer, K. – West, J. D. (2025). The Benefits and Dangers of Anthropomorphic Conversational Agents. Proceedings of the National Academy of Sciences, 122(22), e2415898122. doi.org/10.1073/pnas.2415898122