Tools that nudge people to double-check AI answers mostly help those who already like to think hard — can they be redesigned to help everyone else too?

This explores whether cognitive forcing functions (small pieces of friction, like making you commit to your own answer before you see the AI's) can be redesigned so they help everyone, not only people who like to think hard. The short answer: the corpus documents the problem clearly but has no tested redesign. What it does have are some useful clues about why the problem exists.

The problem first. In a 199-person experiment, forcing functions cut overreliance on AI on average, but most of the gain went to people high in 'Need for Cognition,' the trait of enjoying effortful thinking Do cognitive forcing functions help all decision makers equally?. So an intervention meant to protect people from blindly trusting AI ended up widening the gap between careful thinkers and everyone else. The same study found a second problem: participants rated the forcing designs least favorably of all the conditions, even though they worked best, and human+AI teams still did worse than the AI alone Do forcing functions reduce AI overreliance without harming team performance?. Together these findings point to the core design tension. Friction works only if people engage with it, and the people least inclined to engage are the ones who most need the help and who most dislike the friction.

A lateral clue comes from health chatbots. When three major LLMs were tested on behavior-change conversations, they did well with users who already had clear goals but failed to notice ambivalence or low readiness to change Why can't chatbots detect when users are ambivalent about change?. That is the same failure in a different setting: support built for the already-motivated quietly fails everyone else. It suggests one direction for a redesign. Forcing functions today apply the same friction to everyone. A better version might first figure out where a person is (engaged, skeptical, or just trying to get through the task) and adjust. But that study also shows current models are poor at reading those states, so 'just make it adaptive' is not a solved problem.

A more speculative angle comes from research on training AI models, where scalar scores turn out to carry only part of the useful signal. Feedback contains an evaluative part (how well did that go?) and a directive part (what should change?), and these do different jobs Can scalar rewards capture all the information in agent feedback?. Models stuck on a plateau improve when they're told why they failed rather than simply that they failed Can natural language feedback overcome numerical reward plateaus?. Treat this as an analogy, not evidence about people. Still, a typical forcing function only makes you pause, which is a kind of bare evaluative signal. For someone who won't generate their own reasoning during that pause, a nudge that names what to check might do more than a blank 'think first' gate.

The surprise worth taking away: making people think harder is not neutral. It can become an inequality machine, rewarding the people who were already doing fine. Whether a redesign can fix that is still an open question in this collection.