Designers add 'speed bumps' so people don't blindly trust AI — but do those same speed bumps make people reject good advice too?

This explores whether friction designed to stop people from blindly accepting AI advice also makes them reject advice that was correct. The short answer from the corpus is yes. In a 400-person study with GPT-4o, three 'cognitive forcing' interventions, such as making people commit to their own answer before seeing the AI's, reliably cut over-reliance. They also pushed people into under-reliance: participants started turning down advice they should have taken. Only one intervention reduced both errors, and only on one type of task Can we design interventions that reduce reliance on bad LLM advice?. Friction doesn't make people better judges of when the AI is right. It mostly makes them more skeptical across the board.

The cost isn't spread evenly, which you might not expect. Forcing functions help people who already enjoy effortful thinking ('high need for cognition') much more than everyone else. So an intervention that looks good on average can widen the gap between careful and casual users Do cognitive forcing functions help all decision makers equally?. The people most at risk of over-relying on AI, the ones who would rather not think hard, are the ones friction helps least. A separate line of work shows that even correct AI interruptions carry a price: they break a person's concentration, and rebuilding focus can hurt overall reasoning more than the suggestion helped Does AI assistance always help reasoning or does it carry hidden costs?. Friction and interruption share this hidden tax.

This matters because the default without friction is not neutral either. LLMs produce fluent, fast, intuitive-sounding answers that line up with known human biases: mistaking the description for the thing itself, treating gut feeling as reasoning, and seeking confirmation. When those biases stack, people drift toward trusting outputs they shouldn't Why do people trust AI outputs they shouldn't?. Long-term passive reliance may also wear down people's own thinking. An EEG study found weaker brain connectivity and worse recall of one's own work among habitual LLM users Does AI assistance weaken our brain's ability to think independently?. So the real design problem is calibration: getting people to accept good advice and reject bad advice, rather than just accepting less.

The more promising answers in the corpus change what the AI offers instead of slowing the user down. A review of 73 friction designs sorts them into three approaches: asking users to contribute something, restricting access, and reshaping the AI's output. The best of these redirect effort toward the user's own reasoning rather than simply discouraging use How can designers slow down AI chatbot interactions productively?. 'Learning to Guide' goes further. The AI never hands over a verdict to accept or reject. It points out which parts of the input matter, and this removes anchoring bias while leaving the decision with the human Can AI guidance reduce anchoring bias better than AI decisions?. A philosophical counterpart argues that AI output should count as one piece of evidence weighed alongside others, not a replacement for judgment Should AI outputs replace or supplement human judgment?.

The takeaway: blanket skepticism has a price, because good AI advice is valuable. In a 1,500-person experiment, even a measurably sycophantic model moved people away from their initial biased leanings, because its advice carried real information Can sycophantic AI advice still push people away from polarized views?. Friction that makes people reject that information throws away the benefit along with the risk. The open question in the corpus is how to build friction that helps people tell good advice from bad, not just say no more often.