Generative AI is becoming part of everyday knowledge work, but a study from Microsoft and Carnegie Mellon University points to a tradeoff: the more workers lean on AI tools, the more their role may move away from solving problems directly and toward verifying AI output.
The study surveyed 319 knowledge workers and collected 936 real-world examples of generative AI use across IT, design, administration, and finance. Its central concern is practical: AI can make work easier, but convenience may reduce the chances people get to exercise critical thinking.
What the Study Examined
The research looked at six categories of critical thinking: knowledge, understanding, application, analysis, synthesis, and evaluation. Those categories matter because they cover more than whether a worker can spot a bad answer. They also describe how people gather information, connect ideas, apply judgment, and decide whether an output is good enough to use.
According to the study, generative AI is changing the shape of those activities. Workers are not simply using a new tool to do the same old job. In many cases, the task itself becomes different once AI is involved.
The researchers identified three major shifts:
- Instead of gathering information independently, workers increasingly focus on verifying AI outputs.
- Instead of developing their own solutions, workers integrate AI-generated answers.
- Instead of executing tasks directly, workers monitor AI systems.
That does not mean every use of AI weakens thinking. The source study does not make that broad claim. Its warning is narrower and more useful: when AI handles too much of the thinking process, people may get fewer chances to practice the skills that make them effective problem solvers.
The Convenience Problem
The study describes a hidden cost in routine or less critical tasks. When the stakes feel low, people may be more likely to accept AI help without much questioning. Over time, that pattern can matter because everyday work is also where judgment is practiced.
The researchers describe this risk as "raising concerns about long-term reliance and diminished independent problem-solving." They also refer to an "irony of automation": by taking over mundane work, AI tools can remove the very moments where people would normally use their judgment and strengthen their "cognitive muscles."
The source article uses the term "cognitive offloading" for this process. In plain language, that means shifting thinking to an outside system. A calculator, a search engine, or an AI assistant can all reduce mental effort in some way. The issue raised here is whether generative AI shifts too much of the process, especially when workers stop asking hard questions about the output.
This matters because verification is not the same as original problem-solving. Checking an answer can require skill, but it may still start from the AI system’s framing. If the worker accepts that framing too quickly, important alternatives may never be considered.
Confidence, Motivation, and Barriers
The study found that self-confidence may offer some protection. Workers who feel more confident in their own abilities tend to be more skeptical of AI outputs. The researchers did not establish a definitive causal relationship, so the finding should be read carefully.
Still, the pattern is important. A worker who trusts their own judgment may be more willing to challenge an AI answer, ask for a better response, or reject the output entirely. A worker who feels less certain may be more likely to accept what the system provides, especially under pressure.
The researchers identified three main factors that drive critical thinking when people use AI:
- The desire to improve work quality.
- Error avoidance.
- Personal development.
Those motivations are straightforward. People question AI when they want better work, when they want to avoid mistakes, and when they want to grow their own skills. But the study also points to barriers that make critical thinking harder in practice.
Time constraints are one barrier. If a worker is rushing, verification can become superficial. Lack of problem awareness is another: if someone does not know what can go wrong, they may not know what to check. A third barrier is the difficulty of improving AI responses in unfamiliar domains. When the topic is outside a worker’s expertise, it is harder to judge whether the AI is producing a strong answer or simply a fluent one.
What Companies Can Do
The researchers recommend that companies actively promote critical thinking among employees. That means training workers on how to review AI results, not just how to prompt a tool or move faster through tasks.
This recommendation changes how organizations should think about AI adoption. The goal is not only access to generative AI. It is also the creation of work habits that keep human judgment engaged.
The study also suggests that AI tools should be designed to support critical questioning rather than replace it. That point is central. If a tool makes answers look final, workers may treat them as final. If a tool encourages review, comparison, and reflection, it can keep the user more actively involved.
For employers, the practical lesson is that efficiency should not be the only measure. A workflow that produces fast AI-assisted output may still create risk if people lose the habit of asking whether the answer is complete, relevant, and correct.
Why Younger Users May Need Extra Attention
The source article also cites a separate study by the Swiss Business School involving 666 participants. That study reported similar findings in January. Young people aged 17-25 showed the highest AI tool usage while scoring lowest on critical thinking tests.
Education level appeared to be a significant protective factor in the Swiss study. Participants with higher education questioned AI-generated information more frequently and maintained stronger critical thinking skills despite using AI tools.
Together, the two studies point in the same direction: AI use is not automatically harmful, but uncritical reliance can become a problem. The people and organizations that benefit most from generative AI may be the ones that treat it as a tool to interrogate, not an authority to obey.
The future of knowledge work may depend less on whether workers use AI and more on how they stay mentally active while using it. Verification is necessary, but it should not become the whole job.