Generative AI is becoming a routine part of work for people who write, research, summarize, analyze, and ask for advice. A recent study by researchers from Microsoft and Carnegie Mellon University looked at what that shift may mean for critical thinking.
The central concern is practical: when workers let AI produce the first answer, their mental effort can move from solving the problem to deciding whether the AI output is acceptable. That may sound efficient, but the study argues that this change can reduce the regular practice people need to keep their own judgment sharp.
What the study examined
The study focused on 319 people who said they used generative AI at least once a week at work. Each person was asked to describe three examples of how they used the technology in their job.
Those examples fell into three broad categories. Some involved creation, such as writing a formulaic email to a colleague. Others involved information work, such as researching a topic or summarizing a long article. A third group involved advice, including asking for guidance or making a chart from existing data.
For each task, respondents were asked whether they used critical thinking skills. They were also asked whether generative AI made them use more or less effort to think critically. The researchers also asked how confident respondents were in themselves, in generative AI, and in their own ability to evaluate AI outputs.
That design matters because the study was not only asking whether people used AI. It was asking what kind of thinking remained after AI entered the workflow.
The shift from thinking to checking
The study describes a change in where human effort goes. Instead of spending as much time creating, evaluating, and analyzing information, workers may spend more of their effort verifying that an AI response is good enough to use.
Verification is still a form of judgment. A worker who checks an AI-generated answer is not necessarily acting passively. But the paper warns that a workflow built around checking can leave fewer routine chances to practice deeper problem-solving.
The risk becomes clearer when AI works most of the time but fails at moments that matter. If people only step in when the AI answer is inadequate, they may have had fewer opportunities to build the judgment they need when those exceptions appear.
Put simply, the concern is not that using AI once weakens a worker's mind. The concern is that repeated reliance can change habits. When the tool handles the first pass, the human may get less practice doing the harder intellectual work from the start.
When workers still use critical thinking
The study found that about 36% of participants reported using critical thinking skills to reduce possible negative outcomes from AI use. These cases show that workers often recognize AI output as something that must be handled carefully.
One participant used ChatGPT to write a performance review, then checked the result because she worried that submitting the wrong thing could get her suspended. Another respondent used AI-generated emails for messages to his boss, but edited them because the boss's culture placed more emphasis on hierarchy and age, creating a risk of a faux pas.
Other participants checked AI-generated responses through general web searches, including resources such as YouTube and Wikipedia. That extra verification may reduce risk, but it can also weaken the original efficiency argument for using AI. If a worker must use broad web searches to confirm the output, the AI may be adding another step rather than removing one.
These examples point to a broader pattern: AI can speed up a draft or suggestion, but the user still carries responsibility for the outcome. The more sensitive the task, the more important it becomes to know what could go wrong.
Confidence changes the amount of effort
The study also found an important link between confidence and critical thinking. Participants who reported confidence in AI used less critical thinking effort than those who reported confidence in their own abilities.
That finding suggests a subtle danger. If a worker trusts the system too much, they may examine its output less carefully. If they trust their own ability to evaluate the situation, they may be more likely to question, revise, or verify what the AI produces.
The paper also notes that workers need to understand the shortcomings of generative AI in order to compensate for them. Not all participants were familiar with those limits. That matters because possible downstream harms can only prompt critical thinking when the user is aware those harms may exist.
In workplace settings, this creates a clear implication. Critical thinking depends not only on access to AI, but on understanding when AI output might be incomplete, inappropriate, or risky to use as-is.
The real warning for AI at work
The researchers do not make the simple claim that generative AI tools make people dumber. The study is more careful than that. Its warning is about overreliance and the gradual weakening of independent problem-solving when workers stop practicing it regularly.
For employers and workers, the message is not to avoid generative AI altogether. The source material supports a narrower point: AI-assisted work should still leave room for people to create, evaluate, and analyze, not merely approve or reject machine-generated output.
A healthier workflow would treat AI responses as material to examine, not as a finished answer by default. The worker's role remains central because context, judgment, and awareness of consequences still determine whether an output is actually fit for use.
Generative AI can help with workplace tasks, but the study shows why convenience has a cognitive cost when it replaces practice. If people want to stay ready for the moments when AI is wrong or insufficient, they need regular chances to exercise the very skills the tool can make easier to skip.