The Thinking We Can’t Afford to Hand Over to AI
As I write this, I am deliberately making my own thinking harder.
I am spending time in an environment where history, politics and culture are often explained through a different lens from the one I am accustomed to. Events I recognize are framed differently. Familiar assumptions are not always shared. I find myself having to hold two interpretations in my head at once, work out what each one emphasizes or leaves out, and decide which facts remain stable even when their meaning changes with the frame.
At the same time, I have been developing two articles about AI, productivity and critical thinking. That process has required a different version of the same intellectual work: connecting ideas that arrived separately, testing whether the connections actually hold, discarding statistics that were interesting but did not support the argument, and eventually recognizing that I was trying to force too many good ideas into one article.
AI has been involved throughout. But the most valuable moments have not come from asking it to produce an answer. They have come from using it to make my own thinking more difficult: challenge this premise, show me what I am missing, tell me what does not belong, argue the other side.
That experience has made me increasingly skeptical of one of the most common pieces of advice about critical thinking with AI: check the output.
Of course we should check the output. But by then, some of the most consequential thinking may already be over. The harder question is whether we have retained enough ownership of the problem, the evidence, the connections and the argument to know what deserves to be checked in the first place.
The most dangerous AI answer may be the one that looks right
Hallucinations are relatively easy to understand. A fact is wrong. A source does not exist. A number cannot be substantiated. Those failures matter, but at least they give us something visible to challenge.
A more difficult failure is an answer that is polished, coherent and plausible.
Nothing is obviously false. The logic flows. The evidence sounds credible. The paragraphs fit neatly together. Yet the question may have been framed too narrowly, an important perspective may be missing, or several related ideas may have been compressed into an argument they do not genuinely support.
This is where fluent AI creates a particular problem. Good writing can make weak thinking look resolved.
Research from Carnegie Mellon University and Microsoft Research helps explain why this matters. Their study of 319 knowledge workers and 936 examples of AI use found that generative AI shifts cognitive effort. Information gathering increasingly becomes information verification; problem-solving shifts toward integrating AI responses; and analysis, synthesis and evaluation move from direct task execution toward oversight or “task stewardship.”
That shift is important, but it still begins after we have already accepted a task worth stewarding.
Before an AI response exists, someone has decided what the problem is, which assumptions belong inside it, what evidence matters, which perspectives deserve attention and what would constitute a useful answer. If AI establishes that frame and we simply evaluate what follows, we may be exercising judgment inside a structure we never examined.
Critical thinking therefore has to begin further upstream.
Critical thinking requires some cognitive discomfort
The environment I am in now has made this especially visible to me. I am deliberately putting myself in situations where I cannot rely on the historical, cultural and political frames I already know. I am confronted with points of view I didn’t know existed, hear familiar events described from another vantage, encounter interpretations I wouldn’t hear at home, and am forced to keep an open mind for added data points to add to my mental rolodex.
That takes considerably more effort. I have to ask what each interpretation wants to convey and why, what it leaves out, what assumptions sit underneath it, which facts remain relatively stable across accounts, and which meanings change depending on who is telling the story. Sometimes the useful outcome is not choosing one interpretation over another, but understanding why two people could look at much of the same evidence and construct different meanings from it.
That kind of thinking is uncomfortable because it leaves things unresolved for longer. It asks us to hold competing possibilities in our heads rather than immediately selecting one, and it takes time.
This matters with AI because AI is exceptionally good at removing that discomfort. Ask a sufficiently capable model to reconcile two positions and it will usually find the common ground. Ask it to summarize a debate and it can quickly give you the principal arguments. Ask which interpretation is stronger and it will offer a conclusion. All of that can be useful, but sometimes the thinking we need to exercise is precisely the work that happens before everything has been reconciled.
AI can also help us create that productive discomfort. We can ask what someone from another industry might see, what a regulator might question, how someone with opposing incentives would interpret the problem, or what someone who rejects our premise altogether would say. The value is not in having AI decide which perspective wins. It is in giving ourselves more material to think with.
Some of the best thinking happens between ideas
We tend to describe critical thinking as an evaluative skill: challenge the premise, scrutinize the evidence, identify the bias, find the flaw. But critical thinking is also about making connections.
The argument for this article did not originate with me sitting down and asking AI, “What should I write about critical thinking?” It developed through several separate conversations and ideas that began to connect. Someone else had raised an idea with me about AI supporting creative thinking. I was separately reading McKinsey research about the productivity gains organizations expect from AI and the possibility of freeing people from more routine work. At the same time, I had been thinking about something that has concerned me in my own AI use: the ease with which we can begin to hand over not simply execution, but judgment itself.
The useful idea was not contained in any one of those inputs. It emerged from putting them beside one another: if AI is expected to create more time for creative, strategic and higher-value thinking, what happens if we are simultaneously using AI to perform more of that thinking for us?
McKinsey’s 2026 research makes the question more than theoretical. When AI frees employees’ time, only 30% of leaders currently see that time being redirected toward higher-value activities such as critical thinking and creativity. Its broader research on future skills suggests that people may spend less time preparing documents and conducting basic research and more time framing questions and interpreting results.
That sounds promising, but framing questions and interpreting results are not simply the activities left over after AI performs the easier work. They are difficult capabilities in their own right, and they are capabilities we only strengthen by exercising them.
AI can help us think, but we have to decide what role to give it
That realization has changed some of the ways I use AI. There are times when I want the answer. There are other times when I deliberately want resistance.
While working through these ideas, I have asked AI to identify where my argument does not hold, what opposing evidence I have overlooked, which assumptions I am making, what I might be forcing together and which parts of the argument deserve to be separated. Those prompts are useful because they change the role AI plays. Instead of asking it to complete my thinking, I am using it to put more obstacles in its path.
That is not a trivial distinction. A 2025 experiment on automation bias found that participants who received faulty AI assistance answered fewer than half as many Cognitive Reflection Test questions correctly as participants who received no AI assistance. Adding a warning that encouraged people to reflect critically improved performance substantially, although it did not raise performance above the group working without AI. Self-reported AI literacy did not significantly protect participants from automation bias either.
Research from Carnegie Mellon University and Microsoft Research adds another dimension. In their study of knowledge workers, greater confidence in AI performing a task was associated with less critical thinking. Greater confidence in one’s own ability to perform the task and evaluate the AI response was associated with more.
The implication is not that we should distrust AI or make simple work unnecessarily laborious. It is that we should become much more deliberate about which cognitive work we want AI to remove and which cognitive work we need to continue doing ourselves.
The harder part was deciding what I was wrong about
This article itself became an example. At one point, I had a large collection of ideas that all seemed relevant: AI productivity, critical thinking, cognitive offloading, creative thinking, AI agents, instructional design, opposing perspectives, human judgment and the changing nature of work. There was research to support many of them.
AI could make them fit. That turned out not to be the same thing as having a strong argument.
The more I worked on it, the clearer it became that some of my own connections were too loose. A fact could be accurate without advancing the idea I was trying to develop. Two concepts could be related without belonging in the same argument. An interesting tangent could still be a tangent.
Eventually I realized I was not struggling to structure one complicated article. I had two different articles. One became a practical piece for learning professionals about how to design learning that requires people to compare perspectives, develop ideas, challenge assumptions and exercise judgment with AI. This article became something different: an examination of the thinking process itself and how easily AI can allow us to bypass parts of it.
That decision took far more thinking than asking AI to generate another structure. It required me to recognize that something I had already spent considerable time building did not work. That, too, is critical thinking. It is not simply finding fault with someone else’s argument; it is being willing to dismantle your own.
Synthesis is not the same as making everything fit
The process also exposed another trap I now watch for when using AI. Generative AI is exceptionally good at combining ideas. Give it five related concepts and it can usually produce an elegant paragraph that contains all five.
But combining is not necessarily synthesis. Real synthesis produces something new. My conversation with someone about creative thinking, the McKinsey evidence about freeing human capacity, and my concern about cognitive offloading became useful because the connection generated a question that none of the pieces provided independently.
If AI creates a smooth paragraph containing three different ideas, by contrast, that may simply be compression. The tension has disappeared, but no new understanding has taken its place.
I now find myself asking: Did connecting these ideas reveal something I could not see before, or did I just make them sound as though they belong together? That question slows the process down, but it also improves it.
Critical thinking is not a final check on AI
This is why I think “check the AI output” is far too limited as advice. Of course we should verify facts and sources. But critical thinking runs through the entire process, often before there is an output to verify. It influences how we frame the problem, what we think before AI enters the conversation, which perspectives we seek, which ideas we connect, which challenges we invite, and what we are ultimately prepared to discard.
There is no neat sequence for doing that. My own experience certainly has not followed one. But several questions repeatedly force me to do more of the thinking myself:
- What problem am I actually trying to solve?
- What do I think before AI gives me something to react to?
- Whose perspective have I not considered?
- What becomes visible when I connect this idea with something from somewhere else?
- What would make me conclude that my argument is wrong?
- Am I creating a genuine synthesis or merely a coherent one?
- What should I remove, separate or leave unresolved?
AI can help with every one of those questions. What it should not do is make the questions unnecessary.
The Carnegie Mellon University and Microsoft Research found that knowledge workers generally perceived less cognitive effort when using generative AI. Importantly, they note that this reduction can mean different things: AI may be helping people perform the thinking more efficiently, people may be offloading some of the thinking to AI, or they may simply be doing less of it.
That matters because the promise of AI is that we can spend less time on some forms of work and redirect our capacity toward work that requires more judgment, creativity and thought. If we want that promise to hold, we cannot also optimize all of the effort out of thinking.
Critical thinking takes time. It can be uncomfortable. It requires exposure to ideas we did not generate ourselves, connections that are not immediately obvious, arguments against our own position and the willingness to abandon work that seemed promising. AI can make that process richer, but if we use it primarily to make the process easier, we may gradually hand over precisely the thinking we hoped AI would free us to do more of.