Cognitive Debt: How to Tell If AI Is Quietly Eroding Your Thinking

Here is the AI risk I think is much harder to notice than hallucinations: the work keeps getting done. The analysis gets finished. The presentation looks good. The email goes out. The recommendation sounds convincing, and you may even be producing more than you used to.

Then somebody asks a question you were not expecting. “Why did you choose that assumption?” “What would happen if churn moved five points?” “Which of these two arguments do you believe?” “Where did that number come from?” And you notice a strange hesitation. Six months ago, you would have had a rough answer almost immediately. Now your first instinct is to open Copilot, Claude or ChatGPT.

Nothing dramatic happened. You did not wake up unable to think. You simply stopped getting as many repetitions of a skill because AI quietly started doing that part first.

That is the version of cognitive debt I think professionals should pay attention to. Not “AI is making us stupid.” The research is nowhere near strong enough to support a claim that broad. A more defensible concern is narrower.

Those are different things, and the distinction changes what we should do about the problem. The answer is not avoiding AI. The answer is becoming much more deliberate about what we offload, what we keep practicing, and which judgments we still need to own when the AI is not there to help us.

Cognitive offloading is not automatically a problem

Humans have been outsourcing cognition for a long time. We write things down instead of remembering everything. We use calculators instead of doing long division in our heads. GPS handles navigation, and search engines retrieve information we once might have memorized. That process is called cognitive offloading, and it can be useful. A 2026 Acta Psychologica experiment, for example, let people type notes during a memory task and found that note-taking helped most when the load exceeded what working memory could hold, although people also introduced copying and reading errors into their own notes.^[1] That study was not about generative AI, but the broader lesson matters: reducing mental load is not inherently bad.

I do not want employees wasting cognitive energy memorizing boilerplate, reformatting tables or manually sorting information simply to prove their brains still work. The question is what gets offloaded. If AI takes over work you do not need to retain as a human capability, excellent. If AI handles a skill you are still expected to use under pressure, defend in a meeting, supervise in someone else’s work or recover when the system fails, I would pay closer attention. That is where convenience can become debt.

What the research actually tells us so far

There are reasons to take this concern seriously. There are also good reasons not to sensationalize the evidence.

Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI at work. Higher confidence in AI was associated with less self-reported critical-thinking effort, while higher confidence in one’s own ability was associated with more critical thinking. Critical thinking also did not disappear when people used AI. It shifted toward verifying information, integrating responses and supervising the task.^[2]

That finding is more interesting to me than “AI reduces thinking,” because AI changes where the thinking happens. Sometimes that is exactly what we want. I do not need to spend 20 minutes producing a first draft if AI can give me one in 30 seconds and I can spend those 20 minutes challenging the argument. But there is a catch: that only works if I actually challenge the argument. If the output arrives polished, sounds plausible and gets accepted because I am busy, the intended shift from creation to evaluation never fully happens. That is where the risk starts.

The MIT study is interesting, but do not turn it into “ChatGPT damages your brain”

You may have seen headlines about an MIT Media Lab study described as evidence that ChatGPT “weakens the brain.” That is a much bigger claim than the study supports. Researchers had people write essays in three conditions: with an LLM, with a search engine, or with no digital assistance. Fifty-four participants completed the first three sessions, and 18 returned for a fourth crossover session. The team measured EEG activity and examined the essays, recall and participants’ sense of ownership. The LLM group showed the weakest connectivity patterns during the task, reported the lowest ownership of their essays and had more difficulty accurately quoting their own work. The authors use the phrase cognitive debt in the paper’s title.^[3]

Interesting? Absolutely. Proof that regular workplace AI use damages cognition? No. The paper is a preprint rather than a peer-reviewed workplace study, the sample is small, the task was essay writing, and the findings should not be generalized to every kind of professional AI use. The useful question it raises is narrower.

That is worth investigating.

Learning research gives us another warning

A 2025 randomized controlled trial assigned 120 undergraduate business students to study using either ChatGPT or traditional methods. Eighty-five returned for a surprise retention test 45 days later. The ChatGPT group scored 57.5% and the traditional-learning group scored 68.5%.^[4] Again, this is an educational study. I would not use the result to claim that a finance professional using Copilot for a monthly report will lose 11 percentage points of memory. What it does support is a principle learning designers have known for a long time: effort is sometimes part of how learning happens. If AI removes every difficult part of acquiring a capability, we should not be surprised if the capability develops differently.

Then there is a striking medical example

A 2025 multicenter observational study followed experienced endoscopists at four centers in Poland before and after their clinics introduced AI-assisted polyp detection. When those clinicians later performed standard colonoscopies without AI, adenoma detection fell from 28.4% before AI exposure to 22.4% afterward. The authors described this as a possible negative effect of continuous AI exposure, while the observational design cannot establish that AI caused the decline.^[5]

This is a specialized medical task, not ordinary office work, and AI-assisted colonoscopy itself has demonstrated benefits. A separate meta-analysis of 28 randomized trials found that AI assistance raised adenoma detection by about 20% relative to routine colonoscopy.^[6] That combination is exactly why the issue is interesting. AI can improve assisted performance while raising questions about unaided capability. Both ideas can be true.

So, is AI making us worse thinkers?

The best answer I found comes from a 2026 article in Trends in Cognitive Sciences that asks almost exactly that question. Its authors argue that offloading cognition to AI can impede skill acquisition and lead to skill decay, which is a reasonable worry if a skill develops through practice and the practice gets handed away. They are more tentative about whether our basic cognitive abilities erode, and suggest those may prove more resilient. And they stress that the risks depend on how the AI is used.^[7] Reviewing, challenging or learning with AI is different from delegating the cognitive process completely.

That is the distinction I would keep. The question is not “Is AI making my brain worse?” Ask instead: “Which skills am I getting fewer repetitions of because AI now goes first?” That is specific enough to do something about.

The danger is not dependence. It is invisible dependence.

I rely on a calculator, and that does not worry me. If somebody asks me to multiply 76,483 by 19 in a leadership meeting, I am perfectly happy admitting I need a calculator. There are other abilities I would be uncomfortable losing. If I am responsible for a recommendation, I want to be able to explain the argument without reopening the AI conversation. If I approve a forecast, I want enough numerical intuition to notice when the output is wildly implausible. If I manage somebody, I want to form my own read of the situation before an AI summary tells me what the situation means. If I publish an article, I want to know why I believe the argument.

That gives me a practical rule, and it is a much more realistic goal than telling busy professionals to do everything manually first.

The Cold Capability Check

Choose two or three capabilities that matter enough in your role that you would feel exposed if you could not perform them without AI. Then test three things.

1. Can I start it?

Give yourself ten minutes with a real piece of work and no AI. Not a toy exercise. Can you produce a credible first pass? It does not need to be polished. You are checking whether the machinery still starts.

2. Can I explain it?

Take a recent AI-assisted deliverable and close the document. Can you explain the core argument, the important numbers, the assumptions and the recommendation in your own words? If you need the document open simply to remember what you approved, pay attention.

3. Can I challenge it?

Take a plausible answer, preferably one you did not write. Can you identify where the reasoning may fail? What assumption is carrying too much weight? What evidence is missing? What would change the answer? This third test matters because modern AI work increasingly moves humans from producing toward supervising. A supervisor who cannot challenge the output is not supervising much.

You do not need a numerical score. The useful signal is the mismatch. If AI is your default for a capability and one of those three questions suddenly feels uncomfortable, that is the skill I would put back into practice.

What should you test?

This depends on your job, which is why I would not give everyone the same cognitive-debt assessment. These are the capabilities I care about, not whether someone remembers the syntax of a prompt framework.

Three signs you are accumulating cognitive debt

AI always gets the first word

This is one of the easiest habits to notice. You receive a problem, and your hand goes straight to the AI tool. Nothing is inherently wrong with that. For plenty of tasks, asking AI first is exactly what I would do. The problem comes when AI becomes the default first move on decisions where your own initial judgment matters. The first plausible answer creates an anchor, and now you are reacting to someone else’s framing rather than discovering your own.

You approve arguments you cannot restate

This one bothers me more than imperfect recall of exact wording. If you created a 20-slide deck with AI yesterday, can you explain the argument today without opening the deck? Not every detail. The logic. What are we recommending? Why? What evidence carries the case? What are we assuming? What is the strongest objection? If you cannot answer those questions, you may have reviewed a deliverable without fully owning the reasoning inside it. The MIT essay experiment is relevant here because the LLM-assisted participants were less successful at quoting their own work and reported lower ownership.^[3] That does not prove the same thing happens in a strategy deck, but the workplace version is easy to test yourself. Close the file and explain the argument. It takes two minutes.

Polished output lowers your suspicion

I call this the polish effect. A badly written answer invites scrutiny, while a beautifully written answer often feels finished. That is dangerous because fluency and correctness are separate properties. The Microsoft and Carnegie Mellon participants described needing to shift critical thinking toward verifying, integrating and supervising AI-generated material, and greater confidence in AI went with lower self-reported critical-thinking effort.^[2]

So I would not tell employees “Always double-check everything AI produces.” Nobody will sustain that. I would teach proportional checking. Ask: What happens if this is wrong?

• If the answer is “almost nothing,” move on.
• If the answer is “a customer gets the wrong price,” check the price.
• If the answer is “leadership makes a decision from this number,” trace the number to source.
• If the answer is “this affects somebody’s employment,” raise the standard again.

Good AI judgment includes knowing where your attention is worth spending.

Draft first or ask first

The sign about AI getting the first word leads to a simple default, and it is a default, not a law.

When I would draft first

Suppose you are an HR business partner reviewing interview feedback. You have notes from three panelists. You can paste everything into AI and ask, “Should we hire this candidate?” I would not. Before AI summarizes anything, I would write three lines:

• My current read: Strong hire, with one concern around ambiguity.
• Why: Excellent case performance and strong evidence of analytical discipline.
• What would change my mind: Evidence that the candidate requires overly defined briefs to perform.

Now use AI. Ask it to challenge your read against the interview material. Did you overweight one interview? Did you miss conflicting evidence? Is your concern supported? Did another interpretation fit the evidence better? AI now has something valuable to work against, and you have a record of what you thought before the machine supplied an answer. That makes the interaction much more useful. I go further into this in my guide to using AI as a thinking partner to sharpen judgment instead of replacing it.

When I would happily ask AI first

Now imagine you have a 42-page regulatory document you did not write and need to understand before tomorrow morning. I am not going to insist you spend two hours manually summarizing it to preserve your brain. Ask AI first. Ask for the structure, which sections apply to your question and where the source locations are. Then check the claims that affect your decision against the original document. The same goes for formatting, cleaning text, boilerplate, converting notes into a table, comparing versions, extracting dates and organizing material you already understand. There is no prize for manually performing cognitive labor that your role does not need you to preserve.

That is why I think the simplistic advice to “always think before using AI” goes too far. The better question is this.

If yes, draft first. If no, let AI save you the work.

A five-minute “my read before AI” note

You do not need to write an essay before every important prompt. For consequential work, five minutes is enough to jot down the following.

Then give that to AI and ask for opposition, using a prompt like the one below. It is a much better thinking workflow than asking “What should I do?”

One weekly rep is enough to learn something

You do not need a 30-minute cognitive fitness regime. Choose one capability you want to keep, and once a week use that capability before AI. If you are an analyst, make one estimate cold. If you are a marketer, write one campaign thesis yourself. If you are an HR partner, form one independent read before asking AI to summarize the case. If you are a leader, frame one decision before asking AI for options. Then compare. Did AI catch something you missed? Good. Did AI change your view? Why? Was the machine wrong, and did you notice? Did your unaided attempt feel easier than last month? The goal is not beating AI. The goal is maintaining enough capability to use AI intelligently.

This is not about keeping humans artificially busy

“Keep practicing unaided” can turn into bad management fast. Imagine telling an employee, “I know Copilot creates this report in ten minutes, but once a week I need you to spend two hours doing the whole thing manually so your skills do not decay.” No. The point is not recreating obsolete labor. Keep the smallest meaningful repetition of the underlying capability. If AI writes the report, perhaps the analyst still writes the initial interpretation of the three largest variances. If AI generates the presentation, perhaps the leader still writes the one-sentence decision and three reasons before asking AI to structure the deck. If AI summarizes interviews, perhaps the hiring manager still writes an independent candidate read before opening the summary. You do not need to preserve the old process. You need to preserve the judgment the new process still depends on, which is a much cheaper form of insurance.

Capability insurance

This does not apply to everything. I do not need “formatting insurance,” or a weekly exercise where I manually rewrite meeting notes into a table. I might need estimation insurance if I approve forecasts, argument insurance if I advise executives, diagnostic insurance if people come to me when normal processes fail, or interpersonal judgment insurance if my job affects hiring, performance or leadership decisions. Protect what the role still needs. Offload the rest.

What this means for managers

Managers cannot see cognitive debt in an AI adoption dashboard. Prompt volume can rise, weekly active users can rise, and everyone can become a “power user.” None of those numbers tells you whether the team still understands the work. So I would not respond by monitoring people’s brains or forcing “AI-free Fridays.” I would put thinking checkpoints inside existing workflows.

The manager question I would use in reviews

Instead of asking “Did you use AI?” I would ask: “What did AI do, what did you still decide, and what did you challenge?” That question tells you much more. If the employee says, “AI summarized the material, but I chose the recommendation and rejected its first explanation because the source data did not support it,” good. If the answer is “AI analyzed everything and I reviewed it,” I would ask one more question: “What did you disagree with?” Not because every AI output needs to be wrong, but because disagreement tells you the person is still willing and able to inspect the reasoning.

Protect expertise differently by role

I would not create one organization-wide list of “skills humans must never lose.” That sounds reassuring and becomes meaningless quickly. Finance needs enough numerical intuition to challenge forecasts. Marketing needs enough customer and editorial judgment to reject competent but generic content. HR needs enough independent judgment to avoid letting a polished synthesis steer a people decision. Legal needs source traceability and interpretation. Operations needs exception handling. Leadership needs problem framing, tradeoff judgment and decision ownership. So the useful organizational exercise is not “Which tasks should stay human forever?” We do not know what technology will look like forever. Ask: “Which human capabilities would leave us exposed if nobody here could perform them anymore?” That is a much better talent question.

Organizations can accumulate cognitive debt too

Imagine an organization where almost every first draft comes from AI. Every business case. Every strategy document. Every customer communication. Every analysis. Each individual may still review their own work reasonably well, but gradually fewer people get repetitions at forming an original first view from a blank page. Now think about succession. Who learns to do the difficult part? Where do junior employees acquire the pattern recognition that senior employees built by doing hundreds of first passes badly before they became good? What happens when today’s reviewer becomes tomorrow’s decision-maker if they spent their development years mostly approving machine-generated work?

The evidence on this organizational version of cognitive debt is still thin, so I would present it as a management question, not a proven effect. But I think it is an important one. AI can change not only how experts work. It can change how future experts become experts. That deserves attention from L&D, HR and functional leaders, and it is one reason I think well-designed training programs should include practice evaluating AI output, not only producing it.

This changes how I would train people on AI

I would not build an AI curriculum around generating impressive output. That is too easy now. I would put people into situations where the AI output looks good and ask: What is wrong with this? What did the model assume? What information is missing? Which claim would you verify first? What would change your mind? Would you defend this recommendation in front of the CFO? That is much closer to the skill employees need. As AI gets better at producing, humans need more practice evaluating, challenging, integrating and deciding. The Microsoft workplace research points in that direction too, describing critical thinking in AI-assisted work as shifting toward verification, response integration and task stewardship, so I would train those skills explicitly. I cover the personal version of this in my piece on how to keep thinking for yourself when you use AI every day.

What I would not do

I would not ban AI from important work simply because the work matters. Used properly, AI can improve analysis, challenge assumptions and give people access to capabilities they did not have before. I would not require people to manually perform every task “to keep their skills sharp,” because that wastes capacity. I would not interpret lower mental effort as automatically bad, since reducing unnecessary cognitive load is one of the reasons tools exist. And I would not claim that the neuroscience has settled this question. The 2026 Trends in Cognitive Sciences article is appropriately cautious: skill acquisition and retention deserve concern, but the case that AI erodes basic cognition is much less settled.^[7]

The goal is not maximum mental effort. The goal is putting human effort where we still need human capability.

The practical rule I would leave on your desk

When you are about to hand something to AI, ask: Is this load, or is this a capability I still need? If the work is load, offload aggressively: summarization, formatting, sorting, transformation, searching, first-pass synthesis of material you did not create. If the work contains a capability you will still be expected to defend, supervise or recover without assistance, keep enough of the capability in your hands to know what good looks like. Then use AI around it.

That is the relationship I want. Not human versus AI. Not “AI does the boring work and humans do the creative work,” because real jobs are messier than that. Something more practical.

What I would do this week

Do not audit everything. Pick one. Choose a capability you use often enough that losing fluency would bother you. Maybe estimating, argument building, problem framing, diagnosing an exception, reading a customer situation or evaluating evidence. Before AI touches one real task this week, give yourself ten minutes and produce the rough version yourself. Then bring in AI and compare the two. What did AI add? What did you catch that AI missed? What did AI change your mind about? Could you explain the final answer tomorrow without opening the conversation?

That tiny experiment will tell you more about your relationship with AI than your prompt count ever will. The risk is not that AI suddenly steals your ability to think. The more plausible risk is quieter: you become excellent at getting work done with AI while gradually becoming less practiced at one of the capabilities the work still requires. The best time to notice that gap is while AI is still helping you perform well, not the day somebody closes your laptop and asks, “What do you think?”

Sana Mian, Co-Founder of Future Factors AI

Sana is an AI educator and learning designer specializing in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →