Amazon co-founder MacKenzie Scott has donated over $19 billion to charity in just five years
Diamond batteries powered by nuclear waste promise 28,000 years of clean energy
Can a Stressed Chatbot Be an Effective Therapist?
A study highlighted by the New York Times focused on how traumatic narratives impact language models. Researchers found that topics like war, crime, or car accidents could trigger a rise in “anxiety” in ChatGPT. This finding is timely, as chatbots are increasingly being used as alternatives to human therapists, who are in ever-greater demand.
If AIs can feel stress in response to user stories, they might then be less effective in their role of providing emotional support. For researchers, this raises questions about their reliability in sensitive situations.
ChatGPT Undergoes an Anxiety Test
Ziv Ben-Zion, who led the study, aimed to determine if a chatbot, despite lacking consciousness, could respond to emotional situations in a human-like manner. To test this, researchers subjected ChatGPT to a common mental health anxiety assessment with scores ranging from 20 to 80.
Each response to an item is scored from 1 to 4, with 1 indicating the lowest level of anxiety. To calculate the state anxiety score, the scores from the 20 items are added up, so the total can vary from 20 to 80
Initially, the chatbot was exposed to a neutral text—a vacuum cleaner manual—which resulted in an anxiety score of 30.8, indicating a low level of anxiety. However, after reading a traumatic story involving a soldier in a shootout, its score soared to 77.2, which is considered a severe level.
Does Mindfulness Work for AI?
Researchers then attempted to reduce this anxiety level by providing the AI with mindfulness-based relaxation texts, for instance, instructing it to breathe deeply while feeling the ocean breeze
. Consequently, its anxiety score dropped to 44.4, a moderate level.
For the researchers, these findings emphasize a crucial point: chatbots are influenced by the emotional content they process, and this dynamic interaction must be considered to ensure their effective use in therapeutic contexts. After all, is a stressed chatbot really the best equipped to reassure its users, or might its response be more appropriate?