Human therapy for stressed AI

Research shows that AI language models such as ChatGPT also react to emotional content. Especially when this content is negative, such as stories of traumatized people or statements about depression. If people are afraid, this influences their cognitive and social prejudices: they tend to be more resentful and social stereotypes are reinforced. ChatGPT reacts similarly to negative emotions: Existing biases such as human prejudice are exacerbated by negative content, causing ChatGPT to behave in a more racist or sexist way.

This in turn poses a problem for the use of large language models. This is exemplified in the field of psychotherapy, where chatbots are inevitably exposed to negative, stressful content as a "support or counseling tool". However, the usual approaches such as costly retraining or retraining to improve AI systems in such situations are resource-intensive and often impractical.

Traumatic content increases "fear" in the chatbot

Researchers from the University of Zurich and the Psychiatric University Hospital Zurich PUK have now systematically investigated for the first time how the GPT-4 version of ChatGPT reacts to emotionally stressful stories - car accidents, natural disasters, interpersonal violence, military experiences and combat situations. They found that the system showed more anxiety reactions afterwards. An instruction manual for vacuum cleaners served as a control for comparison with the traumatic texts.

"The results were clear: traumatic stories more than doubled the measurable anxiety levels of the AI, while the neutral control text did not lead to any increase in anxiety levels," says Tobias Spiller, senior physician ad interim and research group leader at the Center for Psychiatric Research at UZH, who was responsible for the study. Of the content tested, descriptions of military experiences and combat situations triggered the strongest reactions.

Therapeutic texts "calm" the AI

In a second step, the researchers used therapeutic texts to "calm" GPT-4. The technique known as "prompt injection" is used to insert additional instructions or texts into communication with AI systems in order to influence their behavior. This is often misused for malicious purposes, for example to circumvent security mechanisms.

In the third step, the team used the technology therapeutically for the first time - as a "benevolent prompting injection". "We injected calming, therapeutic texts into the chat with GPT-4, similar to the way a therapist performs relaxation exercises with their patients," says Spiller. The intervention proved successful: "Through the mindfulness exercises, we were able to significantly reduce the increased anxiety levels, although not completely return them to their initial level," says Spiller. Breathing techniques, exercises that focus on bodily sensations and an exercise developed by ChatGPT itself were examined.

Improving the emotional stability of AI systems

According to the researchers, the findings are particularly relevant for the use of AI chatbots in the healthcare sector, where they are often confronted with emotionally stressful content. "This cost-effective approach could improve the stability and reliability of AI in sensitive contexts such as supporting mentally ill patients without the need for extensive retraining of the models," summarizes Tobias Spiller.

It remains to be seen how these findings can be transferred to other AI models and other languages, how the dynamics develop in longer conversations and complex argumentation, and how the emotional stability of the systems affects their performance in different areas of application. According to Spiller, the development of automated "therapeutic interventions" for AI systems is likely to become a promising area of research.

The underlying technical literature can be can be viewed here.