A survey of 319 workers links confidence in AI with less reported critical thinking. The everyday risk is approving claims you have not verified.
A meeting summary says the launch is confirmed for Friday. The original notes say Friday is possible if testing goes well. It is a small change in wording with a practical cost: someone may promise a customer a date that the team never agreed to. This is the kind of error that makes careless AI use expensive. It does not need to involve a spectacular invention. Removing a condition, confusing a proposed date with a deadline, or assigning a task to the wrong person can be enough. The summary looks finished, so the person sending it spends less time checking the source.
Research on workers' use of generative AI suggests that confidence in the tool is tied to how much checking people report doing. That does not prove that AI is making people less intelligent. It raises a more immediate workplace question: when a tool produces a convincing answer, do we still do the checks that made our work reliable before we started using it?
Hao-Ping Lee and colleagues surveyed 319 knowledge workers and collected 936 examples of AI use. Their CHI 2025 paper found that greater confidence in AI was linked with less reported critical thinking. Confidence in one's own skill was linked with more. Workers also described a change in the work: checking results, fitting them into a task, and keeping watch over the process. This was a survey about what people said they did. It did not measure a fall in intelligence or prove that AI caused a lasting loss of skill.
Read the authors' research page. The finding worth keeping is narrow but useful: trust and effort were linked in these reports.
Here is an invented work example. You ask a tool to turn meeting notes into a short update. The notes say a launch may happen on Friday if a test passes. The summary says the launch will happen on Friday. Only a few words changed. The meaning changed a lot. A reader may now book time, tell a customer, or promise something the team has not agreed to. Focus particularly on wording that changes what someone will do. “May launch” and “will launch” carry different commitments. “Payment due” and “payment received” describe different states. These are meaning changes, even when the rest of the summary is accurate.
A useful summary keeps the limits of the source. A polished summary that strips them away creates work for everyone who trusts it.
Choose the checks before reading the generated summary. For meeting notes, a useful list is names, dates, assigned tasks, and commitments. Deciding that list in advance makes it less likely that the finished draft will determine what you bother to verify. For a short research summary, it would be who took part, what the groups did, what was measured, and what the study found. For a price comparison, it would be the full cost and what is included.
This checklist is a practical suggestion rather than a method evaluated by the survey. Its purpose is to identify errors that would matter to the recipient and compare those claims directly with the original material. If the task is too large to check, shrink the claim. It is better to send a modest update you can stand behind than a sweeping one you cannot trace.
A real link can sit beside a false sentence. The source may study adults while the answer talks about children. It may measure feelings while the answer claims a change in skill. It may report a plan rather than a finished result. Open the page and look for the exact support. You should be able to say, in plain words, how the source backs the sentence. If you cannot, mark it as unconfirmed or leave it out.
An official-looking source may still be irrelevant to the claim. A library’s ordinary opening hours, for example, do not establish when it opens on a public holiday. Check that the source covers the exact case in the summary.
Suppose you know a project has four stages. The AI lists five. Pause before deciding it knows something you do not. Compare the list with the actual plan. Maybe it has split one stage into two. Maybe the plan changed. Maybe it invented a step. Each is possible. Your job is to resolve that one mismatch, not choose between total faith in yourself and total faith in the machine. Keep your first notes when you ask for help. They let you see which claims the tool added. Without them, its version can quietly become the only version you remember.
The answer is not to scrutinize every comma in every generated message. Spend that effort on the parts someone will act on: dates, amounts, responsibilities, promises, and claims about what a source says. A missed stylistic improvement is usually less costly than a missed condition attached to a deadline. The survey cannot tell us that every worker who trusts AI becomes careless. It does give teams a reason to make checking an explicit part of the job. If a workflow saves ten minutes drafting but leaves no time to verify the important claims, the saving has come at someone else's expense.
