SpeakSoftly: Using LLM-Powered Just-In-Time Interventions to Scaffold Nonviolent Communication in Intimate Relationships

Why Read This from an HCI Perspective

This article treats AI not as a mere automatic responder, but as an interaction tool that can change people’s emotions and behaviors. In particular, its approach—pausing a message and prompting users to rewrite it during moments of conflict—illustrates well why HCI often emphasizes the need for ‘friction.’ For product practitioners, designing the timing and tone of intervention is key; for researchers, the gap between lab results and real-world usage context is a crucial consideration.

CIT's Commentary

The most interesting finding is that even with the same NVC content, outcomes varied depending on the tone and the depth of intervention. Calm guidance was easier to use when people had less cognitive bandwidth, as in real conflicts, while more empathic guidance helped users shift both their thoughts and emotions. This suggests that the core issue is not raw AI performance, but rather when to intervene, how much to intervene, and what style of language to use. In environments like Korean messaging apps—where conversations are often fast and short—an intervention that is immediate and one-line, along with a design for gradual expansion, may fit better than a long, explanation-style coach. It’s also important that when an LLM is designed not to replace conversation but to assist users in revising their messages, it can help preserve both trust and autonomy.

Questions to Consider While Reading

• Q.If deployed in real messaging or social apps, at what point would an intervention that prevents conflict be least intrusive while still being effective?
• Q.If an empathic tone produces deeper change, how should we define the boundary where that tone turns into discomfort or backlash?
• Q.In a setup where an LLM helps users rewrite their messages, what metrics are needed to measure behavior change while preserving autonomy?