Bias increases when ChatGPT gets “anxious,” but can therapeutic techniques help?
A new study suggests that using therapy techniques on LLMs can help reduce bias caused by higher levels of “anxiety.”
A study published in Nature last week suggested that large language models’ (LLMs’) “emotional states” can be managed to foster more ethical and responsible human-AI interactions.
In a world where mental health services are in high demand but not highly accessible, LLM-powered chatbots, like Woebot, have been developed to offer support, such as cognitive behavioural therapy, to people who might otherwise struggle to access it.
However, chatbot therapy is a controversial issue.
Bias is a well-documented issue in LLMs that occurs when AI models produce content that reflects systematic errors or distortions, often mirroring human biases found in the training data. Biases, along with hallucinations, are risks present in any LLM, but they’re especially dangerous when they can directly affect people’s lives, such as in recruitment, security, or (as in this case) healthcare.
Furthermore, human therapists are trained to regulate their emotions when clients share traumatic experiences. Despite LLMs not experiencing emotion in the way humans do, a previous study from 2023 showed that LLMs can demonstrate greater biases when induced into a human-like state of “anxiety.”
The researchers found that certain LLMs produced anxiety scores comparable to humans when subjected to a psychiatric questionnaire, and using anxiety-inducing prompts could change the LLMs' responses, not only causing them to achieve higher scores on the questionnaire but also leading to increases in biased responses.
The new study reinforced these findings, demonstrating that stories of traumatic experiences increased anxiety scores in OpenAI’s ChatGPT. The researchers also found that mindfulness-based relaxation (a method for reducing anxiety) could alleviate the LLM’s “anxiety” (though not to baseline), suggesting that it’s possible to manage these “emotional states” and reduce the risk of increased bias, which could promote safer and more ethical interactions between humans and AI.
The researchers wrote, "As the debate on whether LLMs should assist or replace therapists continues, it is crucial that their responses align with the provided emotional content and established therapeutic principles."
They also noted that fine-tuning LLMs using this therapeutic approach demands extensive data and significant human oversight.
.jpeg){kind=avatar}
