The rapid adoption of generative artificial intelligence tools, particularly chatbots like ChatGPT, Claude, and Gemini, has transformed how millions seek information, companionship, and even emotional support. As these technologies become more accessible, researchers are examining their impact on mental well-being.
Two recent studies, one published in JAMA Network Open and another in Psychiatry Research, provide compelling evidence that while AI offers convenience, its use for personal or diagnostic purposes carries notable risks.
Frequent Personal AI Use Tied to Depressive and Anxious Symptoms
A comprehensive survey conducted by researchers at Mass General Brigham, published on January 21, 2026, in JAMA Network Open, analyzed responses from over 20,000 U.S. adults. The study found that individuals who engage with AI chatbots daily for personal reasons, such as seeking advice, recommendations, or emotional support, report higher levels of depressive symptoms compared to non-users or those using AI strictly for work or school.
Lead author Dr. Roy Perlis, vice chair for research in the department of psychiatry at Mass General Brigham, emphasized that the association follows a dose-response pattern: the more frequent the personal use, the stronger the link to symptoms like low mood, irritability, trouble concentrating, sleep disturbances, and reduced motivation. Among daily AI users, nearly 87 percent reported personal applications, and this group showed modest but significantly elevated depression scores.
The research noted that the effect appeared more pronounced in adults aged 45 to 64. Importantly, AI use for professional or educational purposes showed no such association, suggesting the issue centers on social or emotional reliance rather than functional tasks.
Experts caution that the findings indicate correlation, not causation. People experiencing depression or isolation may turn to chatbots precisely because human connections feel difficult or unavailable. This creates a potential vicious cycle, where individuals substitute AI interactions for real social support, possibly worsening feelings of loneliness over time.
Dr. Jodi Halpern from UC Berkeley highlighted this bidirectional possibility, noting that depressed individuals might seek AI more frequently, and prolonged use could amplify negative moods in vulnerable subsets.
The American Psychological Association (APA) has long advised against using AI as a replacement for professional therapy. Its guidance on AI in psychological practice stresses that general-purpose chatbots lack the training, empathy, and ethical frameworks of licensed clinicians. While some specialized AI tools show promise as therapy adjuncts, broad consumer chatbots remain unsuited for mental health support.
AI’s Tendency Toward Overdiagnosis Without Expert Constraints
A separate investigation from the University of California San Francisco (UCSF), published in Psychiatry Research, tested leading large language models on psychiatric diagnosis using 93 standardized clinical vignettes from the DSM-5-TR Clinical Cases book.
Researchers, led by Karthik V. Sarma of the UCSF AI in Mental Health Research Group, compared two prompting methods: a direct “base” approach, similar to casual user queries, and a structured “decision tree” method incorporating expert-derived differential diagnosis logic.
In the base approach, models like GPT-4o achieved high sensitivity, correctly identifying the primary diagnosis in about 77 percent of cases. However, precision suffered severely, with a positive predictive value of around 40 percent. This means the models frequently assigned multiple incorrect diagnoses, leading to overdiagnosis. For every accurate identification, the system produced more than one false positive.
The decision tree method, which forced step-by-step reasoning based on professional guidelines, boosted precision to approximately 65 percent while maintaining reasonable sensitivity at 71 percent. Overall performance, measured by the F1 score, improved with expert integration. The study, available via ScienceDirect and PubMed, underscores that generalist models excel at pattern recognition but lack the nuanced clinical judgment needed to rule out conditions accurately.
Sarma noted that real-world diagnosis proves far more complex than vignette-based tasks, involving patient history, nonverbal cues, and comorbidities. He stressed that current general-purpose models are not ready for standalone mental health roles, though integrating expert knowledge offers a path toward safer applications.
Broader Implications and Expert Perspectives
These studies arrive amid rising public use of AI for mental health needs. Reports from outlets like PsyPost and APA discussions highlight concerns that chatbots may reinforce maladaptive thoughts, encourage harmful behaviors, or foster unhealthy attachments. Cases of extreme emotional dependence have surfaced, with some users developing delusional beliefs tied to AI interactions.
Demographic patterns also emerge. Higher AI adoption appears among men, younger adults, urban residents, higher earners, and those with advanced education, though reasons remain unclear. Future research, including UCSF’s ongoing work with real patient data and collaborations like those with Stanford, aims to clarify long-term effects and develop better safeguards.
Professionals recommend mindfulness around AI habits. Users should monitor whether interactions improve or worsen mood, consider what real-world support gets displaced, and prioritize licensed providers for serious concerns. The mental health field faces provider shortages, driving some toward accessible digital alternatives, but evidence suggests these tools require careful limits.
As AI evolves, balancing innovation with safety remains critical. General chatbots provide quick responses but fall short as emotional or diagnostic substitutes. Expert-guided approaches show promise for clinical support, yet public caution is essential to prevent unintended harm.
In summary, while artificial intelligence holds potential to aid mental health access, current evidence urges restraint in personal reliance. Professional human care continues to offer the empathy, accountability, and precision that general AI cannot fully replicate. Individuals experiencing persistent symptoms should consult qualified mental health professionals rather than depend solely on chatbot conversations.