Chatbots that add small 'speed bumps' feel harder to use — but does that friction actually help you remember more?

This explores whether deliberately slowing chatbots down, with small design 'speed bumps' inside each back-and-forth, makes people retain more and learn better, or whether it just makes the work feel harder. The short answer from the corpus: friction reliably changes how the interaction *feels*, but it only sometimes changes what people remember. The most direct evidence comes from a 24-person study that tested six kinds of turn-level friction. Every probe consistently raised workload, made tasks take longer, and gave people a stronger sense that the work was theirs. Memory gains were selective, though. They showed up when the friction matched what the user was trying to do at that moment and faded when it didn't Does friction in AI chatbots actually improve what users remember?.

That result makes more sense once you see that 'friction' isn't one thing. A review of 73 friction designs across 24 HCI papers sorts them into three mechanisms. Some ask the user to contribute first, for example by drafting an answer before the AI does. Some restrict access, such as holding back the full answer. Some reshape what the AI outputs, like giving hints instead of solutions How can designers slow down AI chatbot interactions productively?. Each one pushes your effort in a different direction. A prompt to write your own reasoning first helps when you're trying to understand something. It's just an obstacle when you only need a quick fact. This is likely why the memory effect depends on 'task fit': the friction has to point your effort at the thing you need to retain.

Most of the field pushes the other way. One line of work treats fewer turns as the goal and shows that proactive assistants, which offer information before you ask, can cut conversations by up to 60% Could proactive dialogue make conversations dramatically more efficient?. A critique of chat-first interfaces argues that for structured, goal-directed work, conversation already adds mental overhead compared with clicking buttons Is chat the wrong default interface for AI products?. Taken together, this suggests a split. Efficiency is the right goal when the value lies in the output. Friction earns its cost only when the value lies in what changes in the person. Pacing research adds a twist: an agent that varies its response timing to fit the conversation, rather than replying at a fixed speed, made people feel better listened to and led them to share more about themselves in support chats Can pacing strategies make conversational agents feel more human?. So slowing down isn't only about learning. It can also change how much people open up.

The corpus doesn't yet show whether these effects last. The friction evidence comes from single sessions, and longitudinal chatbot studies show that effects measured in one session (novelty in particular) fade predictably as people keep using a system. That's a warning against assuming any first-session effect will hold up over weeks Do chatbot relationships lose their appeal as novelty wears off?. One more distinction is easy to miss: 'remembering' in chatbot research usually means the *AI's* memory, not yours. Even there, users prefer friction at the moment of capture, deciding upfront what gets stored, over auditing records afterward Do users prefer controlling what gets remembered over auditing what was stored?. Whether the memory belongs to the person or the machine, a well-placed pause at the right moment seems to do more than review after the fact.