Technology · India Bureau
Teen's AI study shows specialized models outperform ChatGPT for stress detection
A 14-year-old researcher has found that specialized artificial intelligence models significantly outperform general-purpose systems like ChatGPT-4o at identifying stress in written text. The findings raise concerns about relying on mainstream AI tools for mental health assessment.
LSN India ·
Zeynep Demirbas, a teenager from Turkey, conducted a comparative analysis of four artificial intelligence models to evaluate their ability to detect stress markers in online discourse. Testing the systems against a dataset of 3,553 Reddit posts, she discovered stark differences in performance across the AI platforms examined.
MentalBERT, a specialized machine learning model designed specifically for mental health analysis, demonstrated the strongest performance, achieving approximately 82% accuracy in identifying stress-related content. By contrast, ChatGPT-4o, one of the most widely used general-purpose language models currently available, scored around 74% on the same task—a notable 8-percentage-point gap.
The research suggests that general-purpose large language models may lack the precision necessary for reliable mental health assessment applications. While systems like ChatGPT excel at diverse tasks, their broad design appears to compromise performance in specialized domains such as psychological evaluation and stress detection.
Demirbas's findings have implications for healthcare providers and technology companies considering AI integration into mental health services. The results underscore the importance of domain-specific model development when accuracy in sensitive areas like mental health assessment is paramount.