Research on how people judge AI answers found that extra explanation could increase confidence without improving error detection.
An AI answer can include definitions, examples, and a careful-sounding explanation while getting the one fact you need wrong. If you ask about a train on Sunday, an accurate history of the railway does nothing to confirm the departure time. Yet that surrounding detail can make the answer feel researched. This matters because many people judge reliability partly through presentation. A complete explanation suggests that the writer understands the subject. With a language model, fluent explanation is much easier to produce than reliable evidence for every claim in it.
Research by Mark Steyvers and colleagues examined this gap between how accurate AI answers were and how accurate people believed them to be. Longer replies could increase confidence without helping readers identify the errors. The practical response is to separate the claim you need from the explanation around it, then check the claim against something outside the conversation.
Mark Steyvers and colleagues studied how people judged answers from language models. Their paper came out in Nature Machine Intelligence in 2025. People thought the AI was right more often than it was. Long replies could make people feel more sure, but the extra words did not help them spot wrong answers. The study concerns judgments of answers under its tested conditions. It does not mean every long answer is bad or every short one is honest. It shows that length can act as a cue for trust without earning that trust through better error detection.
The full paper is What large language models know and what people think they know. Keep this question nearby. Do I know the answer is right? Or does it just feel complete?
Imagine asking whether a train runs on Sunday. The answer gives the line's history, its usual route, and a list of stations. It then says the Sunday service starts at six. The only part you need for the trip may be that last claim. The history can be accurate while the time is wrong. None of the surrounding detail makes the timetable current. Reduce the answer to the claim your trip depends on: the train leaves a particular station at a particular time on that date. Compare that with the operator’s timetable. Accurate background information cannot substitute for this check.
For any long response, identify the statement you would act on and what would count as evidence for it. This makes verification manageable even when the surrounding explanation is extensive.
Suppose a shop cuts a price by 20% and then raises the new price by 20%. The AI says you are back at the starting price. It explains that the two changes cancel out. That explanation is incorrect because the two percentages use different starting amounts. A price of £100 falls to £80 after a 20% discount. Raising £80 by 20% adds £16, producing a final price of £96. Working through actual numbers exposes the error more clearly than requesting another general explanation. The critical question is which amount each percentage applies to; once that is clear, the conclusion can be checked directly.
This is an invented example. It shows why a chain of reasons is only useful when the steps themselves can be tested.
Instead of asking the tool to explain more, ask it to show the one step you need to inspect. For a number, show the input and the calculation. For a claim about a study, show the relevant result. For a rule, show the source and the part that applies. A useful answer makes the check easier. An answer that adds five more broad claims makes the job harder. Be careful with requests for confidence scores. A model saying it is very sure does not give you a new source. Nor does a second reply from the same model count as an independent check.
When the fact can change, use a current source. A correct explanation of how a service usually works may still send you to a closed office.
Long answers often glide over the very step that matters. They may say therefore without showing why. They may use a study of one group to make a claim about everyone. They may give an average with no sign of how much people differed. Mark the point where the answer jumps. Ask what would have to be true for that jump to work. In the price example, the changes would need to use the same starting amount. They do not. You can use this habit on human writing too. A confident voice is not a special weakness of machines. AI simply makes it cheap to produce that voice at length.
For a factual question, ask for the answer and its source before requesting a long explanation. For a calculation, inspect the numbers and the steps. For a changing fact such as a timetable, check the service's current information. These checks address the thing you need to know rather than the style in which the answer was delivered. Length is useful when it resolves a difficulty, supplies needed context, or explains a step you can verify. It is less useful when it repeats the same claim in increasingly confident language. The study gives a concrete reason to avoid treating a detailed answer as evidence that the central fact has been checked.
