Should an AI's slow-down mode be a toggle you can flip mid-task, not one fixed setting for the whole job?

This explores whether people want to switch AI friction on and off partway through a task, as their goals shift, instead of accepting one fixed setting. The short answer is that no study in the corpus has asked users this directly or tested a toggle. But several findings point toward it, and taken together they suggest the real issue is timing, not whether friction is present at all.

The strongest clue comes from a 24-person study that added six kinds of turn-level friction to a chatbot Does friction in AI chatbots actually improve what users remember?. Friction reliably made people feel more ownership of their work. It also reliably made the task slower and more effortful. Gains in what people remembered were patchy, though. They showed up only when the friction fit what the person was trying to do at that moment. So the same speed bump helped at one point in a task and was just a cost at another. If a friction's value depends on the user's goal at that moment, a fixed design will often be wrong, and that is the case for letting the user control it. It also helps to know there are three kinds of friction to control: asking the user to contribute something, holding back access to the AI's output, and reshaping the output so the user has to work with it How can designers slow down AI chatbot interactions productively?. A toggle could act on any one of these separately.

The cost of getting it wrong shows up in a study of 199 people using "cognitive forcing functions," which make people commit to their own judgment before they see the AI's answer Do forcing functions reduce AI overreliance without harming team performance?. These designs did cut blind reliance on the AI. Participants still rated them the least favorable of all the designs tested, and the human-AI teams still did worse than the AI working alone. Imposed friction can work and still be resented. That gap between effect and acceptance is exactly what a user-controlled toggle might close, though no one has tested whether it does.

The closest working example comes from a neighboring problem: how much independence to give an AI agent. Designers are settling on an "autonomy slider" that users adjust per task, moving between suggest-only, draft, and execute Should agent autonomy be a continuous slider or fixed levels?. That is a friction toggle under another name, since "suggest-only" forces you into the loop and "execute" removes you from it. Human Factors researchers make a similar case for personalization and for clear handoffs of who does what between person and system Can Human Factors design fixes restore GenAI productivity losses?. Both proposals share the same gap: they are shared vocabulary and design ideas, not tested results. Even the idea of "micro-frictions" tuned to a social goal is still only a hypothesis Can micro-frictions boost rivalry without harming collaboration?.

The unexpected part is that friction isn't always something designers add. Some of it is already built into chat, such as having to type out exactly what you want. One argument is that interface elements like canvases, style lenses, and visual intent maps should remove that kind of friction Can task-oriented UI patterns reduce the friction of AI chat interfaces?. So the better question may not be "friction on or off" but which friction to keep. People likely want to drop the friction of putting their request into words and keep the friction that makes them think, and today's interfaces tend to do the reverse.