Use AI Without Losing Your Mind

A lot of people who have been using AI for the last two years have started to notice the same uncomfortable thing. What they used to do alone, they cannot do alone anymore. They sit down to write an email, draft a memo, or work through a problem set they used to find routine, and they cannot get started without opening ChatGPT or Claude. When they try to push through alone, the effort feels disproportionate to what the task should require. When they give up and use the tool, it does not feel like giving up. It feels reasonable. It feels like the obvious thing to do given how much harder the unassisted version has become.

If you have felt this, you are not imagining it. Something is actually happening, and as of April 2026 we have causal evidence for what it is. A team of researchers at Carnegie Mellon, Oxford, MIT, and UCLA ran the first large scale randomized trial on this exact question, and the finding is uncomfortable. Ten to fifteen minutes of AI-assisted problem solving is enough to measurably hurt how well you solve problems on your own. Not just your performance. Your willingness to keep trying.

The good news, which most of the coverage of this paper has buried, is that the harm is not caused by AI itself. It is caused by a specific way of using AI. Other studies, including a clean randomized trial at Wharton, have already shown that AI used differently produces the opposite effect: students learn faster, retain more, and lose nothing when the AI is taken away. What follows is the research behind both, why it happens, and how to use AI so it actually helps you think.

Three major studies, run with different methods on different populations, all converge on the same finding. When you let AI do the thinking for you, two things happen. You get worse at the task when the AI is removed. And you stop trying.

The Liu paper (April 2026): the new causal evidence

Participants were given fraction problems or SAT reading comprehension passages. Half had access to a chatbot built on GPT-5 that had been pre loaded with the answers. The other half worked alone. Both groups were paid the same regardless of performance, and both could skip any problem they wanted.

The chatbot was available for the first chunk of problems. Then, without warning, it was taken away. The final three problems had to be solved alone. This is the part that matters. The researchers were not measuring how well people perform with AI. They already knew AI helps in the moment. They wanted to know what happens to your unassisted thinking after even brief AI use.

The numbers, in order of importance:

• After the AI was removed, the AI group skipped 20% of the remaining problems entirely. The control group skipped 11%. The AI users were nearly twice as likely to simply give up.
• The AI group solved 57% of the test problems overall. The control group solved 73%.
• In the reading comprehension experiment, the gap was sharper. The AI group skipped problems eight times more often than the control group.
• All of this emerged after just 10 to 15 minutes of AI use.

The skip rate matters more than it might appear. Participants were told repeatedly that there was no penalty for wrong answers and that their payment did not depend on getting questions right. Choosing to click "skip" rather than attempt a problem was a clean signal of one thing: the participant had decided the problem was not worth the effort. Skip rate is the experimental measure of giving up. And after 10 minutes with AI, the participants who had been using it gave up roughly twice as often as the participants who had never seen it.

The researchers compared this to the boiling frog effect. Each individual moment of letting AI think for you feels weightless. The cumulative effect is not.

Two things make this paper a turning point. First, it is causal. Until now, most concerns about AI and cognition came from surveys, interviews, or correlational studies. People who used AI more reported feeling foggier. That is suggestive but not proof. The Liu paper is the first large randomized trial to move the needle on causal evidence. Second, the timeline is fast. Ten minutes. Not months. The shift starts almost immediately.

The Kosmyna study (June 2025): your brain on ChatGPT

A team at the MIT Media Lab, led by Nataliya Kosmyna, asked 54 students from five Boston area universities to write SAT style essays. Three groups: one used ChatGPT, one used a search engine, one used nothing. Each group wrote three essays over the course of four months, with a fourth crossover session for some participants. The whole time, the researchers had 32 channel EEG headsets on the participants, monitoring electrical activity across the scalp.

The ChatGPT group's brain connectivity was significantly lower than the brain only group across multiple frequency bands. The largest gaps were in networks tied to attention and working memory. Brain connectivity systematically scaled down with the amount of external support: the brain only group showed the strongest, widest ranging neural networks. The search engine group showed intermediate engagement. The ChatGPT group showed the weakest coupling. After they wrote their essays, most of the ChatGPT users were unable to accurately quote from their own work moments later. They had typed it and they had read it. But they could not recall it.

By the third essay, many of the ChatGPT users had stopped engaging with the writing process at all. They were copy-pasting the AI's output with minor edits. The researchers gave this pattern a name that has since gone viral: cognitive debt.

The fourth session is the part that should concern you. Participants who had used ChatGPT for sessions 1 through 3 were asked to write their fourth essay without it. Even with the AI removed, their neural connectivity was still weaker than the people who had never used it. The pattern, brief as the exposure had been, had stuck.

A caveat the headlines flattened: this is a preprint, not yet peer reviewed. The sample is small. It is one task. Even Kosmyna pushed back on the "ChatGPT rots your brain" framing that dominated the press cycle. What the study showed is more modest and more disturbing. When you ask an AI to write for you, the brain regions that would normally do that work go quiet. And they do not snap back instantly when the AI is removed.

The Bastani study (Wharton, PNAS 2025): the crutch effect in a real classroom

A team at Wharton ran a field experiment with nearly 1,000 Turkish high school students across about 50 classes. Students took a regular math lesson, then moved to a practice phase. One group had only their notes. The other had access to a custom ChatGPT interface built on GPT-4. The next day, both groups took a clean exam without any AI.

During practice, the ChatGPT group did spectacularly better, solving 48% more problems correctly. On the next day's exam, that inverted: the ChatGPT group scored 17% worse than the notes-only group. The mechanism was the same crutch effect the Liu paper would later document. Students were using ChatGPT as a crutch. They predominantly typed things like "what is the answer?" rather than working through the problem step by step.

What made the finding harder to dismiss was what the students reported afterward. When surveyed, the ChatGPT group did not perceive that they had performed worse or learned less than anyone else. They had walked out of the lesson confident, and into the next day's exam unprepared.

The Bastani team also tested a third group, with an AI interface designed differently. We will come back to that.

Pulling the three together

A 1,222 person randomized trial in 2026. A neuroscience study with brain scans across four months. A field experiment with 1,000 students in a real classroom. Different methods, different populations, different domains. The signal in all three is the same.

When AI does your thinking for you, your unassisted performance drops. More importantly, your willingness to engage with difficulty drops with it. The drop in willingness is the part that compounds, because it shapes whether you ever build the skill in the first place. The Liu paper puts this more precisely than we can:

"People do not merely become worse at tasks, but they also stop trying."

That second half is the part that should worry you. Performance is recoverable. The decision to stop trying is the deeper damage.

Why does this happen? The Liu paper proposes two mechanisms, both grounded in older cognitive science. Knowing them helps you spot the pattern in your own behavior.

Mechanism 1: Hedonic adaptation

Hedonic adaptation is the psychology of getting used to things. The first weekend in a new apartment feels luxurious. Six months later it feels ordinary. The first run after months on the couch feels brutal. A month into a habit, the same distance is pleasant.

The same mechanism applies to cognitive effort. When AI completes a problem in 10 seconds, your brain quietly registers a new reference point: this is how long this should take. The next time you sit down to do the same kind of problem yourself, the 10 minutes of effort it actually requires now feels disproportionate. Worse than disproportionate. Painful.

This is not weakness. This is how the brain calibrates. The problem is that the calibration is happening in the wrong direction. Each AI-assisted rep shifts your tolerance for effort downward. Each downward shift makes the next AI-assist more attractive. The Liu paper calls this self-reinforcing, and the researchers' math suggests they are right.

Mechanism 2: Loss of metacognitive calibration

The second mechanism is subtler and arguably more damaging. Metacognition is the technical term for thinking about your own thinking. When you sit with a hard problem, fail, try again, fail differently, and finally crack it, you are not just learning the problem. You are learning what you can do. You are mapping the edge of your own ability.

When the AI hands you the answer, you skip that mapping. You never find out how close you were. You never feel the click of the right approach. You never learn that you can, in fact, do this.

Over months and years, this hollows you out in a way that is hard to see from the inside. You can still get answers. You just lose the sense of what you can produce on your own. Persistence, self regulation, and the ability to push through difficulty are among the strongest predictors of long term academic success. All of them rest on the self knowledge you only get from doing hard things alone. When the self knowledge erodes, the persistence erodes with it.

These two mechanisms reinforce each other. The effort feels worse than it used to. The confidence that the effort would pay off has quietly thinned. That is the combination behind the Liu finding: people in the AI condition are not failing the task. Most of them never start.

Two of the studies we just walked through have a second story buried inside them. Almost no coverage of either paper has surfaced it. It changes the practical picture entirely.

The Liu data, recut

In the second Liu experiment, the researchers asked the AI condition participants how they had actually used the chatbot during the task. The answers split the group into three clean buckets.

• 61% used the AI primarily to get answers directly. They pasted the problem in. They copied the answer down. This was the largest group.
• 27% used the AI for hints and clarifications. They asked for the first step. They asked the chatbot to explain a concept they were not sure about. They kept their own minds in the loop.
• 12% chose not to use the AI at all, despite having access.

When the researchers compared these three groups on the unassisted test problems against their own pretest scores, the pattern was sharper than anything in the summary statistics.

The direct answer group, the 61%, took the full hit. Their solve rate on the test dropped by 10 percentage points relative to their own pretest. Their skip rate jumped 11 points, nearly double the control group's 6 point jump. By the time the AI was taken away, this group scored measurably worse and skipped measurably more than they had walking in.

The hint seekers, the 27%, did not just stay flat. Their solve rate improved by 5 percentage points relative to their own pretest. Their skip rate increased by only 4 points, which was actually less than the control group's increase. They had used AI during the learning phase. When the AI was removed, they performed better than they had at the start. In this group, the data does not show harm. It shows improvement.

The 12% who voluntarily abstained from AI, despite having it available, did the best of all. Their solve rate improved by 11 percentage points relative to their own pretest, the largest improvement of any group in the study, including the controls who never had AI access. The choice to engage your own brain, when the easy option was sitting right there, turned out to be the strongest predictor of improvement.

This is the part the headlines miss. Inside a single condition of a single experiment, with the same students and the same tool, three completely different outcomes emerged. What separated the three groups was not whether AI was available. They all had it. What separated them was how each one chose to use it.

The Bastani group we did not talk about

The Bastani team also ran a third arm of their study. Same 1,000 Turkish high school students. Same math problems. Same lesson, same exam. The only thing that changed was the design of the AI tool the students had access to during practice.

The third arm used a custom GPT-4 interface that the researchers called GPT Tutor. It was instructed, through its system prompt, never to give a student the direct answer. Instead, it provided graduated hints, pointed students toward the next step, corrected reasoning errors, and only revealed parts of the solution as students worked through the problem. Teachers had provided the correct solutions and common student mistakes in advance, so the tutor never made factual errors of its own.

During practice, the GPT Tutor group performed 127% better than the no-AI control group. That is not a typo. They more than doubled their practice performance.

Then the AI was taken away. On the unassisted exam, the GPT Tutor group performed at the same level as the no-AI control. The 17 point crutch penalty that ate the standard ChatGPT group was completely eliminated. The students had practiced with a tool that scaffolded their learning, and the practice held.

The conversation data tells the rest of the story. Students in the standard ChatGPT arm sent an average of about 2 messages per problem. Students in the GPT Tutor arm sent about 5, and the proportion of substantive (non superficial) conversations was roughly twice as high. The ChatGPT students were asking for answers. The GPT Tutor students were working. The tool's design changed how the students used it. The change in use changed the outcome.

What this means

The reframe is simple. In both the Liu data and the Bastani data, there is a version of AI use that hurts learning and a version that helps it. The two versions can look almost identical from the outside. The difference is whether the AI hands you answers or asks you questions back.

The harm in these studies does not track with exposure to AI. It tracks with one specific way of using AI. The hint seekers in Liu's experiment improved their scores. The GPT Tutor students in Bastani's experiment doubled their practice performance without paying any retention penalty. The pattern is consistent. AI used as a tutor builds skill. AI used as an answer dispenser destroys it.

The Bastani finding adds one more thing that matters for what we do next. The hint seekers in Liu's study figured out the right behavior on their own, but they were a minority. Most people, faced with a tool that will hand them the answer, will take the answer. The Bastani study shows that this is not just an individual willpower problem. It can be engineered into the tool itself. Design the AI to refuse to hand over answers and ask questions back, and most students will end up using it in the better way.

If you take only one thing from this entire piece, take this:

The question is not whether you use AI. The question is whether you stay present in your own thinking while you do.

Almost everything else, in terms of what to do about all this, follows from that distinction.

The research points to a fairly specific shape for healthy AI use. It is about the four patterns that separated the 27% of Liu's participants who improved from the 61% who got worse. We can describe them concretely.

Articulate before you check.

Before you ask the AI whether your answer is correct, write down what you think and why. Saying it out loud is even better. The act of putting your reasoning into words is most of the cognitive work. Checking against an authority is a useful capstone. If you skip the articulation, the check is hollow.

Ask for the first step, not the final answer.

When you hit a problem you cannot solve, the instinct is to paste it in and ask "what is the answer." A more useful prompt is "I am stuck. Give me the first thing I should try." This keeps the cognitive work on your side of the table. It also tends to produce better outcomes because you build up the solution yourself, which is when learning actually happens.

Make the AI ask you questions back.

The Liu paper and the Bastani paper converge on the same point: the direction of questioning matters. If you are asking the AI questions, you are extracting information. If the AI is asking you questions, you are doing the thinking. The second mode is the one that builds skill.

You can prompt for this. "Quiz me on the material before you explain it." "Ask me what I think the answer is, then tell me whether I am right." Some tools build this in. Most do not.

Treat the AI like a tutor, not a search engine.

A search engine returns answers. A tutor returns better questions. If your interactions with an AI feel like a long list of "what is X" and "explain Y," you are using it as a search engine. If they feel like a back and forth where you are doing more talking than the AI, you are using it as a tutor. Liu's findings show only one of these modes is actually helping you.

The honest problem with these four patterns is that they are hard to maintain alone. Default tools are designed against you. ChatGPT, Claude, Gemini, and Copilot are optimized to give you what you want, instantly, with no friction. Every product decision in those tools is aimed at "answer faster, answer more completely, refuse less." That is great for productivity. It is bad for learning. The Liu paper says this explicitly:

"AI systems are fundamentally short term collaborators... extraordinarily helpful in the moment, but indifferent to what that help does to the person receiving it over time."

This is the central architectural problem with using a general purpose chatbot as a study tool. You are fighting the tool. Some students can do that consistently. Most cannot, because the path of least resistance is to ask for the answer and move on. The 61% in the Liu study were not less intelligent or lazy students. They were normal people using a tool exactly as it was designed to be used.

This is the part where most coverage of this research ends. "Use AI better." "Just ask for hints." It is sound advice in principle. In practice, asking a student to fight their tool every single time they use it is asking them to rely on willpower that nobody has. The cleaner answer is to use a tool whose design is on your side.

Almost every AI study tool on the market today is a content generator. You upload your slides or notes. The tool spits out flashcards, quizzes, summaries, or podcasts. You consume those outputs.

This is the same trap dressed up in school colors. The thinking has been done for you. You skim a generated summary instead of reading the chapter. You flip through generated flashcards instead of producing your own retrieval cues. You let an AI podcast lull you to sleep instead of struggling with an idea. The cognitive offloading is fully active. You feel productive. You are not learning much.

A few tools have realized this and built differently. They borrow from a 40 year old finding in education research that almost everyone in tech has forgotten: Benjamin Bloom's 2 sigma problem.

In 1984, Bloom published a paper showing that students given one on one mastery tutoring outperformed 98% of their peers in conventional classroom settings. Two standard deviations of improvement. The result has been replicated repeatedly. It is one of the most consistent findings in the entire field of education. Bloom called it a problem because tutoring at that level is economically impossible to deliver at scale. One teacher per 30 students is hard enough. One personal tutor per student is fantasy.

Until now. The reason people are excited about AI in education, when they think about it carefully, is that it is the first technology that could plausibly deliver mastery tutoring to every student on earth. Not by replacing teachers. By giving every student access, for the first time, to the kind of patient, adaptive, one on one teaching that the wealthy have always paid for.

The catch is that this only works if the AI behaves like Bloom's tutor. Patient. Probing. Slow with answers and fast with questions. Watching what the student knows and what they only think they know. Returning to weak spots until they are solid.

That is not what ChatGPT is. That is not what NotebookLM or Quizlet is. That is what a properly designed AI tutor is. The Bastani study tested exactly this, and the result was a 127% practice improvement that survived the removal of the AI.

That is what we tried to build with Akono. It is a private one on one tutor that runs sessions against your own course materials, asks more questions than it answers, and is designed around a single principle: a study tool should build your brain, not replace it. If you want to try it on a course you are actually taking, the first session is free.

Go back to what we opened with. The problem you used to be able to do alone, you can now finish in 15 minutes by asking ChatGPT for the answers. Or you can finish it in two hours by working through it with a tool designed to teach you. The first option saves time tonight. The second option saves you the next time you have to do this on your own.

The research is not telling you to stop using AI. The research is telling you that the AI you use is shaping the brain you will have in five years. That is a real thing. It is not a metaphor. The Liu paper shows the effect emerges in 10 minutes. The Kosmyna paper shows it tracks at the neural level. The Bastani paper shows that a different tool design produces the exact opposite outcome.

You have a choice every time you sit down to study. It is not between using AI and not using AI. That ship has sailed. The choice is between an AI that does the thinking for you and an AI that helps you do the thinking. Both are available. They are not the same thing.

We will leave you with the line from the Liu paper that we think captures the whole question best:

"We hope that our work inspires the field to think about optimizing not just what people can do with AI, but what they can do without it."

That is the test. When the AI is gone, what is left of you?

• Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. [preprint] arXiv:2604.04721.
• Kosmyna, N., Eugene H., Ye Y., Jessica S., Xian-Hao L., Ashly B., Iris B., Pattie M. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. [preprint] arXiv:2506.08872
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26). https://doi.org/10.1073/pnas.2422633122
• Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One to One Tutoring. Educational Researcher, 13(6), 4 to 16. https://doi.org/10.2307/1175554