Does friction in AI chatbots actually improve what users remember?

Chen and Gero built six turn-level friction probes for AI chatbots — intentional constraints added to each chatbot exchange — and tested them against a conventional chatbot in a within-subject study with 24 participants doing short writing tasks. All six friction probes increased workload, task duration, and perceived ownership relative to a conventional AI chatbot, while their effects on recall and recognition were more selective. Their discussion section states it more specifically: "all six friction probes increased perceived ownership relative to ChatGPT, while their effects on memory were more selective: most improved recall, but only Elenchus and Less is More significantly improved recognition of participants' own wording."

The six probes instantiate three mechanisms the authors derive from prior HCI friction research — eliciting user contribution, restricting access, and reshaping system output — two probes per mechanism. Their explanation for why memory gains and participant reactions varied is fit between the friction and the user's current goal: "friction could support involvement when it kept users engaged in formulating, interpreting, or assembling the response, but could instead become obstructive when effort shifted toward satisfying the constraint or managing the interface." The clearest case is Elenchus, a Socratic-questioning probe: participant P6, who had already decided their argument's direction, found continued questioning "interfered with rather than supported the assistance they were seeking," while participants who had not yet formed a position found the same questioning useful. Participants did not simply comply with constraints either; they "adapted by decomposing tasks, bringing their own ideas earlier into the interaction, compressing model responses, or working around the friction to restore familiar chatbot workflows."

This study supplies the test that Can micro-frictions boost rivalry without harming collaboration? says is missing, though for a different population and a different outcome frame — ownership, workload, and memory rather than rivalry and collaboration. Here, friction reliably costs effort and reliably raises ownership, but the added effort does not reliably convert into better recognition or into a uniformly preferred interaction, which complicates any assumption that one micro-friction design would move every user toward the same outcome. It also sharpens the turn-level/sustained-use distinction raised by Do static AI evaluations miss harms that build over time? from the opposite direction: that paper argues static single-turn evaluation misses harms that build up only through sustained interaction, while this paper, built strictly at the turn level, finds in its own discussion that "applying the same friction across turns can overlook how users' goals change over the course of a task," with participants wanting friction they could switch on and off by task stage. The underlying mechanism taxonomy is set out separately in How can designers slow down AI chatbot interactions productively?.

The study used short, bounded writing tasks with 24 participants in a controlled setting, not longer or more naturalistic chatbot use, and the authors are explicit that "these results should not be interpreted as identifying universally effective or ineffective frictions." It does not establish whether the same six probes would hold up, fatigue, or get disabled over repeated real-world use, nor does it generalize beyond writing to other chatbot activities such as coding or search. The implication the excerpt supports is narrower than add friction to curb overreliance: friction is productive only when it keeps users doing work still relevant to their current goal, which argues for friction as a switchable repertoire rather than a fixed default.