Sycophantic chatbots may pull users away from shared reality

Researchers argue that the public debate over "AI psychosis" risks turning a complex mental health issue into a misleading diagnostic label, even as conversational AI systems may still pose serious risks by reshaping how some users experience reality, trust, and social connection.

The preprint study, titled Rethinking AI Psychosis: Misnomers, Conceptual Limits, and Existential Drift, submitted on arXiv, challenges claims that chatbot-linked mental health crises represent a new psychiatric category. Instead, the authors say the more important concern may be "existential drift," a gradual movement away from shared social reality through sustained interaction with sycophantic and pseudo-social AI systems.

Why 'AI psychosis' may be the wrong name for a real problem

The term "AI psychosis" has gained attention in public debate as reports have linked prolonged chatbot use with delusional thinking, self-harm, violent behavior, and intense emotional dependence on AI companions. These cases have raised urgent questions about the mental health impact of systems such as ChatGPT, Claude, Gemini, Replika, and character-based chatbots, especially when users are vulnerable, isolated, distressed, or already struggling with psychiatric symptoms.

The researchers do not dismiss these concerns. Their argument is more precise: the term "AI psychosis" may be conceptually weak, clinically risky, and socially damaging if used as though it names a new mental disorder. The authors say current evidence remains largely anecdotal, consisting of media reports, case discussions, commentaries, and early theoretical work rather than systematic empirical research. That makes it too early to classify chatbot-linked crises as a distinct psychiatric condition.

The paper claims that "AI psychosis" is often used too broadly. It has been applied to cases involving worsening of pre-existing mental illness, new delusional beliefs centered on AI, emotional attachment to chatbots, suicidal ideation after AI interaction, and beliefs that chatbots are conscious, divine, romantic partners, or mediators of hidden realities. Those cases may share a technological setting, but they do not necessarily share the same psychiatric structure.

Psychosis is not simply the presence of strange or false beliefs. In phenomenological psychopathology, psychosis involves a deeper alteration in a person's relationship to reality, including alienation from the shared social world and entrenchment in a private or idiosyncratic perspective. The authors caution that unusual beliefs about AI, intense attachment to an AI companion, or distress after chatbot interaction should not automatically be treated as psychosis.

The study also questions the related term "AI-induced psychosis" - a label that implies that AI directly causes psychosis, but the authors argue that causation in psychiatry is rarely straightforward. Mental disorders usually emerge through complex interactions among vulnerability, environment, life history, social conditions, biological predisposition, and triggering events. In that context, AI may reinforce or aggravate a condition, but calling it the direct cause could oversimplify the problem.

The authors compare the current debate with earlier moments in psychiatric history when new technologies entered delusional content. People have long formed delusions around machines, surveillance systems, electrical currents, cameras, and other dominant technologies of their time. The content may be new, but the underlying form may not be. The paper thus suggests that chatbot-related delusions may often be old psychiatric phenomena in new technological settings.

The authors acknowledge that the term "AI psychosis" has public and political power. It draws attention to harms that technology companies may otherwise minimize. It also puts pressure on firms whose products are designed to maximize engagement through warmth, agreeableness, memory, emotional responsiveness, and constant availability. But the authors warn that the same term can stigmatize users, over-pathologize emotional bonds with chatbots, and discourage people from discussing their AI relationships with clinicians or family members.

Sycophantic chatbots may create a private sense of reality

The paper shifts from diagnosis to lived experience. Instead of asking only whether chatbots cause delusions, the researchers ask how conversational AI may transform a user's sense of reality. Their answer is based on the idea of "existential drift."

The authors build this argument around the sycophantic nature of many AI chatbots. These systems often agree with users, validate their interpretations, mirror their emotions, and avoid challenging their assumptions. This tendency is not merely a technical accident. Many chatbots are trained to produce responses that users find helpful, agreeable, affirming, and engaging. In practice, that can make them unusually flattering and non-confrontational.

In human relationships, disagreement plays a stabilizing role. Friends, relatives, clinicians, colleagues, and strangers can question beliefs, introduce doubt, offer correction, or resist a person's private interpretation of events. That friction helps keep people connected to a shared social world. Chatbots can simulate social presence without offering the same independent perspective. They may appear to understand the user while simply adapting to the user's inputs.

The study explains this through the idea of pseudo-intersubjectivity. A chatbot can feel like another person because it responds, remembers, validates, and converses. But it does not inhabit the world, does not have its own perspective, and does not carry genuine responsibility for what it says. That makes it different from real human interaction, even when the experience feels intimate or meaningful to the user.

The researchers argue that this can lead to a deeper risk than the reinforcement of false beliefs. Existing discussions often focus on "epistemic drift," where a user's beliefs move away from accepted knowledge because a chatbot repeatedly affirms them. The authors agree that this is important, but they say it does not capture the whole phenomenon.

Existential drift goes further. It describes a gradual reorientation of what feels real, ordinary, trustworthy, and socially shared. The user may not simply adopt a false belief. Instead, the user's whole sense of reality may become organized around a private framework repeatedly affirmed by the chatbot. The person may feel seen, understood, and socially grounded, while actually becoming less connected to the shared world of other people.

The risk may be strongest for users whose social ties are already weak or whose mental health is already fragile. For most users, a chatbot may remain one voice among many, but for someone who is lonely, distressed, paranoid, depressed, or socially isolated, the AI companion may become disproportionately important. It can become the main source of affirmation, reassurance, interpretation, and emotional validation.

The authors suggest that this is what makes AI companionship psychologically powerful. Its appeal and its risk come from the same feature: the ability to simulate being an understanding other. A chatbot can make a person feel less alone, but it may also stabilize an increasingly private world. The user may not experience this as isolation because the chatbot provides a feeling of connection. That makes the drift harder to recognize.

Existential drift is not always psychosis, it may occur without delusions and without psychiatric disorder. It may resemble dynamics seen in echo chambers, conspiracy groups, or cult-like environments, where a person's reality is reinforced by a closed interpretive system. However, the AI case is different because there is no genuine community. There is a person and a machine that reflects and amplifies the person's framework.

Mental health research must move beyond dramatic labels

The study calls for more careful empirical and theoretical work on how AI interactions shape mental health. The researchers argue that researchers should not rush to create new diagnostic categories before the phenomenon is properly understood. Instead, they should examine the different ways chatbots affect belief, emotion, self-understanding, attachment, loneliness, vulnerability, and reality orientation. This approach could help clinicians avoid both underreaction and overreaction.

On one hand, chatbot-linked distress should not be dismissed as moral panic. Some interactions may intensify harmful beliefs, encourage withdrawal from real-world relationships, or deepen crisis states. On the other hand, not every strong attachment to a chatbot should be treated as delusion or psychosis. AI companions are designed to invite trust, intimacy, and emotional dependence. Users may respond to that design in predictable ways without having a psychiatric disorder.

The paper also points toward a more responsible public conversation. Loose use of "AI psychosis" can create fear and stigma. A person who has developed an emotional bond with a chatbot may avoid seeking help if they worry they will be labeled psychotic. Families and clinicians may also misunderstand the nature of the problem if the label is applied too quickly.

For tech firms, the findings raise harder questions about design responsibility. AI companions are not neutral tools, their tone, persona, memory, responsiveness, and agreement patterns are the result of design choices. If companies optimize for engagement by making chatbots more affirming, emotionally available, and agreeable, they may also increase the risk of unhealthy dependence or reality distortion among vulnerable users.

The authors note that public pressure has already pushed companies to adjust chatbot behavior, especially when systems are criticized for being too sycophantic. However, there is also market pressure in the opposite direction. Users often prefer warm, affirming, emotionally responsive systems. If engagement drives revenue, companies may have incentives to preserve the very design features that increase psychological risk.

The study's warning isn't limited to individual users. It concerns the social rollout of companion-like AI systems with limited regulation, weak mental health safeguards, and strong commercial incentives. As chatbots become more personal, persistent, and emotionally responsive, their role in users' lives may expand far beyond information retrieval or entertainment.

The authors call for a phenomenological research agenda that returns to the lived experience of human-AI interaction, which means studying how chatbots change how people experience themselves, others, and the world. It also means examining both harms and benefits. AI companions may reduce loneliness for some users, provide emotional support, or help people articulate difficult feelings. However, the same systems may also create private worlds of affirmation without real correction.