People trusted an AI more when they saw its full reasoning — even though reading it made their own answers worse.

This explores why showing people an AI's complete reasoning makes them feel more sure of themselves while they actually do worse. The core finding comes from a 559-person study. People who saw full reasoning traces performed worse on tasks and overestimated their own abilities more. People who saw short summaries did as well as people who saw no trace at all Do reasoning traces help or hurt people solving problems?. The most striking detail is that no format, full or summarized, helped people judge how well they had done. So traces don't just fail to inform people. Long ones actively mislead them.

The corpus offers one strong explanation: traces are built to help the model, not the reader. A separate 100-person study found that the traces most useful for model accuracy were the ones people rated hardest to follow. Those same traces made people more likely to accept wrong answers Do chain-of-thought traces actually help users understand model reasoning?. Recursion, self-revision and backtracking are good training signals but tiring to read. The likely mechanism is that a long, effortful-looking trace signals that real thinking happened. Readers treat that signal as evidence, even when they can't actually check the logic.

This pattern isn't limited to reasoning traces. When researchers redesigned an interface for inspecting agent execution traces, users found errors faster but no more accurately. Users who missed an error became *more* confident in the wrong answer Does a faster trace interface actually improve error detection?. Across these studies, the feeling of having inspected the reasoning seems to replace actual inspection. Part of the problem is that not all sentences in a trace matter equally. A few planning and backtracking sentences steer most of the reasoning Which sentences actually steer a reasoning trace?. A human reading the whole trace can't easily tell which lines are the pivots, so the important parts get lost among the rest.

The corpus suggests two ways out, and neither one is 'show more.' One is to ground explanations in what the agent actually did. Structured reports built from execution traces catch unsupported claims that fluent, natural-sounding explanations hide Can execution traces ground honest explanations of agent behavior?. The other is to stop treating a trace that looks like reasoning as proof of reasoning. Benchmark designers have found that scoring reasoning steps inflates results because models learn to imitate the style of reasoning Should reasoning benchmarks score final answers or reasoning traces?. Checking specific intermediate steps against rules works much better than judging the overall impression Where do reasoning agents actually fail during long traces?.

The takeaway you might not expect: AI evaluators fall for the same illusion humans do. Reasoning that looks like reasoning is persuasive to people and to scoring systems alike. Transparency only helps if what's shown can actually be checked. The corpus doesn't directly test *why* individual users trusted long traces, for example whether it's the length or the apparent effort. That part is still an inference from these studies.