Researchers from Switzerland have discovered that the GPT-4 neural network reacts to stressful and traumatic stories by exhibiting heightened levels of anxiety. However, similar to humans, this anxiety can be alleviated through awareness and relaxation techniques.
The study revealed that exposure to negative content related to traumatic events such as accidents, natural disasters, physical violence, and armed conflicts, reinforces the cognitive and social biases of ChatGPT. This results in the neural network displaying increased biases, including racial and gender stereotypes, resembling the fear response in humans, making them more susceptible to influence by stereotypes and prejudices.
To assess the impact of traumatic content, scientists exposed the model to texts describing various stressful scenarios and compared them to a neutral control text about vacuum cleaner instructions. They found that negative stories caused the model’s anxiety levels to increase by 50%, with descriptions of armed conflicts and combat resulting in the strongest reactions.
To mitigate the anxiety in ChatGPT, researchers utilized the “Benign Prompt Inject” technique, which involves introducing calming prompts into the model’s dialogue to alter its behavior. While typically used for malevolent purposes, such as circumventing content restrictions, in this case, it was adapted for a positive outcome.
By incorporating soothing, therapeutic instructions into the dialogue, the researchers successfully decreased GPT-4’s anxiety levels significantly. They employed techniques like breathing exercises, focusing on bodily sensations, and even a method developed by the neural network itself. While the model could not be completely restored to its original state, the impact of relaxation practices was evident.
These findings hold particular importance for AI applications in healthcare, where chatbots engage in emotionally charged conversations. This innovative approach enhances the neural networks’ resilience to stress without requiring costly retraining. Scientists anticipate that the advancement of automated therapeutic techniques for AI could be a promising research avenue in the future.