AI Is Making Us Think Less Without Us Noticing

The Machine That Makes You Feel Smart

Here is a paradox that should unsettle you: the more you trust an AI, the less you think, but the more confident you feel about your own thinking.

This is not a philosophical provocation. It is a measured finding from a 2025 survey of 319 knowledge workers, published in a peer-reviewed venue, which collected 936 firsthand accounts of people using generative AI at work. The researchers, Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch, and Ian Drosos, asked a straightforward question: When people use AI to do their jobs, do they still engage in critical thinking, or do they outsource the effort?

The answer is disturbing. People who reported higher confidence in generative AI also reported putting less cognitive effort into their work (Lee et al., 2025). They did not notice the drop. They felt they were producing better results with less strain. And they were right about the strain. But the quality of their thinking? That was another matter.

The Confidence Trap

The study measured two kinds of confidence: confidence in oneself and confidence in the AI tool. These two variables pulled in opposite directions.

Higher self-confidence predicted more critical thinking. People who trusted their own abilities were more likely to question the AI's output, verify its claims, and integrate its suggestions with their own knowledge. They treated the AI like a junior colleague whose work needed checking.

Higher confidence in the AI predicted less critical thinking. People who believed the tool was reliable stopped asking questions. They accepted outputs at face value. They reduced the mental effort they applied to tasks they previously would have thought through carefully (Lee et al., 2025).

This is the trap. The better the AI gets, the less we scrutinize it. The less we scrutinize it, the more our critical thinking muscles atrophy. And the whole process is invisible to us, because the AI produces plausible answers that feel correct.

How 319 People Actually Use AI at Work

The researchers recruited knowledge workers through Prolific, a platform for academic studies. The participants came from a range of industries: technology, consulting, education, healthcare, marketing. Each person described between one and three specific instances where they used generative AI for a work task. In total, the researchers collected 936 examples.

This is important. The study did not ask people to guess how they might use AI. It asked them to report what they actually did. That gives the findings a grounded quality that hypothetical surveys lack.

The researchers coded each example for whether the person engaged in critical thinking during the task. They defined critical thinking as "the careful, deliberate application of one's cognitive abilities to evaluate, analyze, and synthesize information." They measured effort, verification behaviors, and the nature of the thinking that occurred.

What they found was a shift in the type of critical thinking people did. It was not that critical thinking disappeared entirely. It changed shape.

The Three Remaining Modes of Thought

When people used AI, their critical thinking narrowed to three specific activities (Lee et al., 2025):

• ▸Information verification: Checking whether the AI's output was factually correct. This sounds good, but it often meant people spent their cognitive energy fact-checking rather than generating ideas.

• ▸Response integration: Figuring out how to combine the AI's output with their own work. This is a logistical task, not a creative one.

• ▸Task stewardship: Managing the interaction with the AI itself: writing prompts, selecting the right tool, deciding when to trust the output.

Notice what is missing from this list. Generating original hypotheses. Questioning assumptions. Connecting disparate ideas. Recognizing when a problem is framed poorly. These are the higher-order cognitive skills that make knowledge work valuable. They were systematically reduced.

The researchers describe this as a "shift in the nature of critical thinking." A more honest description might be a narrowing. People stopped thinking about what to think and started thinking about whether the AI had thought correctly.

The Illusion of Effortless Mastery

There is a particularly insidious finding buried in the qualitative responses. Several participants reported that using AI made them feel more competent, even as they exerted less cognitive effort (Lee et al., 2025).

This makes sense if you think about how the brain works. Cognitive effort feels bad. It is metabolically expensive. It produces uncertainty and doubt. When a machine removes that effort, what remains is a smooth feeling of confidence. You see the output. It looks good. You did not have to struggle to produce it. Therefore, you must be good at this.

The researchers found that this feeling of confidence was a reliable predictor of reduced critical thinking. The people who felt the most confident about their AI-assisted work were the ones who had thought the least about it.

This is not a failure of individual willpower. It is a feature of how human cognition interacts with reliable tools. We are wired to conserve mental energy. When a tool consistently produces good results, we stop questioning it. That is rational in the short term. It is dangerous in the long term.

What the Study Does Not Prove

Before we declare a crisis of critical thinking, we should acknowledge what this study does not show.

It does not show that AI causes a permanent decline in critical thinking ability. The study measured self-reported effort in specific tasks, not long-term cognitive decline. It is possible that people can switch critical thinking back on when they need to, like a muscle they choose not to use for routine tasks.

The study also relies on self-report. People are notoriously bad at accurately describing their own cognitive processes. Someone who says they "thought carefully" about an AI output may have done less thinking than they imagine. Conversely, someone who says they "just used the output" may have done more unconscious evaluation than they realize.

And the sample, while decently sized at 319 participants, is not representative of all knowledge workers. The participants were recruited from a platform for online studies, which skews toward people who are comfortable with digital tools and may be more trusting of AI than the general population.

These are real limitations. But they do not weaken the core finding. They simply define its boundaries. Within those boundaries, the evidence is clear: people who trust AI reduce their cognitive effort, and they do not seem to notice.

The Design Problem

The researchers frame their findings as a design challenge. They argue that current generative AI tools are optimized for one thing: producing outputs that satisfy the user. The tools reward acceptance. They penalize scrutiny.

Think about the interface of ChatGPT or Claude. There is a text box. You type a prompt. The AI generates a response. The default behavior is to accept the response. There is no built-in mechanism that asks, "Are you sure this is correct?" There is no friction that encourages you to verify, to question, to push back.

The researchers point out that this design creates a perverse incentive. The more fluent and confident the AI's output appears, the less likely the user is to engage in critical thinking. The tool is designed to be trusted. Trust reduces effort. Reduced effort reduces critical thinking. The system optimizes for the wrong outcome.

They suggest new design directions: tools that prompt users to verify key claims, interfaces that surface uncertainty, systems that require users to articulate their reasoning before accepting an output (Lee et al., 2025). These are not anti-AI suggestions. They are pro-human ones.

The Organizational Risk

There is a broader implication that the researchers do not fully explore. If individual knowledge workers are reducing their critical thinking, what happens to organizations that depend on that thinking?

Companies are deploying generative AI at scale. They are measuring productivity gains: faster writing, quicker code generation, more efficient data analysis. What they are not measuring is the slow erosion of judgment. The quiet acceptance of plausible but wrong outputs. The gradual loss of the skill of thinking from first principles.

A marketing team that uses AI to draft copy may produce more content in less time. But if nobody on the team is checking whether the copy is strategically sound, the volume comes at a cost. A software team that uses AI to write code may ship features faster. But if nobody is questioning whether the feature should exist in the first place, speed becomes a liability.

The researchers found that critical thinking shifted toward verification and integration. These are second-order tasks. They assume the first-order thinking has already been done by the AI. But AI does not think. It generates patterns based on data. It has no understanding of context, strategy, or long-term consequences.

When organizations outsource first-order thinking to machines, they are not just saving time. They are losing the capacity to think differently than the data suggests.

What This Actually Means

The study by Lee and colleagues is not a prediction of doom. It is a measurement of what is already happening. The question is what to do with that information.

• ▸If you use AI, treat every output as a draft from a junior employee. Verify the claims. Question the framing. Ask yourself: would I have written this differently? The researchers found that people who maintained self-confidence engaged in more critical thinking. That confidence comes from practice. You have to keep thinking to stay good at it.

• ▸Design your workflow to include friction. The researchers suggest interfaces that prompt verification. You can do the same thing manually. Before you accept an AI output, write down one thing you think might be wrong. Force yourself to find a mistake. If you cannot find one, you have not looked hard enough.

• ▸Watch for the feeling of effortless confidence. That feeling is a warning sign. It means you are not thinking. The researchers found that confidence in AI predicted reduced effort. When the work feels too easy, assume you are missing something.

• ▸Organizations should audit for critical thinking, not just productivity. If your team is producing more output but asking fewer hard questions, you have a problem. Measure whether people are verifying, questioning, and integrating AI outputs. If they are not, the efficiency gains are hollow.

• ▸Teach people how to use AI without losing their cognitive edge. This is a new skill. It is not about prompting better. It is about maintaining the habit of independent thought while using a tool that makes independent thought feel unnecessary.

The researchers end their paper with a call for new design paradigms. That is fair. But the deeper lesson is personal. AI is not making us dumber. It is making us feel smart enough that we stop trying to be smarter. The difference is small. The consequence is not.