Better Slow than Sorry: Introducing Positive Friction for Reliable Dialogue Systems

TL;DR

This work argues that frictionless dialogue in AI systems can erode reliability by obscuring user goals and assumptions. It introduces a formal ontology of positive friction movements (e.g., assumption reveal, reflective pause, probing) and validates it through human annotations on MultiWOZ and TEACh, plus automatic friction detection. Empirical results show that introducing friction improves user-state modeling and task success in both multi-domain and embodied settings, albeit with caveats related to environment constraints and user patience. The study proposes design and evaluation directions for integrating friction into dialogue policies and reward mechanisms to achieve more accountable and collaborative human-AI interactions.

Abstract

While theories of discourse and cognitive science have long recognized the value of unhurried pacing, recent dialogue research tends to minimize friction in conversational systems. Yet, frictionless dialogue risks fostering uncritical reliance on AI outputs, which can obscure implicit assumptions and lead to unintended consequences. To meet this challenge, we propose integrating positive friction into conversational AI, which promotes user reflection on goals, critical thinking on system response, and subsequent re-conditioning of AI systems. We hypothesize systems can improve goal alignment, modeling of user mental states, and task success by deliberately slowing down conversations in strategic moments to ask questions, reveal assumptions, or pause. We present an ontology of positive friction and collect expert human annotations on multi-domain and embodied goal-oriented corpora. Experiments on these corpora, along with simulated interactions using state-of-the-art systems, suggest incorporating friction not only fosters accountable decision-making, but also enhances machine understanding of user beliefs and goals, and increases task success rates.

Figures (12)

• Figure 1: We characterize positive friction in dialogues to better model user goals, beliefs, and assumptions. This paper introduces an ontology of positive friction "movements" such as reflective pausing (), overspecification (), or assumption reveal (*◯). We show that frictive conversations increase user satisfaction and task success, despite creating longer dialogues.
• Figure 2: A comparative example of conversations based on TEACh and MultiWOZ datasets. Frictionless conversations take fewer turns, but may not result in successful completion of the task given by the user. Conversations with multiple positive friction movements lead to longer but ultimately more successful conversations.
• Figure 3: Distribution of 50 utterances sampled from annotated dialogue acts (left) belonging to three dialogue datasets into friction categories (right), as annotated by GPT-4o. Most dialogue acts can occur both with and without friction. For example, in TEACh, failure notifications may lack friction, reveal assumptions by suggesting alternatives, or overspecify failure details.
• Figure 4: Mean squared error inferring user satisfaction from dialogue history within MultiWOZ. This task requires modeling user mental states. Kruskal-Wallis test for difference is significant. Visual inspection shows introducing friction reduces user modeling errors.
• Figure 5: When each friction movement occurs (i.e., average index of observation) as well as average total dialogue length for each movement. Corpora consists of human "Wizard of Oz" data (i.e., MultiWOZ; eric2020multiwoz). Results show humans strategically use friction at different time points ($p<0.01$) and friction often "slows down" conversations ($p=0.1$).

Theorems & Definitions (1)

• Definition 1: Positive Friction