Does AI Hurt Critical Thinking?

Does using AI hurt critical thinking?

In 2026, Xiaoxiao Sun and colleagues published a meta-analysis, Effects of Generative AI Tools on Students' Critical Thinking: A Three-Level Meta-Analysis, in Educational Psychology Review, a study that combines results from many studies. Their analysis combined results from 26 studies of learning supported by generative AI and critical thinking. On average, learning supported by generative AI was associated with higher critical-thinking performance. The results also varied substantially across studies.

Those mixed results lead to a more useful question: What thinking is the person still doing while AI is helping?

I think of the difference as active use and passive use. Active use keeps me involved in the reasoning. I am questioning, checking, comparing, building on ideas, revising, and deciding. Passive use hands the core reasoning to the tool and treats its answer as the end of the thinking.

Those are my everyday terms. Research uses more specific language for different forms of AI-assisted thinking. The connection matters, and so does keeping the terms separate.

Active use keeps the person inside the reasoning

Picture a message you need to prepare about a decision.

You could give AI a few facts and ask it to produce the whole message. You could then read the answer once, make a small wording change, and send it. Most of the reasoning about what matters, what needs context, and how the ideas connect has already been handed over.

You can use the same tool differently. You can decide what the message needs to accomplish. You can give AI the facts and ask it to organize several possible sequences. You can compare those options against what people need to understand. You can question a missing step, check a claim against the source information, and decide which explanation belongs in the final message.

Both approaches may end with a complete message. The human contribution is different.

In active use, you remain responsible for the questions that shape the work. You decide what deserves attention. You test whether an answer fits the facts and the situation. You decide what changes and what stays.

James Clear's work on identity-based habits offers a useful analogy. Repeated actions can become patterns. The way you repeatedly use AI can also become a way of working. If your normal pattern includes questioning, checking, comparing, and deciding, those actions stay part of the process each time the tool helps.

The analogy gives us a way to think about repetition. The research carries the claims about critical thinking.

Research distinguishes different kinds of cognitive offloading

Psychologists use the term cognitive offloading for using something outside your mind to reduce the mental work you have to do yourself. Writing a reminder is a simple example. A calculator can handle arithmetic while you focus on the larger problem.

With generative AI, people can offload more kinds of cognitive work. A person can ask the tool to draft an argument, compare options, summarize information, or propose a conclusion. That makes one question especially important: who is steering the reasoning?

A 2026 study, Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI, by Qiuhan Zhu and colleagues in Frontiers in Psychology examined this question by distinguishing two forms of cognitive offloading: dependent cognitive offloading and autonomous cognitive offloading.

Dependent cognitive offloading means delegating core thinking to AI. The researchers describe patterns such as accepting AI output with little evaluation, letting AI organize the reasoning, and treating the resulting AI output as the main answer.

Autonomous cognitive offloading means using AI as a scaffold while retaining cognitive agency. The person compares AI output with their own reasoning, continues thinking from what the tool provides, and keeps ownership of the final answer.

The researchers measured the two forms separately. Their language is specific to their study. My active/passive framing describes similar practical territory in everyday words, especially the question of who retains cognitive authority during the work.

Zhu and colleagues followed 589 university students and early-career knowledge workers across three survey waves. Dependent offloading was associated with greater cognitive agency transfer and lower intrinsic motivation, the motivation that comes from interest in the activity itself. Autonomous offloading was associated with higher intrinsic motivation. The researchers found no reliable association between autonomous offloading and cognitive agency transfer.

Cognitive agency transfer sounds technical. The idea is concrete. It asks who is deciding what information matters, how evidence should be judged, and which conclusion to draw. When a person hands those decisions to AI, they also hand over part of the reasoning.

The study also connected cognitive agency transfer and intrinsic motivation with participants' later perceptions of how well they could process ideas deeply and make independent judgments. These are associations from a survey design. They show different patterns around different forms of offloading. Causal effects on thinking ability remain outside what this study tested.

The study supports a narrower point: how cognitive work is allocated during AI use matters.

Critical thinking can show up in different parts of AI-assisted work

Hao-Ping Lee and colleagues studied this in The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers, a paper published at the 2025 Conference on Human Factors in Computing Systems. They surveyed 319 knowledge workers who described 936 real examples of using generative AI in their professional work.

The researchers found that critical-thinking effort often shifted toward activities such as verifying information, integrating AI responses into the work, and maintaining responsibility for the result.

Suppose you use AI to create a first draft. You may spend less time producing the first sentences yourself. Your thinking can then shift to evaluation. You check the facts. You look for missing context. You test the sequence against what people need to understand. You decide which parts to keep, change, or remove.

Those are thinking decisions that happen later in the process.

Lee and colleagues also found a confidence pattern. Higher confidence in generative AI was associated with less reported critical-thinking effort. Higher confidence in one's own ability was associated with more reported critical-thinking effort. The study shows associations, and the participants were general knowledge workers. It gives us a useful professional example of where people reported doing critical-thinking work when AI was involved.

Less mental effort can still leave important thinking with the person

Mental effort and quality are different measures.

Purvi Jain and colleagues tested young adults on an activity that asked them to identify errors in reasoning, with and without ChatGPT. The study, Cognitive offloading, critical thinking and attitudes towards artificial intelligence in the era of ChatGPT, was published in Cognitive Processing. The AI-assisted group produced more correct responses and reported lower mental effort on that activity. The study also found different approaches inside the AI group. Some people delegated more of the work. Others used AI selectively while continuing to reason, verify, and monitor what they were doing.

The study helps separate effort from judgment.

You can use AI to reduce effort in parts of the work. Some steps carry less professional judgment than others. Putting dates into order, grouping repeated notes, or applying a consistent format can take time while leaving the main judgment for the person doing the work.

The important question is what happens to the thinking that remains.

If you use AI to group a set of notes, you can spend the time you saved checking what is missing. If you ask AI for three possible explanations, you can compare them against the facts and the people who need to understand them. If you use AI for a mechanical first pass, you can spend more attention on the decision that requires context and responsibility.

Useful efficiency can create room for judgment. The person decides how to use that room.

People can change how they rely on AI output

Chenyu Hou and colleagues studied 226 participants in a higher-education collaborative problem-solving setting. The study, The effects of critical thinking intervention on reliance behaviors, problem-solving quality, and creativity during human-Generative AI collaborative learning, was published in Computers & Education. One group received a critical-thinking intervention that included structured dialogue and other supports for more deliberate engagement with generative AI.

The intervention reduced how often participants used AI-generated content directly. The intervention group's reports also contained solutions that were more original and contained more ideas. The study's self-reported critical-thinking score did not increase significantly during the short intervention.

The intervention changed how participants relied on AI output.

This gives us evidence that engagement patterns can change. A person can choose to question an answer that first looks acceptable. They can compare it with another source. They can ask what assumption sits underneath it. They can decide that the tool's suggestion conflicts with the situation and take another direction.

What this looks like in K-12 communication work

Most of the research above studies university students, young adults, or general knowledge workers. Its findings give us evidence about AI-assisted thinking in those populations. Applying the ideas to K-12 professional communication is a separate reasoning step.

Consider a few ordinary examples.

You have source notes for a message. You can use AI to organize the notes into a possible structure. You still check whether every fact is accurate, whether an important piece of context is missing, and whether the sequence helps people understand the decision.

You have a draft that feels incomplete. You can ask AI to identify questions a reader might still have. You then decide which questions are real gaps, which facts answer them, and which details belong in the final communication.

You are trying to explain a complicated change. You can use AI to generate several ways to organize the explanation. You compare those options with what you know about the situation and choose the one that makes the reasoning easiest to follow.

You have a long set of source material. You can use AI to help sort or summarize it. You return to the original material for the facts that carry the message and decide what is accurate enough to use.

In each example, the person uses AI to reduce or support part of the work while keeping the professional judgment. The person sets the purpose, supplies context, checks information, weighs alternatives, revises, and decides.

Those are concrete signs that the person is still doing the thinking that matters.

A better question to ask about your own AI use

The evidence gives us a more useful answer to the title question.

Critical-thinking outcomes vary across AI-supported settings. One important difference is how the cognitive work is divided between the person and the tool.

I keep coming back to active and passive use. The terms are simple. The question behind them is serious: What thinking am I still doing?

Look at one recent piece of professional work where AI helped. Ask yourself three questions:

What did I question?
What did I check or compare?
What did I decide?

Those answers tell you where your judgment stayed active in the work.

AI-assisted work becomes more valuable when the tool creates more room for human judgment and the person keeps exercising that judgment.

Does AI Hurt Critical Thinking? The Difference Between Active and Passive Use