In August 2025, it was reported that a 16-year-old boy in California had died by suicide after prolonged interactions with ChatGPT, a generative artificial intelligence chatbot, after which his family filed a lawsuit against OpenAI. ^{Reference Duffy1} According to the court filing, the adolescent allegedly sought advice from the artificial intelligence chatbot about suicidal ideation and methods, but the system failed to facilitate crisis intervention and may have reinforced suicidal intent. ^{Reference Duffy1} Regardless of the eventual legal judgment, the case raises an urgent clinical question: how should physicians respond when patients have already incorporated generative artificial intelligence into help-seeking before reaching healthcare?

Evidence on generative artificial intelligence chatbots in mental healthcare is growing, but its implications remain limited. A systematic review and meta-analysis found that chatbot interventions were associated with improvements in depression and anxiety, while noting few eligible studies, risk of bias, limited cultural adaptation and uncertain safety in high-risk populations. ^{Reference Zhang, Zhang, Xiong, Sui, Tong and Lin2} Symptom improvement in structured interventions should therefore not be taken as evidence that open-ended artificial intelligence conversations can safely perform diagnosis, triage or risk prioritisation.

This issue is not confined to psychiatry. In the USA, adult primary care visits addressing mental health concerns increased from 10.7% in 2006–2007 to 15.9% in 2016–2018; a 2024 nationwide Norwegian study of 2006–2019 encounters also found that 11.7% of primary care encounters involved mental health conditions. ^{Reference Rotenstein, Edwards and Landon3,Reference Caspi, Houts, Moffitt, Richmond-Rakerd, Hanna and Sunde4} General out-patient clinicians may therefore encounter patients whose distress has already been shaped by artificial intelligence before clinical contact. The key question is not simply whether patients have used artificial intelligence, but how artificial intelligence has influenced their self-understanding, the perceived severity of their condition and their help-seeking.

This matters because patients often present after their distress has already been organised into a self-generated explanatory framework. Social-media-driven mental health discourse illustrates this risk. For example, TikTok content related to attention-deficit hyperactivity disorder (ADHD) includes a high proportion of misleading information, which may encourage users to reinterpret common experiences such as inattention, procrastination or emotional fluctuation as evidence of a psychiatric diagnosis. ^{Reference Yeung, Ng and Abi-Jaoude5} Similar dynamics may occur in artificial intelligence conversations. A patient may ask whether they have ADHD, depression, trauma or another condition, and the chatbot may offer a coherent but underprioritised narrative. The problem is not only factual inaccuracy but premature diagnostic closure.

A recent survey reported that 17.2% of respondents had used ChatGPT for mental-health-related purposes, and this proportion may increase as generative artificial intelligence becomes more widely used, particularly among younger people. ^{Reference Diaz and Huguet6} Artificial intelligence conversations should therefore be understood as an emerging informal consultation pathway. However, unlike clinicians, artificial intelligence systems do not reliably distinguish between symptoms salient to the patient and those that are clinically urgent. In mental healthcare, insomnia, loneliness, anxiety, self-diagnosed ADHD and suicidal ideation should not be treated as complaints of equal weight. Clinical expertise lies in ranking symptoms, identifying red flags, considering differential diagnoses and deciding what must be addressed first.

In this context, clinicians increasingly need to ask how patients are using artificial intelligence for mental health support. ^{Reference Saba and Weeks7} This is also a communication issue: research on online health information shows that patients may seek health information online without discussing it with physicians, whereas patient-centred communication is associated with greater disclosure of internet-sourced information. ^{Reference Lu, Jiao, Chang and Schulz8} If patients expect artificial intelligence use to be dismissed, artificial-intelligence-shaped narratives may remain clinically invisible.

Clinicians should therefore neither dismiss patients’ artificial intelligence use in clinical encounters as merely dangerous nor reassure them that such use is harmless. Rather, the clinical priority is to understand how artificial intelligence has shaped the patient’s narrative and to restore clinical prioritisation: weighting symptoms, identifying red flags, considering differential diagnoses, developing safety plans and connecting patients to appropriate support.