Can AI provide emotional support to humans? Researchers at the Digital Emotions Lab have undertaken an intriguing project, focusing a process called cognitive reframing—a strategy of reducing negative emotions by changing the interpretation of emotional situations. To compare the ability of humans and AI to perform cognitive reframing of negative situations, they developed 18 vignettes, such as “My friend forgot my birthday after saying we’d go to dinner together. I feel unwanted.” They then trained both humans and GPT-4 on how to perform cognitive reframing, based on a well-established training process that has been validated in tens of thousands of people. Human raters then evaluated the quality of these efforts based on their effectiveness, empathy, novelty, and specificity.
Key Insights:
Performance Insights: Out of 4195 human attempts, GPT-4 consistently performed better across most dimensions, ranking in the 85th percentile for the quality of its responses, with the exception of specificity (“This rethinking is specific to the following scenario”).Â
Effort vs. Skill: To determine if effort influenced outcomes, researchers offered human participants up to 150% of their base pay for exceptional reappraisals. Despite spending more time on the task, there was no noticeable improvement in their responses. This suggests that GPT-4’s superior performance is likely due to ability rather than more effort.
Analyzing the Content: Using advanced analysis techniques, the study measured how closely the rethought scenarios aligned with the original vignettes. It was found that human responses tended to stick closely to the original scenarios, while GPT-4’s were generally broader. Notably, GPT-4 produced better-quality responses when they were more aligned with the vignettes, whereas humans excelled when they ventured beyond the original context, showing a distinct difference in approach between humans and AI.
This research sheds light on the significant differences in how AI and humans approach the task of cognitive reframing. It reveals that AI can excel in closely following the specific emotional context, while humans show strength in applying a broader perspective. These findings not only help us understand AI’s potential in tasks requiring emotional insight but also suggest ways AI might support human efforts in fields like therapy and customer service, where understanding and reshaping emotions are crucial.