Does slowing people down before an AI gives advice actually make them think more — or just make them like the tool less?

This explores whether friction built into AI decision support, meaning features that slow people down so they think before accepting a recommendation, costs users their sense of engaged flow. The corpus has no study that measures 'immersion' directly. What it does have is a close substitute: how much people like these designs. On that measure the answer is a clear yes. In a 199-person study, 'cognitive forcing functions' (asking users to decide on their own first, or making them wait for the AI's answer) cut overreliance on AI more than standard explainable-AI displays did. Participants also rated these designs the least favorably of everything they tried Do forcing functions reduce AI overreliance without harming team performance?. People preferred the designs that protected them least.

The less obvious part is that the friction did not make the human-AI team better. Teams still did worse than the AI working alone. So friction shifted *who* did the thinking without improving the result. That turns the question from 'does friction hurt the experience?' into 'what is the friction buying?' Here it bought less blind copying of the AI, at a cost to how pleasant the tool felt to use, and with no gain in accuracy.

The cost also falls unevenly. The same experiment found that forcing functions helped people who enjoy effortful thinking (high 'Need for Cognition') much more than people who don't Do cognitive forcing functions help all decision makers equally?. One plausible reading is that people who like thinking treat the pause as an invitation, while everyone else experiences it as an interruption. If so, one fixed amount of friction can't serve everyone, and the users who most need protection from overreliance get the least help from it.

A nearby line of work points to an alternative to constant friction: friction that adapts to the person. Research on behavioral signals shows that AI systems can estimate a user's mental state from gaze, typing hesitation, and interaction speed. That lets a system time its interventions without stopping people to ask questions, which keeps their flow intact Can AI systems read cognitive state from interaction patterns alone?. In principle, a decision tool could apply friction only when someone seems to be rubber-stamping the AI's answer and stay out of the way otherwise. The same research gives a warning, though. The signals that make well-timed help possible also make detailed, manipulative profiling possible.

The broader lesson is that the question is less about friction versus immersion than about where the friction lives. Flow probably matters most for experienced users and for the 'creation tasks' where human-AI teams do well, and constant friction is a blunt instrument. The next study the corpus doesn't yet have would measure immersion directly while testing adaptive, signal-triggered friction against the fixed version.