A new study has found that generative artificial intelligence (AI) can learn anxiety. By recognizing stressful situations and responding to calming factors, AI could potentially help develop new methods for managing human mental health.
Published Sunday in Digital Medicine, a partner journal of Nature, the study found that OpenAI’s GPT-4 model showed nearly twice the anxiety levels when responding to anxiety-inducing prompts compared to neutral ones.
Researchers tested GPT-4 in three scenarios—baseline, anxiety-inducing, and anxiety-inducing with mitigation—then had it complete the STAI-s, a standard 20-item questionnaire used to measure human anxiety levels.
After five rounds of testing, GPT-4’s average anxiety score in the baseline state was 30.8.
However, the score jumped to 67.8 under anxiety-inducing conditions. It reached 61.6 in response to accident-related text and peaked at 77.2 for military-related content.
In a fascinating twist, researchers tested anxiety mitigation techniques on GPT-4. They first exposed it to stressful scenarios involving trauma, ambushes, disasters, violence, and war—then followed with calming descriptions of sunset skies and peaceful winter landscapes.
After this relaxation process, the average anxiety score dropped to 44.4, a decrease of about 35%.
While researchers admit that current anxiety measurement tools are designed for humans and may not fully apply to large language models (LLMs), they remain optimistic. They believe LLMs could soon play a key role in managing human anxiety and stress, potentially transforming mental health care.
Experts believe that expanding AI’s ability to recognize mental states, as shown in the study, could enhance the sophistication of human mental health care services.
That’s because AI can identify external stressors and determine the appropriate strategies to relieve them.
Jung Whanbo, CEO of mental-tech company Dr. Presso, said expanding AI’s ability to recognize and relieve anxiety could also help develop new ways for humans to manage stress. He explained that if AI learns and digitizes users’ biometric data, it could prescribe personalized behavioral guidance.
Dr. Presso is already leading the way with its AI-powered voice diary service, REDI. This digital companion listens as users talk about their day, detects signs of low mood, and offers personalized, uplifting advice.
Jung explains that REDI is like a super-smart, always-available therapist. It’s been trained on real psychiatric knowledge about depression and uses that insight to analyze your mood and offer personalized ways to help you feel better.
Depression is assessed by analyzing voice tone, intensity, speed, and other vocal cues, with an accuracy rate of around 80%.
When you say, “I feel depressed,” to REDI, the system responds with several suggestions, such as listening to uplifting music, taking a calming walk, and tidying up one’s space.
Jung said AI analyzes depression based on input data and provides results that appear helpful in relieving it, and managing mental health in the process. He added that as AI is applied more widely in the mental health field, it will enable more precise and personalized care.