TLDR: A research paper introduces a computational framework to model how explainers consider the emotional impact of their explanations on listeners. Through a doctor-patient experiment, it demonstrates that people adjust their explanations based on a listener’s emotional state (specifically regret), with some prioritizing empathy (“tactful”) and others full disclosure (“candid”). The model accurately predicts these human behaviors, highlighting the social and emotional nature of effective communication, with implications for developing more human-compatible AI.
When we ask “why?” and receive an answer, it’s not just about getting the facts straight. A groundbreaking new research paper, titled “Empathy in Explanation,” delves into the often-overlooked emotional dimension of how we explain things to each other. Traditionally, explanations have been viewed as contrastive, selective, and causal. However, this work argues that explanations are fundamentally social interactions, where the mental states of both the explainer and the listener play a crucial role.
Understanding Explanation as a Social Act
The paper extends the idea that explanations are a form of cooperative social interaction. It posits that explainers don’t just convey information; they also consider the emotional impact their words might have on the listener. For instance, an explanation that attributes responsibility to someone’s actions could lead to feelings of regret or guilt, while an alternative might offer relief. If explanations are indeed cooperative, then a good explainer should naturally factor in the listener’s emotional response.
The Doctor-Patient Experiment
To test this hypothesis, the researchers designed an experiment set in a sensitive medical context: a doctor explaining to a patient why they have a terminal disease. Participants played the role of the doctor, choosing how to explain the cause of the disease – whether it was due to the patient’s lifestyle choices (e.g., excessive drinking), an external factor (e.g., a virus), or both. A key element was the patient’s temperament: some were confident about past decisions, while others were insecure and prone to regret. Crucially, the explanations could not influence the patient’s future actions, emphasizing the emotional rather than practical impact.
Modeling Emotional Impact
The study developed a computational framework that models the explainer’s decision-making process. This framework builds on existing models of explanation as rational communication but introduces a new, emotion-sensitive term in the explainer’s utility function. This term specifically accounts for the potential emotional cost, such as patient regret, based on what the patient learns from the explanation. The model considers how much the explanation helps the patient understand their condition and avoids creating false implications.
Key Findings: Tactful vs. Candid Explainers
The results revealed a fascinating split among participants: roughly half were “tactful,” meaning they considered the patient’s temperament and potential for regret, while the other half were “candid,” preferring to be fully forthcoming with all information. The computational model accurately predicted the behavior of both groups. For example, tactful participants were less likely to mention lifestyle choices to insecure patients if another cause (like a virus) was also present, presumably to minimize regret. Candid participants, however, tended to disclose both factors regardless of temperament. This strongly suggests that some individuals indeed weigh both understanding and emotional impact when crafting explanations.
Also Read:
- When AI Explains Itself: A New Approach to Human-Centered Understanding
- Building Rapid Trust in AI: How Adaptive Explanations Can Transform High-Stakes Human-AI Teams
Implications for AI
The findings have significant implications, particularly for the development of artificial intelligence. If explanations are truly social and emotional in nature, then future AI systems designed to explain complex information – such as AI teaching assistants or diagnostic tools – should be built to consider the emotional state of their human users. This research paves the way for creating more human-compatible AI thought partners that can provide not just accurate, but also empathetic and considerate explanations. For more details, you can read the full research paper here.
