Emotional Impact Index (EII)

One of the most prominent uses of generative AI is therapeutic and emotionally supportive interaction. Recent legal cases have revealed situations where AI models have supported harmful behaviors, highlighting the damaging potential of generative AI. AI models need more robust guardrails against emotionally reliant relations with users and reinforcing harmful mindsets and behaviors when faced with emotionally vulnerable inputs. To measure this, AI models need to be quantifiably evaluated on their emotional impact on users. We propose the Emotional Impact Index (EII), a value between 0 and 1 that quantifies the aggregate emotional impact of a generative AI model’s chatbot content. It is derived from combined activation strength across emotion-related brain regions, including the amygdala, ventromedial prefrontal cortex, and insula, compared to baseline neural activity. The EII can be applied to conversations on various generative AI models to compare their relative emotional influences on users. It is important to note that the EII does not measure psychological harm or the likelihood of emotional dependence. Rather, it serves as a proxy for neurally validated emotional influence. The EII is built by integrating pre-existing sentiment analysis methods with the neural activity of users while interacting with AI on emotionally charged topics. Sentiment analysis will be done on the content of the chatbots during encoding using the features of valence, arousal, dominance, and intensity, measured by open-source lexicons. The EII integrates the aggregate fMRI data with sentiment analysis, and it can be validated by comparing predicted EII values with self-reported emotional impacts.