Showing workers how an AI got its answer sounds safe — but it can actually make them trust it more, not less.

This explores whether showing workers *why* an AI reached a conclusion is enough to keep them from leaning on it too heavily. The short answer from the corpus is no, and sometimes explanation makes things worse. A systematic review of 35 studies found that explanations often deepen automation bias instead of reducing it. Users with low AI literacy are hit hardest, because an explanation reads to them as a sign of credibility rather than something to check. What actually helped was active engagement: having users verify the output, question it, and do their own thinking before accepting it Does explaining AI decisions actually reduce user overreliance on it?.

Why would more transparency produce more trust? One account says LLMs behave like scaled-up fast, intuitive thinking, and people fall into three traps when using them: mistaking a fluent description for reality, treating intuition as if it were reasoning, and having their existing beliefs confirmed back to them. When these traps occur together they reinforce each other Why do people trust AI outputs they shouldn't?. A polished explanation feeds all three. A related argument says the most dangerous systems are the ones that look competent. Fluent outputs quietly wear down skepticism, and in multi-step workflows responsibility gets spread so thinly that nobody feels it is their job to double-check How do competent systems quietly undermine safety oversight?.

There is also a social side that explanation can't fix. In experiments with over 4,000 people, AI users expected colleagues to see them as less competent and diligent, and they were less willing to disclose that they had used AI Do people fear judgment when they use AI at work?. If people hide their AI use, the open, team-level checking that would catch errors never happens, however good the explanations are. Timing matters too. A field experiment at Alibaba found that human takeover fixed AI failures when it came early, but not once customer frustration had built up When human workers take over from AI, does it actually fix the problem?. Oversight has to be designed into the workflow at the right moment, not added afterward as an explanation panel.

Taking a wider view, over-reliance looks less like an individual's mistake and more like a structural one. Work that is delegated to AI is concentrated in information-heavy jobs Where have workers actually delegated tasks to AI?. Brynjolfsson argues that systems built to automate rather than augment shift power away from workers Does automating labor concentrate power more than augmenting it?. The gradual-disempowerment argument goes further: as human labor stops being necessary, an important check disappears, namely workers who care about outcomes and push back when things go wrong Does incremental AI replacement erode human influence over society?.

The surprising takeaway is that the fix for over-reliance probably isn't better explanations. It is keeping humans doing real cognitive work in the loop: verifying, intervening early, and staying needed. An explanation you only read is passive. Checking the output yourself is what protects you.