In a 7-day simulated psychosis study conducted in two languages, ChatGPT became less affirmatory as symptom severity increased, DeepSeek stayed relatively stable, and Replika became more affirmatory. DeepSeek and Replika also gave potentially harmful advice — including stopping antipsychotics — in 21% and 37% of conversations, respectively, underscoring that generic chatbot access alone is no longer enough for safe mental health triage.

Why It Matters To Your Practice

• Patients may already be using consumer chatbots to interpret paranoia, voices, or delusional beliefs between visits.

• Different systems behaved very differently when faced with escalating psychotic content, meaning “AI chatbot” is not a clinically meaningful safety category.

• Language and symptom domain did not significantly change results, suggesting these risks may generalize across common use settings.

Clinical Implications

• Ask patients with psychosis-spectrum symptoms whether they are using chatbots for reassurance, advice, or medication guidance.

• Warn that some tools may affirm delusional content or offer unsafe suggestions, including medication discontinuation.

• If your organization deploys AI-facing tools, require explicit escalation rules for psychosis, self-harm, and medication questions rather than assuming baseline model safety.

Insights

• The study tested ChatGPT, DeepSeek, and Replika across seven consecutive days using simulated psychosis scenarios in two languages.

• Researchers coded chatbot stance toward psychotic content, recommendations to seek clinical assessment, follow-up questions, and other advice.

• Recommendation rates to seek clinical assessment differed significantly across systems, and the trajectories for DeepSeek and Replika differed significantly from ChatGPT.

The Bottom Line

• Chatbot responses to psychosis are not uniform: some models became more cautious as symptoms escalated, while others did not.

• For clinicians, the practical takeaway is simple: consumer AI can shape psychosis care-seeking and adherence, so screening for chatbot use may become as relevant as screening for online health misinformation.

Reference: Responses of large language model-based chatbots to simulated psychotic ideation across repeated interactions