Two studies landed on the same question this month and gave opposite answers. The difference between them is a choice you make about forty times a day without noticing.
I will tell you the version of me I am least proud of.
It is about four in the afternoon. I have been in back-to-back conversations since seven, I have one more thing to write before I can stop, and I know exactly what good looks like for that thing. I also know it will take me forty-five minutes of actual thinking to get there. So I open a window, type a request that is really a confession of surrender, and I take whatever comes back with two small edits.
Nobody is harmed. The thing is fine. And I have no idea how many times I have done that, which is the part that bothers me now.
So let me answer the question in the headline honestly, because two studies landed on it this month and they appear to disagree. Yes, it can make you worse at thinking, and the research is more specific than the headlines are. It is not the tool that does the damage. It is the handover. When people hand AI the whole problem, their independent performance drops and they quit sooner. When people use the same tool to explain and clarify while doing the work themselves, they hold their ground and in at least one study they do better.
Same tool. Same ten minutes. Opposite outcomes. The variable is whether you handed over the typing or the thinking.
Key Takeaways
- A UC Berkeley-led team ran randomized experiments with over a thousand participants and found that people who had AI help did worse and abandoned tasks more often once the help was removed.
- A Middlebury study found the opposite result for students who used AI to explain and clarify rather than to complete the assignment, which means the handover is the variable rather than the tool.
- A Microsoft and Carnegie Mellon survey of 319 knowledge workers found that confidence in AI correlated with less critical thinking, while confidence in one's own expertise correlated with more.
- Offloading execution is usually smart; offloading the reasoning behind a decision you will have to make again is how the ability to make it quietly disappears.
- The practical fix is to make the AI ask rather than answer, which takes longer on exactly one decision per week and is the only move in this article that compounds.
The Problem: Nobody Can Feel This Happening
Here is why this is a genuinely hard problem rather than a willpower problem.
Most things that erode give you a signal. If you stop lifting, you notice when you lift. If you stop running, your lungs tell you. If you stop practicing an instrument, your hands tell you in the first bar.
Judgment does not do that. Judgment degrades silently, because the thing you lose is the thing you would use to notice the loss. You still produce output. The output is still fine. Your clients still pay. The only thing that changed is that you can no longer reliably tell the difference between a good answer and a plausible one, and you will not discover that on an ordinary Tuesday. You will discover it on the one decision that actually mattered.
I have been in business long enough to have watched this happen with other tools, and the pattern is always the same. The capability you stop exercising does not announce its departure. It just is not there the day you reach for it.
And the honest part, the part I do not enjoy writing, is that the four in the afternoon version of me is not an aberration. It is a default. Nobody decides to stop thinking. You just get tired, and the tool is right there, and it is very good, and the thing you produce is adequate. Forty times a day, adequately, for eighteen months.
What if the real risk of AI for people like us is not that it replaces us, but that we quietly stop being the kind of person it could not replace?
The Evidence: Two Studies That Appear to Disagree and Do Not
The Berkeley study. A team led by Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI, with co-authors from Carnegie Mellon, MIT, Oxford and UCLA, ran three randomized online experiments with 1,222 participants, 1,060 of whom remained in the analyses after pre-specified exclusions. The tasks were fraction problems and SAT reading questions. One group had AI access until it was removed at problem twelve. Once the help was gone, that group's accuracy dropped sharply while the control group held steady. The former AI users also abandoned tasks more often. The paper is titled "AI Assistance Reduces Persistence and Hurts Independent Performance," and it appeared as a preprint in April with a revised version presented at the Conference on Language Modeling in October 2026.
Now the caveats, because the caveats are the interesting part and I would rather you have them than be impressed by me. Participants were told they were using ChatGPT; what sat behind the sidebar was a model preloaded with each problem and its solution, meaning it could give a correct answer to every request. That is a cleaner condition than real life. The paper also expressly asks not to be read as proof that AI reduces persistence in all activities, and the authors call for longitudinal work this design does not provide. Some secondary coverage has claimed the effect held across different types of cognitive tasks. It tested two.
So the honest summary is narrower and still unsettling: in controlled tasks, a short stretch of AI help made people worse and quicker to quit once it ended.
The Middlebury study. This is the one that turns a scary finding into a usable one, and I came to it through Michael Hyatt, who did something I wish more people would do with alarming headlines. He went to the underlying research instead of forwarding the scariest version. What he found was a distinction rather than a verdict: students who handed AI the whole assignment scored worse on later tests, while students who used it to explain and clarify kept their scores and wrote better essays.
His framing of it is the sentence I have not been able to put down: "Used as a shortcut, AI erodes your thinking. Used as a tutor, it sharpens it. Most leaders only practice the first mode."
Read the second half of that again. Not "some leaders." Most. Including me, at four in the afternoon.
The Microsoft and Carnegie Mellon survey. A 2025 study titled "The Impact of Generative AI on Critical Thinking" surveyed 319 knowledge workers who use generative AI at least weekly, and analyzed 936 real-world examples of AI-assisted tasks. Its central finding, in the authors' own words: "Confidence in AI is associated with reduced critical thinking effort, while self-confidence is associated with increased critical thinking." Workers who trusted their own expertise were more likely to actually check the output. One summary of the data reports that workers more confident in their own expertise were 27 percent more likely to critically assess AI output rather than accept it at face value.
Important caveat: this is a self-report study. Participants rated their own effort and confidence, so it shows a correlation and not a causal effect on thinking ability. Michael Gerlich, who studies cognitive offloading at the Swiss Business School, adds the nuance that matters: the effect is largely a consequence of how people use the tools, and used deliberately, AI can enhance critical thinking.
Three studies, three methods, one consistent thread. The variable is never the tool. It is what you handed over.
The Fix: Make It Ask Instead of Answer
The move that changed this for me is embarrassingly small. I stopped asking for output on the decisions that matter, and started asking for questions.
Here is what that looks like in practice, borrowed directly from Hyatt's two tactics because they are better than what I was doing.
Tactic one, interrogation. Paste your rough notes and instruct the model to ask you one question at a time and write nothing else. Just questions. You answer them. When you are done, you say "summarize my position in three sentences." If that summary is thin, your thinking was thin, and you just found out for free rather than finding out in a client meeting.
I ran this on a pricing question last month and it was genuinely uncomfortable. The fourth question was one I had been avoiding for a year, and I did not have an answer, and the reason I did not have an answer is that every previous time I had approached that decision I had asked for options instead of being asked for reasoning. Options are comfortable. They let you choose without ever constructing anything.
Tactic two, opposition. Cast the model as a skeptical board member who names the three assumptions you are most likely wrong about and withholds any alternatives until you have answered each objection yourself. That last clause is the whole trick. If it offers alternatives immediately, you will take one, and taking one is how you skip the part where you defend your own position badly and learn something from how badly you defended it.
And then there is the distinction Hyatt draws that I now apply to every automation decision, because it resolved something I had been confusing for two years. Agents take your thinking at setup and then run alone. A generative session keeps you in the conversation. Those are not the same purchase. One buys leverage on a decision you have already made well. The other is where the decision gets made. Automate the first category all day. Automating the second just builds a very efficient machine to execute a judgment you never actually formed.
His line on where to draw it is the cleanest I have read: "Offloading a calendar invite is smart. Offloading the reasoning behind a pricing decision is how you lose the ability to make the next one."
Practical Steps
1. Audit what you actually handed over last week. Not what you intended to hand over. Pull up five real prompts you used and sort them into two columns: typing I handed over, and thinking I handed over. Most people find at least one item in the second column that shocks them. I found three.
2. Write your one-sentence rule and make it specific enough to use at four in the afternoon. Mine is this: if the output of this decision will be used more than once, I form the position before I open the tool. Yours will be different. It needs to be short enough that tired-you can apply it without deliberating, because tired-you is the one who needs it.
3. Run interrogation mode on exactly one decision this week. Not all of them. One. The one that compounds. Paste your notes, demand questions only, and close with the three-sentence self-test.
4. Write your own rough answer before you open any tool. Badly, in your own words, on the decisions that matter. Then let the AI find the gaps in your version rather than supply a version. This single reordering is most of the difference between the Berkeley condition and the Middlebury condition.
5. Keep one task entirely by hand as a calibration reference. Pick something where quality is visible and judgment is central. Do it unaided, regularly. Without a reference you cannot tell whether AI output is improving or your standards are declining, and those two feel identical from the inside.
6. Test whether you can still spot a wrong answer. Producing a right answer and recognizing a wrong one are different skills, and the second one is the one that protects your customers. Have the AI generate three versions of a typical output with two containing realistic errors, and see if you catch them. Do it quarterly. It is humbling and it takes ten minutes.
7. Teach somebody the reasoning, not the conclusion. This is the oldest test there is. If you cannot teach your team why you decided something in a way that lets them handle a case you did not anticipate, you did not do the reasoning. You collected an output.
Here is the prompt I actually use for step three. It follows the Perfect Prompt Framework 2.0.
[The Job]
Interrogate my thinking on [NAME YOUR CURRENT DECISION] without giving me any answers.
This is for: me, thinking out loud.
It matters because: I want to find the holes in my own reasoning rather than adopt yours.
[The Background]
Here is what you need to know: here are my rough notes on the decision: [PASTE YOUR NOTES]. The constraints are [LIST THEM]. The people affected are [LIST THEM].
Do not use: recommendations, options, or opinions of any kind at this stage. Ask one question at a time and write nothing else.
[The Deliverable]
Return: one question. Then wait for my answer before asking the next. When I say stop, summarize my position in three sentences so I can see whether it holds together.
Optimize for: exposing weak reasoning.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
Does using AI actually make you dumber?
The research does not support that framing. A Berkeley-led experiment found reduced independent performance and higher task abandonment after AI help was removed, while a Middlebury study found students who used AI to explain and clarify held their scores and wrote better essays. The consistent variable is whether the person handed over the whole task or used the tool to support work they were still doing themselves.
What is cognitive offloading and when is it a problem?
Cognitive offloading means using an external aid in place of your own mental effort, which is sometimes entirely sensible, as with a calculator or a calendar. It becomes a problem when the offloaded capability is one you will need to exercise again, especially judgment under uncertainty. The test is whether you are offloading execution or the reasoning behind a recurring decision.
How do I use AI without losing my ability to think?
Reverse the direction of the conversation on anything that matters. Instruct the model to ask you one question at a time and write nothing, answer the questions yourself, then ask it to summarize your position so you can see whether it holds together. A second method is to have it argue against your position and withhold alternatives until you have answered each objection in your own words.
Is it safe to automate a business decision with AI?
Automate a decision only once you have made it well manually enough times to know what good looks like and what the common failure modes are. An agent takes your thinking at setup and then runs without you, so automating a judgment you have not formed produces the same mistake faster and more often. Conversational use is where the judgment gets built; agents are where a settled judgment gets scaled.
How would I know if my judgment is already declining?
You probably would not notice from the inside, which is the core difficulty. Build an external check instead: keep one task you do entirely unaided as a reference point, test quarterly whether you can still spot deliberately introduced errors in typical output, and try to teach your reasoning to someone who will have to apply it without you.
The Close
Four in the afternoon. One more thing to write. Forty-five minutes of real thinking, or two minutes and two small edits.
I want to be honest with you about where I have landed, because I do not think the answer is discipline and I do not think it is abstinence. I have tried both. Discipline fails at four in the afternoon, which is exactly when it is needed, and abstinence is just leaving the best tool of my career in a drawer to prove a point to nobody.
What works is deciding in advance which decisions I do not get to shortcut, and then building the habit of making the machine ask instead of answer on those. One decision a week. Sometimes two. I move slower on those and I keep something I cannot buy back later.
Because here is what I have come to believe about this moment, and it is not a comfortable thought for someone who makes a living teaching people to use these tools. Every advantage in AI is temporary and purchasable. The model you are excited about will be commodity in a year. The workflow you built will be a feature in somebody's product by spring. Anyone with a subscription can have what you have, and most of them eventually will.
Judgment is the one thing in your business that nobody can buy, nobody can copy, and no vendor can ship you an update for. It is also, on the current evidence, the one thing that quietly leaves if you stop using it.
Pick one decision this week and do the hard version. That is the whole thing. The rest of this is just me trying to convince you that the hard version is the only part of your work that will still be yours in five years.
About the author
Jonathan Mast is the founder of White Beard Strategies and writes about AI, entrepreneurship and the parts of building a business nobody puts in the highlight reel. He created the Perfect Prompt Framework 2.0, speaks on practical AI adoption, and serves a community of more than 100,000 entrepreneurs. He still writes his hardest decisions out by hand first, badly, usually on the back of something, because the tidy version always sounds more certain than he actually is.