AI Should Stand for Artificial Intelligence, Not Actual Ignorance

How to use AI as a power tool without letting it take your skills.

Every claim links to its source.

Let me start by telling on myself

I use AI every day.

I write code with it. I write content with it. And yes, I used it to research, draft, and help build the article you are reading right now. I am not going to hide that. It would be a strange thing to write about honest AI use while covering up my own.

So here is what it actually does for me.

For coding, it saves an enormous amount of time. Writing tests. Reading code I did not write. Checking my own work. Looking up the twenty small things I used to open twenty browser tabs for. Work that used to eat an afternoon now takes an hour.

For research, same story. I can read five papers in the time it used to take me to find them.

And for my own business, I run managed AI agents. They draft replies to email. They set up meetings. They take a photo of a business card and turn it into a clean CRM record. Those three jobs alone save tens of hours a week. That is not a sales line. That is my actual calendar.

But I follow one rule and I do not break it.

The agent does not act until I approve the step.

It drafts, I read. It suggests a meeting time, I confirm. It builds the CRM record, I check the record. Nothing leaves my office with my name on it that I have not looked at first.

I used to think of that rule as me being careful. Then I started reading the research, and it turns out that rule is the whole ballgame. It is the single thing that separates people who get faster from people who get hollow.

Which brings me to the question I actually want to ask.

After a hundred drafts that I only had to approve, am I still someone who can sit down and write a hard email cold?

That is what this article is about. Not “is AI good or bad.” That question is boring and already settled. It is a tool. The useful question is much narrower and much more practical. Which parts of the work can you hand over safely, and which parts get expensive when you hand them over?

The research on this is better than most people think. Let me walk you through it, including the parts that argue against me.

Part 1: The worry is real, and one study proves it cleanly

Most “AI is making us dumb” articles lean on shaky sources. So let me start with the strongest one instead.

In June 2025, researchers from the University of Pennsylvania and a Turkish school system published a study in PNAS. That is one of the most respected science journals in the world. They ran a real experiment with about 1,000 high school students learning math.

They split the students into three groups.

1. No AI. The control group.
2. Plain ChatGPT. Full access, no rules.
3. A tutor version of ChatGPT. Same model. But it was told to give hints one step at a time and never hand over the final answer.

Then they measured two different things. How students did on practice problems with the AI. And how they did later on an exam without it.

Here is what came back.

Read that top row twice.

The students with plain ChatGPT looked like they were learning. They were faster and more accurate on every practice problem. Then the tool was taken away, and they did worse than students who never had it at all. The researchers said those students were using it as a “crutch.”

Now look at the second row. Same model. Same students. Same subject. The only difference was how the tool was set up. And the harm went away completely.

Source: Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS 122(26), 2025.

That is my whole argument in one experiment. The danger is not the AI. The danger is the default setting. Ask, receive, move on.

It is not just students

A team from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real cases of using AI at work. Their finding, published at a top peer reviewed conference in 2025: the more confidence people had in the AI, the less critical thinking they reported doing. The more confidence they had in themselves, the more they did.

They found something subtler too. AI does not take thinking out of your job. It changes what kind of thinking you do. You move from making things to checking things. That is a different skill, and most of us have never practiced it.

Source: Lee, Sarkar, Tankelevitch and others, “The Impact of Generative AI on Critical Thinking,” CHI 2025. One caveat worth knowing. This is a survey of what people say about themselves, not a test of their actual thinking. The direction is clear. The size is not.

The trap that catches smart people

Harvard Business School researchers ran an experiment with 758 consultants at Boston Consulting Group. On 18 realistic tasks that AI is good at, the consultants using GPT-4 finished 25% faster and produced work rated more than 40% higher in quality. That is a huge win.

Then the researchers gave them one task built to sit just outside the AI’s ability. The kind of problem where the model gives you a confident, sensible sounding, wrong answer.

On that task, consultants using AI were 19 percentage points less likely to get it right than consultants with no AI at all.

They called this the “jagged technological frontier.” AI is not evenly smart or evenly dumb. It is excellent at some things and terrible at things that look identical from the outside. And it gives you no warning about which one you are standing in.

Source: Dell’Acqua, McFowland, Mollick and others, “Navigating the Jagged Technological Frontier,” Organization Science 37(2), 2026.

The study you have already seen, and why I am not leaning on it

You may have run into headlines about an MIT Media Lab study called “Your Brain on ChatGPT.” It put EEG caps on 54 people writing essays. The ChatGPT group showed the weakest brain connectivity and could barely quote their own essays back.

It is an interesting study and it made the phrase “cognitive debt” famous. But I want to be straight with you. It has not been peer reviewed. It has 18 people per group. The authors themselves say the results should be treated as preliminary. A formal critique has since been published questioning the methods.

I bring it up because you will see it everywhere. And knowing which evidence is strong and which is just loud is exactly the skill this article is about.

Part 2: Why our brains work this way

None of this is new, and none of it is really about AI. It is one of the oldest findings in learning science.

You remember what you make

When you produce an answer yourself, you remember it far better than when you read the same answer. This was shown in 1978 and has held up across 86 studies with more than 17,000 people.

Reading an AI’s answer is the “read” condition. Writing your own answer first and then comparing is the “make” condition. Same amount of time. Very different result.

Effort is the price of keeping something

Psychologists Robert and Elizabeth Bjork call this “desirable difficulties.” Their main point is uncomfortable. The conditions that make learning feel fast and smooth are usually the ones that make it disappear.

Here is the classic test. Students read a passage. Then some reread it and some took a practice test. Five minutes later, the rereaders scored higher. One week later, the test group scored 56% and the rereaders scored 42%. And the rereaders had been the more confident group.

Source: Roediger and Karpicke, “Test-Enhanced Learning,” Psychological Science 17(3), 2006.

That gap between confidence and ability is the whole problem in miniature. Easy feels like learning. It usually is not.

What “use it or lose it” looks like

London taxi drivers who memorize the city’s streets have measurably more grey matter in the part of the brain used for navigation. The difference grows with years on the job.

Source: Maguire and others, PNAS 97(8), 2000.

Run it the other way and you get the mirror image, though the evidence is weaker. Drivers with heavy lifetime GPS use had worse spatial memory. In a small follow up group, the heaviest users declined the most over three years.

Source: Dahmani and Bohbot, Scientific Reports 10:6310, 2020. Caveat: the follow up group was only 13 people. Suggestive, not settled.

The detail that saves the whole story

This is my favorite study in the pile, because it tells you exactly what to do.

Researchers walked people through a museum. Some photographed the objects. Some just looked. The photographers remembered the objects worse.

But the effect disappeared when people zoomed in on a detail before shooting.

Source: Henkel, “Point-and-Shoot Memories,” Psychological Science 25(2), 2014.

Handing the job to the camera hurt when it replaced attention. It did not hurt when it directed attention.

That is the difference between a tool and a crutch. And it is a difference you control.

Part 3: The other side. AI really does save time.

If I only told you the scary half I would be doing the same thing the panic articles do. The time savings are real and well documented.

• Writing. 444 college educated professionals doing realistic writing tasks. With ChatGPT, time dropped about 37%, from 27 minutes to 17. Quality went up. (Noy and Zhang, Science, 2023)
• Customer support. 5,172 support agents at a large software company. 15% more issues resolved per hour. Almost all of the gain went to newer and lower performing agents. (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025)
• Software development. Three field experiments across Microsoft, Accenture and a Fortune 100 manufacturer, with 4,867 developers. About 26% more tasks completed. Again, most of the gain went to juniors. (Cui and others, Management Science, 2026)

Notice the pattern. AI helps the least experienced person the most. For a small business that matters a lot. It means a new hire can answer a customer well in week one instead of month three.

Three honest complications

1. People are bad at knowing whether AI helped them.

A research group called METR ran a careful trial with 16 experienced open source developers working on their own real code. Before starting, they predicted AI would make them 24% faster. They were actually 19% slower. Afterward, having done the work, they still believed AI had made them 20% faster.

Source: METR, 2025. Caveat: 16 developers, all experts in code they knew deeply. Do not read this as “AI makes developers slower.” Read it as: your feeling of speed is not proof of speed.

2. The average effect is real but smaller than the headlines.

A 2026 review that pooled 23 studies of AI coding assistants found a real but moderate productivity gain, and no measurable learning benefit at all. Gains were much bigger in lab settings than in real workplaces.

Source: Maier and others, 2026 (preprint).

3. Hours saved is not the same as value created.

A study of about 25,000 Danish workers across 11 AI exposed jobs found average time savings of 2.8% of work hours. Pay and hours worked barely moved. And most of the new work AI created was oversight. Reviewing, checking, integrating, staying compliant.

Source: Humlum and Vestergaard, NBER WP 33777, 2025 (working paper).

That last point deserves its own line. AI does not delete work. It turns making into checking. And checking only works if you still know what good looks like.

Part 4: Where AI actually makes you better

Now the useful part. There is real evidence that AI can build skill instead of replacing it. But only in specific setups.

A tutor that refuses to give answers

Harvard ran a randomized trial with 194 physics students. One group got a normal, well taught class using active learning. The other got a GPT-4 tutor.

The AI tutored students learned more, in less time. Median 49 minutes against about 60. The effect was large, and they reported higher engagement.

But look at how the tutor was built, because that is the whole trick. Its instructions included:

• “DO NOT give away the full solution”
• “Only give away ONE STEP AT A TIME”
• “Encourage them to give it a try first”
• Keep replies to a few sentences
• And this one matters most: the correct worked solution was written into the instructions by the instructors, so the model was not making up physics on the fly

Source: Kestin, Miller, Klales, Milbourne and Ponti, “AI tutoring outperforms in-class active learning,” Scientific Reports 15:17458, 2025.

Same technology as the students who got worse. Opposite result. The difference was the guardrails.

An AI that asks instead of tells

In a study of 204 people, an AI that turned its information into questions was better at helping people spot flawed logic than an AI that simply explained things.

Source: Danry, Pataranutaporn, Mao and Maes, “Don’t Just Tell Me, Ask Me,” CHI 2023.

That lines up with a much older finding. Asking people to explain things to themselves produces a solid gain in understanding across 69 studies, with no new content taught at all.

Source: Bisra, Liu, Nesbit, Salimi and Winne, “Inducing Self-Explanation: a Meta-Analysis,” Educational Psychology Review, 2018.

An AI that makes the human better at the job

Stanford built a tool called Tutor CoPilot. It gave 900 human tutors real time AI suggestions during live sessions with 1,800 students. Student mastery rose 4 percentage points overall, and 9 percentage points for students of the lowest rated tutors. Cost was about $20 per tutor per year. Looking at 550,000 messages, those tutors asked more guiding questions and gave fewer direct answers.

Source: Wang, Ribeiro, Robinson, Loeb and Demszky, 2024 (preprint, not yet peer reviewed).

The AI was not teaching the student. It was coaching the human who was teaching the student. That is the shape worth copying.

Part 5: Nine rules that keep the thinking in your head

Here is the practical part. None of these mean using AI less. They change how you use it.

1. Draft first, then ask. Write two rough sentences, a rough outline, or a rough guess before you open the chat. Then ask. You have now made something instead of just receiving something, and the research on making versus reading is on your side. Cost: ninety seconds.

2. Decide before you look. For anything that matters, like a hiring call or a pricing decision, write down your own answer before you ask the AI. Harvard researchers tested this exact design with 199 people. It cut blind acceptance of wrong AI answers by a lot. They also found people disliked the designs that worked best. Effective and annoying tend to travel together.

Source: Buçinca, Malaya and Gajos, “To Trust or to Think,” CSCW 2021.

3. Ask it to question you, not answer you. Try this. “Do not give me the answer. Ask me three questions that will help me find it myself.” Or this. “Argue the opposite of what I just said, as hard as you can.” You just turned an answer machine into a sparring partner.

4. Explain it back. After the AI explains something, close the window and say it out loud in your own words. If you cannot, you did not learn it. You just felt like you did.

5. Verify like a professional, not like a fan. LLMs make things up with total confidence. The word for this is hallucination. A Stanford study found that general purpose chatbots got legal questions wrong at least 58% of the time when asked about specific real cases. Purpose built legal AI tools, the expensive ones, still hallucinated 17% to 33% of the time.

Sources: Dahl and others, Journal of Legal Analysis, 2024 and Magesh and others, Journal of Empirical Legal Studies, 2025 These tested 2023 and 2024 models on hard lookup tasks. Today’s tools are better. Better is not the same as trustworthy.

Simple rule. Anything with a name, a number, a date, or a citation in it gets checked. Those are exactly the things models invent.

6. Keep your manual reps. Aviation figured this out decades ago. The FAA told airlines to have pilots hand fly regularly, because constant use of autopilot “could lead to degradation of the pilot’s ability to quickly recover the aircraft from an undesired state.”

Source: FAA SAFO 13002, 2013, updated as SAFO 17007.

Pick one thing a week you do entirely by hand. Write one proposal cold. Do one estimate without a model. It does not have to be efficient. It has to keep the muscle.

7. Hand over the boring. Keep the judgment. Formatting. Reformatting. Transcribing. Summarizing a document you already understand. Turning a messy list into a clean table. Drafting the fourth version of a routine email. Hand all of that over and do not feel bad about it. Nobody’s mind was ever built by reformatting a spreadsheet.

Keep for yourself: deciding what matters, choosing between real options, judging whether an answer is right instead of just plausible, and anything you would have to defend to a customer.

8. Notice the confidence gap. The METR developers felt 20% faster while being 19% slower. Once a month, actually time something. Feelings are not data.

9. Watch the jagged edge. The moment a task feels a little unusual, an odd rule, a weird edge case, a customer situation you have never seen, assume you may have stepped off the frontier. That is the moment to slow down and think first. Not the moment to ask faster.

Part 6: The honest bottom line

Let me give you both halves, because you deserve both.

The evidence that AI helps is stronger than the panic suggests. A 2026 review that pooled 35 experimental studies with 4,193 people found a solid positive effect of ChatGPT on student learning, including critical thinking (Wu and others, Humanities and Social Sciences Communications, 2026). Across the measured outcomes, the research currently leans positive.

And the evidence for harm is narrower, but sharper, than the boosters admit. The harm shows up in one specific place. Unaided performance, later. Take the tool away, wait a while, and see what is left. That is where the Turkish math students lost 17%. That is also where the OECD’s 2026 education report lands, with a phrase worth memorizing: performance is not learning. They call the failure mode “metacognitive laziness,” which is a fancy way of saying the brain stops managing itself when something else is doing the managing.

Source: OECD Digital Education Outlook 2026.

So the claim I will actually defend is not “AI makes you stupid.” It is this.

Handing over the effortful part of a task degrades that skill. Handing over the tedious part does not.

Every study that looked carefully found a knob that flipped the result. The museum photos stopped hurting memory when people zoomed in. The math students stopped losing ground when the tutor withheld answers. The physics students actually gained when the AI was built to make them work.

It is always the same knob. Did the tool replace your thinking, or point it somewhere?

What this means if you run a small business

You do not have the option of refusing AI. Your competitor with three employees and good automation is beating your five employees without it. That is just true.

But you also cannot afford to become a business where nobody knows why anything works. When the unusual customer shows up, the one the model has never seen, somebody has to actually know the answer.

So the move is not less AI. The move is deliberate AI.

• Automate the repetitive, the administrative, the formatting, the follow up.
• Keep human judgment on pricing, hiring, exceptions, and anything a customer will hold you to.
• Build your tools with guardrails on purpose. Hints instead of answers. Questions instead of conclusions. A human in the loop wherever being wrong actually costs something.
• Train your people with the AI, not around it. The evidence is consistent. The newest person on your team gains the most.

That is why my own agents draft and I approve. Not because I do not trust the technology. I trust it quite a lot. I approve because approving is the part that keeps me sharp, and because my name is on the output.

Artificial intelligence is a spectacular lever. Levers only work if somebody is standing on the other end who knows which way to push.

Sources

Evidence on cognitive cost

1. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., and Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26).
2. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., and Wilson, N. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025.
3. Dell’Acqua, F., and others (2026). Navigating the Jagged Technological Frontier. Organization Science, 37(2).
4. Fan, Y., and others (2025). Beware of metacognitive laziness. British Journal of Educational Technology.
5. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1):6. Survey based, so correlation rather than cause. A correction was published.
6. Kosmyna, N., and others (2025). Your Brain on ChatGPT. arXiv preprint. Not peer reviewed. See the published critique.

How learning actually works

7. Bertsch, S., Pesta, B., Wiscott, R., and McDaniel, M. (2007). The generation effect: A meta-analytic review. Memory & Cognition, 35(2).
8. Bjork, E. L., and Bjork, R. A. (2011). Making Things Hard on Yourself, But in a Good Way.
9. Roediger, H. L., and Karpicke, J. D. (2006). Test-Enhanced Learning. Psychological Science, 17(3).
10. Bisra, K., Liu, Q., Nesbit, J., Salimi, F., and Winne, P. (2018). Inducing Self-Explanation: a Meta-Analysis. Educational Psychology Review, 30(3).
11. Henkel, L. (2014). Point-and-Shoot Memories. Psychological Science, 25(2).
12. Maguire, E. A., and others (2000). Navigation-related structural change in the hippocampi of taxi drivers. PNAS, 97(8).
13. Dahmani, L., and Bohbot, V. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10:6310.
14. Risko, E., and Gilbert, S. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9).

Evidence on time and productivity gains

15. Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654). Open PDF.
16. Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2).
17. Cui, K. Z., and others (2026). The Effects of Generative AI on High-Skilled Work. Management Science.
18. Becker, J., Rush, N., Barnes, E., and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR.
19. Maier, S., and others (2026). A meta-analysis of the effect of generative AI on productivity and learning in programming. Preprint.
20. Humlum, A., and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. NBER WP 33777.
21. Bick, A., Blandin, A., and Deming, D. (2026). The Rapid Adoption of Generative AI. Management Science.

Evidence on solutions

22. Kestin, G., Miller, K., Klales, A., Milbourne, T., and Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15:17458.
23. Wang, R. E., Ribeiro, A., Robinson, C., Loeb, S., and Demszky, D. (2024). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. Preprint.
24. Danry, V., Pataranutaporn, P., Mao, Y., and Maes, P. (2023). Don’t Just Tell Me, Ask Me. CHI 2023.
25. Buçinca, Z., Malaya, M. B., and Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI. CSCW 2021. Open PDF.
26. Parasuraman, R., and Manzey, D. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
27. Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6).
28. FAA (2013). SAFO 13002: Manual Flight Operations. Updated as SAFO 17007 (2017).

Reliability and limits of AI systems

29. Dahl, M., Magesh, V., Suzgun, M., and Ho, D. E. (2024). Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models. Journal of Legal Analysis, 16(1).
30. Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., and Ho, D. E. (2025). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies, 22(2).

Institutional guidance

31. OECD (2026). Digital Education Outlook 2026.
32. UNESCO (2023). Guidance for generative AI in education and research.
33. UNESCO (2024). AI competency frameworks for students and teachers.
34. Wu, X., and others (2026). ChatGPT’s impact on student learning outcomes: a meta-analysis of 35 experimental studies. Humanities and Social Sciences Communications.

A note on method. I checked every study above against its original source. Where a study is a preprint instead of peer reviewed, or where the sample was small, I said so in the text. If we are going to write about thinking clearly, we should probably show our work.

AI Abstraction builds and manages AI agents for small businesses in New York and New Jersey. Guardrails on, human approval on every action, and a plain English glossary so you always know what you are buying.

About the author

Jay Taheri

Founder & CEO

Jay Taheri is the founder and CEO of AI Abstraction. He has more than 25 years of experience in system design and engineering, programming, system reliability, and machine learning, with roles including Senior Software Engineer at Bloomberg, Radianz, and systems engineer at Credit Suisse. He holds a certificate in Designing and Building AI Products and Services from MIT Professional Education.

Jay founded AI Abstraction in 2024 to help small and mid-sized businesses put AI to work without the buzzwords and complexity that usually come with it. The company designs, builds, and hosts AI agents that handle real, day-to-day tasks, like answering calls, reading email, and qualifying leads, then stays on after launch to keep every agent monitored, tuned, and running.

When you call, you talk to Jay directly, the person who actually builds and runs your agent, not an account manager reading from a script.