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 goaloriented 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.
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.
Friction Extends Dialogue Acts
Friction can be applied within the same dialogue act using different movements. Further, dialogue acts comprising of requests are inherently frictive in nature, since they probe for information about the environment or the user's preferences. The most common forms of friction applied are Probing and Overspecification. Finally, the prevalence of other friction categories depends on the dialogue data and task.
Distribution of 50 utterances sampled from annotated dialogue acts (left) belonging to three dialogue datasets (MultiWOZ, TEACh, PersuasionForGood) 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.
Friction Helps Model User Mental States
Friction movements slow down the dialogue, inducing higher valence and longer, more thoughtful conversations. Further, model errors at inferring user satisfaction tend to decrease for conversations when certain types of friction are identified.
Friction Helps Accomplish User Goals
For both MultiWOZ and ALFWorld, applying friction categories improves task success in goal-oriented conversations. The results demonstrate that incorporating friction can enhance the agent's understanding of user goals, resulting in higher task success.
