How can designers slow down AI chatbot interactions productively?

Chen and Gero build a design taxonomy for "turn-level design friction" in AI chatbots by reviewing prior HCI interventions. Searching Google Scholar with terms including "design friction, selfcontrol, self-regulation, constraint, restriction, and lockout" across digital wellbeing, social media, decision support, AI-assisted writing, and programming education, they assembled "an exploratory corpus of 24 HCI papers published within the past five years" containing "73 unique designs," with four designs coded under two mechanisms each, yielding "77 design–mechanism entries." From this corpus they define three mechanisms: "M1. Eliciting user contribution requires users to provide additional information, reasoning, judgment, or action beyond what the default interaction would ordinarily require." "M2. Restricting access delays, gates, or conditionally withholds content or functionality that would otherwise be immediately available." "M3. Reshaping system output changes the amount, form, organization, or communicative role of what the system presents, thereby changing how users must interpret or use it."

Designs that did not clearly fit one of the three were coded as "Other" and not treated as a fourth mechanism, "because these designs did not share a sufficiently coherent interaction pattern." The authors distinguish their aim from social-media and digital-wellbeing friction, where discouraging continued use can itself be the goal: "Our goal was not to discourage chatbot use as an end in itself, but to reshape how users engaged with AI assistance." A friction only counted as productive "when it redirected users' effort or attention toward the task, their own reasoning, or the model-generated content, rather than merely making the interaction more difficult or unpleasant." Guided by M1–M3, the authors then built six probes, two instantiating each mechanism, which they tested empirically in the companion study.

The taxonomy sits alongside other recent attempts to name design spaces for structuring AI-chatbot interaction, but at a different grain. Can task-oriented UI patterns reduce the friction of AI chat interfaces? names five levers — intent capture, output shaping, refinement, agentic tasks, workflow embedding — for moving users beyond chat into task-oriented UI; M1–M3 instead describe how to slow down what happens inside a single chat turn, without leaving the chat modality. It also differs from Can micro-frictions boost rivalry without harming collaboration? by naming concrete, corpus-derived mechanisms and building testable probes from them, rather than proposing an untested micro-friction in the abstract. The empirical consequences of applying M1–M3 to real chatbot use are reported in Does friction in AI chatbots actually improve what users remember?.

The corpus review is a qualitative clustering of 73 designs drawn from author-chosen search terms, not a systematic or exhaustive review, and the paper reports no inter-rater reliability figure for the M1–M3 coding beyond noting the four cross-coded designs. The taxonomy also describes mechanisms for adding friction to a single turn, not for removing it or scheduling it across a task; the companion study's own discussion flags that a fixed turn-level mechanism can miss how a user's goal changes over the course of a task. The implication is that M1–M3 is a useful vocabulary for comparing friction designs, not a validated causal account of which mechanism produces which user effect — that mapping is what the six-probe study tests, and only partially confirms.