Scores and ratings showed no change when AI nudged people to think critically — but did their actual thinking shift?

This explores whether looking closely at the thinking process (what people say, question, and notice while working with AI) shows critical-thinking effects that headline numbers miss. The corpus has one study that tests this directly, and its answer is yes. In a 24-person study, short critique prompts were attached to AI suggestions, nudges like 'here's a weakness in this recommendation' Can brief critiques help users think critically about AI suggestions?. On the quantitative measures (how diverse people's shortlists were, how often they reflected, how much they trusted the AI) nothing changed significantly. When researchers coded what participants actually said, though, the critiques had triggered critical and metacognitive thinking at several levels, from noticing problems to reflecting on their own reasoning. Judged by the numbers alone, the intervention failed. Judged by the transcripts, it worked.

This matters because most of the alarming evidence about AI and critical thinking comes from the quantitative side. A 666-person UK survey found that heavier AI use correlated with lower critical-thinking scores, with cognitive offloading (handing mental work to the tool) as the measured mechanism Does frequent AI tool use reduce our critical thinking?. A review of 14 studies describes students swapping effortful reasoning for quick shortcuts Does relying on AI dialogue systems harm student critical thinking?. These findings are real, but they measure end states. If small design changes can stir up real thinking that never reaches the final output, then score-based studies may underestimate how much good interface design can help.

The less obvious lesson comes from AI research itself, where the same gap between outcome and process keeps appearing. Chain-of-thought prompts built on logically invalid reasoning score almost as well as valid ones, so accuracy alone can't tell you whether any real reasoning happened Does logical validity actually drive chain-of-thought gains?. Models trained to imitate ChatGPT fooled human evaluators with confident, fluent style while gaining no factual ability Can imitating ChatGPT fool evaluators into thinking models improved?. Reasoning-trained models look broadly smarter on benchmarks but lose the ability to recognize a badly posed question and stop What critical thinking skills do reasoning models actually lose?. In each case a single number hid what was happening underneath. Taking that number apart, as one study did by splitting chain-of-thought performance into three separate factors, showed what the aggregate concealed What three separate factors drive chain-of-thought performance?.

One practical alternative to choosing between numbers and transcripts is to measure along several dimensions. Work on prompt quality Can we measure prompt quality independent of model outputs? and on training models to ask better clarifying questions Can models learn to ask genuinely useful clarifying questions? finds that breaking quality into named attributes catches improvements a single score misses. Applied to people, critical thinking would be measured as several separate behaviors rather than one score. A caveat: the human-side evidence here is one small study. The corpus strongly suggests qualitative methods catch effects that metrics miss, but it does not yet show whether those coded moments of thinking last or change later decisions.