Some friction in AI tools makes you think harder; other friction just wastes your time — what's the real difference?

This looks at what makes some friction in AI tools useful, because it gets you to think, remember or stay in control, while other friction is just a cost. The corpus points to one main distinction. Productive friction sends your effort toward your own thinking. Annoying friction takes your effort and gives you nothing back. A review of 73 friction designs across 24 HCI papers sorts the useful kind into three moves: asking the user to contribute something before the AI answers, limiting how much the AI hands over, and reshaping the AI's output so you still have work to do How can designers slow down AI chatbot interactions productively?. The point of all three is to keep you engaged with the task, not to make you use the tool less.

The surprising part is that even well-designed friction always costs something, and the benefit depends on timing. In a 24-person study, six kinds of turn-level friction reliably raised workload, slowed people down and increased their sense of ownership over the work. Better memory showed up only when the friction matched what the person was trying to do at that moment Does friction in AI chatbots actually improve what users remember?. So the same speed bump can help one user and irritate another, depending on whether it lines up with their current goal. Friction isn't productive or annoying in itself; fit decides. Research on proactive agents reaches the same conclusion from another direction. An agent that interrupts at the wrong moment or overrides what the user asked for feels intrusive however smart it is. The fix is what the authors call civility: respecting timing, boundaries and autonomy How can proactive agents avoid feeling intrusive to users?.

Removing friction has its own cost, which is the less obvious half of the answer. Four experiments with more than 3,500 people found that after working with generative AI, people felt more in control when they went back to solo work, but they were less motivated and more bored. The AI had taken over the interesting parts and left them the dull leftovers Does AI collaboration drain motivation when workers return to solo tasks?. Low friction can quietly use up the engagement that made the work worth doing. Writers show something similar: those who both compete with and collaborate with AI report the strongest productivity Does balancing rivalry and collaboration with GenAI boost writer productivity?. The proposal to add small frictions to keep that rivalry alive is still an untested idea, though Can micro-frictions boost rivalry without harming collaboration?.

Some friction really is just waste, and it should be removed. Friedman argues that making users type out what they want in a chat box is pure overhead. He suggests interface patterns such as visual intent maps, style 'lenses' for adjusting output and direct manipulation on a canvas, which let people say what they mean faster Can task-oriented UI patterns reduce the friction of AI chat interfaces?. Bichler adds a warning from economics. When AI agents reduce one kind of friction, such as search costs in markets, new kinds tend to appear, such as distorted information and agents guessing at your preferences Do AI agents actually reduce friction in electronic markets?. Friction tends to move around rather than disappear.

In practice, there are two questions to ask. First, does this friction return the effort to you as thinking, ownership or memory, or does it just slow you down? Second, does it arrive when it fits your current goal? The stakes are also uneven across products. One heuristic evaluation of nine deployed agents found that companion and advisory agents were the weakest at preventing user dependency Do productivity agents actually protect users better than emotional ones?. Those are the settings where well-placed friction is probably needed most.