Responsible Use: Dependence, skill, and learning

This content reflects the field as of October 2026.

"AI is making people worse at thinking" has become a familiar headline. So has its opposite, that an AI tutor can teach better than a classroom. Each kind of story usually rests on a real study, and the studies behind them differ a great deal in what they can show.

Two experiments published in 2025 make the point. In the first, researchers based mainly at the University of Pennsylvania gave nearly a thousand high school math students access to AI during practice sessions. Students who had a standard chatbot did 48 percent better on the practice problems than students who had no AI. Then the tool was taken away. On the exam that followed, those same students scored 17 percent lower than students who had never had it (Bastani et al. 2025).

In the second, physics instructors at Harvard University compared an AI tutor they had built with a well-taught class that used active learning, meaning students work on problems in class instead of listening to a lecture. Among 194 students, those using the tutor learned more than twice as much, and in less time (Kestin et al. 2025).

Both were controlled experiments with a comparison group. Someone who read only the first would conclude that AI harms learning. Someone who read only the second would conclude the reverse. Read together, they suggest that the result depends on how the tool is built and how it's used.

The rest of the evidence is more uneven. As of October 2026, it includes:

• a small number of careful experiments like these two, mostly with students
• surveys in which working adults report on their own thinking
• small studies, some not yet reviewed by other scientists, that drew very large headlines

That mix supports a narrower claim than either headline. Handing a task to a tool that supplies answers, while you're still trying to learn the task, can leave you worse at it when the tool is gone. The broad claim that AI use damages adult thinking in general hasn't been established. Neither has the claim that it carries no cost.

This gives you a question to put to any report on the subject before acting on it: what kind of study is this? A controlled experiment, a survey, and an unreviewed brain-imaging study are different kinds of evidence, and a headline rarely says which one it's built on. It helps to know who was studied, what was measured, and what they were compared with.

It also leaves a personal question that no study answers for you. Some of the work you now give to AI involves skills you don't need. Some involves skills you'd miss. The research is most useful once you know which is which.

References

• Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. "Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics." Proceedings of the National Academy of Sciences 122 (26): e2422633122.
  • Free: PubMed Central
  • Publisher: PNAS
• Kestin, Greg, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. 2025. "AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting." Scientific Reports 15: 17458.
  • Free: Scientific Reports

Introduction

Worry about AI and thinking often assumes that leaning on a tool for mental work is something new. People have done it for as long as there have been tally marks, and psychologists have a name for it.

This reading explains what cognitive offloading is, what it gains you, what it can cost, and the distinction that decides whether a given case is a problem.

What Offloading Is

Cognitive offloading is using a physical action, such as writing something down or setting a reminder, to reduce the mental work a task requires. The term comes from a 2016 research review by the psychologists Evan Risko, of the University of Waterloo, and Sam Gilbert, of University College London. Their definition is "the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand" (Risko and Gilbert 2016).

Cognitive demand is the amount of mental work, such as remembering, attending, and calculating, that a task requires. An external aid is a tool or record outside your head, such as a list, a calendar, or a calculator, that does part of a mental task for you.

The review's own opening examples are deliberately ordinary: tilting your head to look at a rotated picture, and setting a phone to remind you of an appointment (Risko and Gilbert 2016). Everyday life is full of others.

The authors note that the behavior is everywhere and had only recently been studied in its own right. Their review asks two questions about it: what leads people to offload, and what the consequences are for the mind (Risko and Gilbert 2016).

What You Gain

The gain is easy to see. Memory and attention are limited. A person holding seven errands in mind has less attention left for anything else, and may still forget one. A list holds all seven perfectly and frees that attention.

Offloading also lets people do things they couldn't do unaided. Few people can multiply two five-digit numbers in their heads, and with a calculator anyone can. Seen this way, offloading is a large part of how people get complex work done at all.

An AI assistant extends the same move to work that earlier aids couldn't take over, such as drafting a paragraph, summarizing a report, or outlining an argument.

What It Can Cost

The cost is harder to see, because it shows up later. Retention is how well you keep a piece of knowledge or a skill over time, and it depends heavily on use. A phone number you dial from memory stays with you. One you always tap from a contact list never gets learned in the first place. Many people who navigate everywhere with a map app find they can't retrace a route they've driven a dozen times.

The general point is that what you don't practice, you may not retain. An external aid does the work for you, and that means you aren't doing the work.

This is the consequence side of the question the 2016 review raised. Applied to AI, it leads to a more specific worry than "AI makes you dumber." If a tool drafts every paragraph you write, the part of writing that the tool does is a part you've stopped practicing.

Two Kinds of Task

Whether that matters depends on the task. There are two different situations.

A task you never need to do unaided. Almost nobody needs to do long division by hand at work, and losing the knack costs nothing as long as a calculator is within reach. Offloading here is close to pure gain.

A task you're trying to learn, or need to be able to do yourself. A student learning to solve equations, a new analyst learning to read a set of accounts, and a teacher who must judge student writing on the spot all need the skill itself. If the aid does the work during the period when the skill would be forming, the skill may not form.

The same tool can sit on either side of that line. A calculator is harmless to an accountant and may be a problem for a child learning arithmetic. An AI assistant that summarizes reports is a convenience for someone who only needs the gist, and it may be a cost for someone whose job is to learn to read such reports critically.

The useful question about any offloaded task is whether you still need the skill, and whether you could check the tool's work without it.

Conclusion

Cognitive offloading is the long-standing practice of moving mental work onto external aids, and AI assistants extend it to more kinds of work. It frees limited attention and memory, and it means the offloaded skill gets less practice. Whether that is a cost depends on whether you need to keep the skill.

Key Terms

• Cognitive offloading: Using a physical action, such as writing something down or setting a reminder, to reduce the mental work a task requires.
• Cognitive demand: The amount of mental work, such as remembering, attending, and calculating, that a task requires.
• External aid: A tool or record outside your head, such as a list, a calendar, or a calculator, that does part of a mental task for you.
• Retention: How well you keep a piece of knowledge or a skill over time.

References

• Risko, Evan F., and Sam J. Gilbert. 2016. "Cognitive Offloading." Trends in Cognitive Sciences 20 (9): 676–688.
  • Free: UCL Discovery (accepted manuscript)
  • Publisher: Trends in Cognitive Sciences

This content reflects the field as of October 2026.

Introduction

Schools and universities have been arguing since chatbots appeared about whether students who use them learn more or less. Two controlled studies published in 2025 came to opposite results, and the reason they differ is the most useful finding in either.

This reading describes each study's design and results, what separates them, and the limits on what they show.

The High School Mathematics Study

The first study was run by researchers at the University of Pennsylvania together with a teacher at the school involved. It was a field experiment, meaning it took place in a real school during regular lessons. It was also a randomized experiment: a study in which people are assigned by chance to different conditions, so that differences in outcome can be credited to the conditions.

Nearly a thousand students at a high school in Turkey took part during four ninety-minute practice sessions in the autumn of 2023. Classrooms were assigned to one of three conditions (Bastani et al. 2025).

• No AI. Students practiced with their usual materials.
• A standard chatbot. Students had a chat tool built on GPT-4, a model from OpenAI, that worked like an ordinary chatbot.
• A tutor version. Students had a tool built on the same model with guardrails: limits built into an AI tool that shape what it will do, such as giving hints in place of answers. It was instructed to give hints without giving away the answer, and teachers had supplied it with correct solutions and common student mistakes.

The study measured two things. Practice performance is how well someone does on a task while the AI tool is available to help. Unassisted performance is how well someone does on the same kind of task later, without the tool. Here the second was an exam taken without any AI.

(Bastani et al. 2025)

The students with the standard chatbot looked much better during practice and did worse afterward than students who had practiced with no AI. The authors report that they used the tool as a "crutch," often asking for the solution and copying it. The tutor version raised practice scores even more and largely removed the later harm.

The University Physics Study

The second study was a randomized trial with 194 students in an introductory physics course at Harvard University, run by instructors of the course. Over two weeks, each student learned one topic from an AI tutor and another in class, in a lesson built on active learning, where students work through problems with guidance. Every student took a short test before and after each lesson (Kestin et al. 2025).

Students learned more than twice as much from the AI tutor lesson as from the class, measured by gains from the first test to the second. They also spent less time. Half of the students finished the tutor lesson within 49 minutes, against a 60-minute class.

The tutor was purpose-built. It ran on GPT-4, and the instructors wrote its instructions to follow the same teaching practices as their class, including keeping students actively working. They also gave it complete step-by-step solutions they had written, so that it wouldn't have to produce the physics itself and risk errors (Kestin et al. 2025).

What Separates the Results

The harm in the mathematics study came from answers on demand. A student who can get the solution by asking has no need to work it out, and working it out is where the learning happens. The benefit in both studies came from guided practice: a tool that held back the answer, gave a next step, and kept the student doing the thinking.

So the two studies don't show that AI helps or harms learning in general. They show that a tool designed to keep the learner working can help, and that an unrestricted one used during practice can hurt.

The Limits

Both studies have limits.

• One subject each. One covered high school mathematics and the other introductory physics. Writing or professional skills may behave differently.
• Short duration. One ran over four sessions and the other over two weeks.
• One setting each. One was a single school and the other a single course at a selective university.
• One model. Both used GPT-4, a 2023 model from one developer. Neither tested models from Anthropic, Google, or other developers, or any later model.

The physics authors add a caution of their own. They "advise against the notion" that AI should replace in-class teaching outright (Kestin et al. 2025).

Conclusion

As of October 2026, two controlled studies of students point the same way despite their opposite headlines. A standard chatbot raised practice scores in high school mathematics and lowered later exam scores, while tutors designed to guide instead of answer improved learning in both studies. The evidence covers two subjects, short periods, and one 2023 model.

Key Terms

• Randomized experiment: A study in which people are assigned by chance to different conditions, so that differences in outcome can be credited to the conditions.
• Guardrails: Limits built into an AI tool that shape what it will do, such as giving hints in place of answers.
• Practice performance: How well someone does on a task while the AI tool is available to help.
• Unassisted performance: How well someone does on the same kind of task later, without the tool.

References

• Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. "Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics." Proceedings of the National Academy of Sciences 122 (26): e2422633122.
  • Free: PubMed Central
  • Publisher: PNAS
• Kestin, Greg, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. 2025. "AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting." Scientific Reports 15: 17458.
  • Free: Scientific Reports

This content reflects the field as of October 2026.

Introduction

Three studies from 2025 are often cited as proof that AI erodes the skills of working adults. Each shows less than its headlines said.

This reading describes what the three studies did and found, what kind of evidence each one is, and where the evidence on skill loss is stronger and weaker. The question is contested, and the reading says where.

A Survey of Knowledge Workers

Researchers at Carnegie Mellon University and Microsoft Research ran a survey: a study that asks a sample of people the same questions and reports the share giving each answer. They asked 319 knowledge workers to describe 936 real examples of using generative AI at work and to rate their own critical thinking in each (Lee et al. 2025).

The main finding was a correlation, a pattern in which two things tend to go together, which doesn't show that one causes the other. In the authors' words, "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking" (Lee et al. 2025).

Two limits follow from the design. The measure is what people said about their own thinking, and the authors note that participants sometimes confused less effort with less critical thinking. The pattern also has more than one explanation. People may think less carefully because they trust the tool, or they may trust the tool on tasks that needed little thought.

An Observational Study of Endoscopists

Deskilling is the loss of a skill through lack of use after a tool takes over the task. The study most often cited for it comes from medicine.

Doctors who perform colonoscopies look for precancerous growths, and AI systems can flag possible growths on screen. A study at four centers in Poland compared how often 19 experienced doctors found such growths in procedures done without AI, before and after their centers introduced an AI system. The detection rate fell from 28.4 percent to 22.4 percent (Budzyń et al. 2025).

This was an observational study: a study that records what happens without assigning people to conditions, so it can't rule out other causes for what it finds. Independent experts who commented on it pointed to several. Venet Osmani, a professor at Queen Mary University of London, noted that the number of procedures rose sharply after the AI system arrived, so heavier workloads could explain the drop. Allan Tucker, a professor at Brunel University of London, said randomized trials would be needed for firmer claims (Science Media Centre 2025).

A Disputed Preprint

The third study drew the largest headlines. Researchers at the MIT Media Lab recorded the brain activity of people writing essays in three groups: with an AI chatbot, with a search engine, and with no tools. They reported that the chatbot group showed the weakest brain connectivity (Kosmyna et al. 2025).

The paper is a preprint, a research paper posted publicly before independent experts have reviewed it for a journal. It had 54 participants, of whom 18 took part in the final session.

Another group of researchers has posted a comment on it, also as a preprint. They raise concerns about the small sample, whether the analyses can be reproduced, the methods used on the brain recordings, and inconsistencies in the reported results. They suggest some results "could be interpreted more conservatively" (Stankovic et al. 2025). A weaker brain signal while using a tool also isn't the same as lasting harm. It may show only that the tool was doing some of the work.

The Three Compared

Stronger and Weaker Evidence

The strongest evidence on AI and skill comes from controlled experiments on learning, in which students were assigned by chance to conditions and tested afterward. Those studies found that a tool giving answers on demand can reduce what students learn.

The evidence on working adults is weaker. As of October 2026 it consists of self-reports, a before-and-after comparison, and an unreviewed study. None establishes that AI use causes lasting cognitive harm in adults, and none shows it's safe. Deskilling is plausible from what's known about practice and skill, and the studies that would settle its size haven't been done.

Conclusion

As of October 2026, the evidence on skill loss among working adults is a survey showing a correlation, an observational study with other possible explanations, and a disputed preprint. None shows that AI causes lasting harm to adult thinking. The stronger evidence concerns students learning with answer-giving tools.

Key Terms

• Survey: A study that asks a sample of people the same questions and reports the share giving each answer.
• Correlation: A pattern in which two things tend to go together, which doesn't show that one causes the other.
• Deskilling: The loss of a skill through lack of use after a tool takes over the task.
• Observational study: A study that records what happens without assigning people to conditions, so it can't rule out other causes for what it finds.
• Preprint: A research paper posted publicly before independent experts have reviewed it for a journal.

References

• Budzyń, Krzysztof, Marcin Romańczyk, Diana Kitala, and others. 2025. "Endoscopist Deskilling Risk After Exposure to Artificial Intelligence in Colonoscopy: A Multicentre, Observational Study." The Lancet Gastroenterology & Hepatology 10 (10): 896–903.
  • Free: University of Warwick repository (accepted manuscript)
  • Publisher: The Lancet Gastroenterology & Hepatology
• Kosmyna, Nataliya, Eugene Hauptmann, Ye Tong Yuan, and 5 others. 2025. "Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task." arXiv:2506.08872. Preprint.
  • Free: arXiv 2506.08872
• Lee, Hao-Ping (Hank), Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. 2025. "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers." In CHI '25: Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. New York: ACM. doi:10.1145/3706598.3713778.
  • Free: Microsoft Research (PDF)
• Science Media Centre. 2025. "Expert Reaction to Observational Study Looking at Detection Rate of Precancerous Growths in Colonoscopies by Health Professionals Who Perform Them Before and After the Routine Introduction of AI." August 2025.
  • Free: Science Media Centre
• Stankovic, Hirche, Kollatzsch, and Doetsch. 2025. "Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Tasks." arXiv:2601.00856. Preprint.
  • Free: arXiv 2601.00856

Introduction

Advice on AI and skills tends to arrive as a warning with no instructions attached. "Don't become dependent" doesn't say what to do on a Tuesday afternoon with a report due.

This reading describes a small set of habits that keep a skill in use while you still get help from AI, where each habit comes from, and how far the evidence for them goes.

Decide Which Skills to Keep

The first step is a decision. People have always moved mental work onto tools, from shopping lists to calculators. Psychologists call this cognitive offloading and treat it as an ordinary and widespread behavior (Risko and Gilbert 2016). Handing a task to a tool costs you something only if you still need to be able to do that task.

So sort the work you give to AI into two groups. Some skills you need to keep, because your job depends on them, because you have to judge the tool's output, or because you're still learning them. The rest you can offload without guilt. Reformatting a table or tidying meeting notes may be skills you're content to lose.

The sorting is yours to do. The habits below are for the skills in the first group.

Before and While You Ask

To attempt first is to do your own thinking on a task before you ask an AI tool for help. Write the outline, sketch the argument, or try the problem, and then bring in the tool.

The reason comes from a randomized experiment with nearly a thousand high school mathematics students. Those given a standard chatbot during practice often asked for the solution and copied it. Their practice scores rose, and on a later exam without AI they scored 17 percent lower than students who had practiced without it (Bastani et al. 2025). The students had skipped the step where the learning happens.

An attempt can be brief and can be poor. Its purpose is to make you produce something from your own head before you see the tool's version.

A hint is a pointer toward the next step that leaves the work of taking it to you. Asking for one is the second habit, and asking for critique of your attempt is a close relative.

This is what the successful tools in the learning studies were built to do. In the mathematics study, a second version of the chatbot was instructed to give hints without giving away the answer, and it largely removed the harm to later exam scores (Bastani et al. 2025). In a separate randomized trial with 194 university physics students, an AI tutor designed to keep students actively working produced more than twice the learning of an active-learning class (Kestin et al. 2025).

Those tools had the restraint built in. A general-purpose assistant doesn't, so the restraint has to come from how you ask, with requests such as "What's wrong with my reasoning in step two?" or "Give me a hint and hold back the answer."

After and Away from the Tool

To explain back is to restate a result in your own words, without looking, to check that you understand it. After the tool helps, close it or look away, and say or write what the answer is and why.

If you can't, you've found out something useful: you have the answer and you don't yet have the understanding.

Unaided practice is doing a task with no AI help, on a regular schedule, to keep a skill in use. It's for the skills you must be able to perform alone, such as writing a clear paragraph, reading a set of figures, or giving feedback on the spot.

The schedule can be light. One piece of writing a week with no assistant, or the first analysis of each month done by hand, keeps the skill exercised. It also tells you whether the skill is still there, which is hard to know while a tool is always present.

How Far the Evidence Goes

These habits are inferences from research. Attempting first and asking for hints follow from what the two learning studies found about answer-giving and guided tools. All four are consistent with the long-standing view that skills are kept through use.

Nobody has tested the four together as a package, and the studies behind them involved students in mathematics and physics over short periods. Whether the same habits protect the skills of experienced professionals over years hasn't been measured.

Conclusion

Keeping a skill while using AI starts with deciding which skills matter to you. For those, four habits follow from the research: attempt first, ask for hints or critique, explain back, and practice unaided. They rest on studies of students and haven't been tested together.

Key Terms

• Attempt first: To do your own thinking on a task before you ask an AI tool for help.
• Hint: A pointer toward the next step that leaves the work of taking it to you.
• Explain back: To restate a result in your own words, without looking, to check that you understand it.
• Unaided practice: Doing a task with no AI help, on a regular schedule, to keep a skill in use.

References

• Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. "Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics." Proceedings of the National Academy of Sciences 122 (26): e2422633122.
  • Free: PubMed Central
  • Publisher: PNAS
• Kestin, Greg, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. 2025. "AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting." Scientific Reports 15: 17458.
  • Free: Scientific Reports
• Risko, Evan F., and Sam J. Gilbert. 2016. "Cognitive Offloading." Trends in Cognitive Sciences 20 (9): 676–688.
  • Free: UCL Discovery (accepted manuscript)
  • Publisher: Trends in Cognitive Sciences

Introduction

Most people who use AI at work couldn't say, at the end of a week, which of their skills got exercised and which got skipped. A short audit makes that visible.

This walkthrough follows one person through an audit of a week's AI use. The analyst, her employer, and her log are invented for the example. Her judgments about which skills to keep are hers, and someone in a different job could reasonably decide otherwise. The walkthrough shows the reasoning and doesn't prescribe the answer.

The Starting Point

Amara is in her second year as an analyst at a market research firm. She uses the AI assistant her employer provides, many times a day. Her manager has started asking her to present findings to clients, and she has noticed that she struggles to explain a chart when someone asks a question she didn't prepare for.

She keeps a simple log for one week. Each time she uses the assistant, she writes one line. By Friday she has eight entries.

1. Summarized a forty-page industry report
2. Wrote a spreadsheet formula to combine two tables
3. Drafted the weekly status email to her team
4. Wrote the "what this shows" paragraph for a chart of quarterly sales
5. Turned rough meeting notes into a list of action items
6. Suggested a structure for a client presentation
7. Checked a memo for grammar
8. Explained a statistical term she had forgotten

Walking Through the Audit

Step 1: List each use and the skill it touched

Amara writes beside each entry the skill she would have used if she'd done the task herself.

The summary would have taken careful reading and picking out what matters. The formula takes knowing spreadsheet functions. The status email is routine writing. The chart paragraph takes interpreting data, which is the core of her job. The action items are formatting. The presentation structure is building an argument. The grammar check is proofreading. The statistical term is background knowledge of her field.

Naming the skill takes some care. "Summarize a report" holds two skills: reading critically and deciding what's important.

Step 2: Mark each skill as one to keep or one to let go

She asks one question of each skill: would it matter if I could no longer do this without help?

People have always moved mental work onto tools. A 2016 review by two psychologists calls the practice cognitive offloading and asks what its consequences are for the mind (Risko and Gilbert 2016). Whether a given case matters depends on whether the skill is still needed. So Amara sorts.

She marks four skills keep: critical reading, interpreting data, building an argument, and statistical knowledge. Clients and her manager expect these of her, and she needs them to judge whether the assistant's output is any good.

She marks four let go: formula syntax, routine email, formatting notes, and proofreading. She can test a formula by checking its results. Nobody is relying on her to format notes by hand.

Another analyst might keep formula writing. Amara's reasoning is that she can verify a formula without being able to write it from memory.

Step 3: For the skills to keep, mark whether she attempted first

For the four keep entries, she notes honestly what she did.

• Report summary: straight to AI. She pasted the report and read the summary. She never read the report.
• Chart paragraph: straight to AI. She described the chart and used the paragraph nearly as it came.
• Presentation structure: attempted first. She had sketched her own outline and asked the assistant what was missing.
• Statistical term: asked for an explanation and read it. She didn't test whether she could restate it.

Two of the four went to the assistant with no attempt of her own.

Step 4: Identify the two uses most likely to erode a needed skill

She looks for uses where a skill she needs is being skipped entirely, and often.

The report summary and the chart paragraph stand out. Both involve core skills, both happen every week, and in both the assistant did all of the thinking. The chart paragraph also matches the trouble she's been having with client questions.

Research on students gives her a reason to take this seriously. In a randomized experiment with nearly a thousand high school mathematics students, those who had a standard chatbot during practice did better on the practice problems and then scored 17 percent lower on a later exam without it. A version of the tool that gave hints without the answer largely avoided that harm (Bastani et al. 2025). Amara isn't a high school student and her work isn't mathematics, so the study doesn't prove anything about her. It does describe her pattern: good results with the tool, and difficulty without it.

The presentation structure worries her less, since she attempted first.

Step 5: Rewrite those two uses with attempt-first and explain-back

She doesn't stop using the assistant for either task. She changes the order.

For reports, she'll read the executive summary and one key section herself and write three bullet points on what she thinks matters. Then she'll ask the assistant for its summary and compare the two. Where they differ, she'll look at the report to see who was right.

For charts, she'll write the "what this shows" paragraph herself first. Then she'll ask the assistant to critique it and to name anything in the data she missed. Before sending, she'll close the assistant and explain the chart out loud in two sentences, as she'd have to in front of a client.

Both revisions take a few minutes more than before. She decides the time is worth it for these two tasks.

Key Considerations

The common mistake is to treat all AI use as equally risky or equally safe. Someone who sees it all as risky gives up the convenience of the four let go tasks for nothing. Someone who sees it all as safe never notices that two core skills have dropped out of the week. The audit's value is in telling the two apart.

Which skills to keep is a judgment, and it changes. If Amara moves into management, interpreting charts herself may matter less and writing clear email may matter more.

The audit also doesn't measure whether her skills have declined. A week's log shows where practice is and isn't happening. To find out whether a skill is still there, she'd have to do the task without help.

Summary

Amara listed her uses, named the skill behind each, sorted the skills, checked whether she had attempted first, and changed two uses. Her audited log follows.

Revised use 1, report summary: Read the executive summary and one key section. Write three bullet points of my own. Then ask for the assistant's summary and check the differences against the report.

Revised use 4, chart paragraph: Write the paragraph myself. Ask the assistant to critique it and to name anything I missed. Close the assistant and explain the chart aloud in two sentences before sending.

1. Each revision begins with her own attempt, so the skill is exercised before the tool is involved.
2. Each uses the assistant for comparison or critique, which keeps the help without handing over the task.
3. The second ends with explaining back, the step that matches what clients ask of her.

References

• Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman. 2025. "Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics." Proceedings of the National Academy of Sciences 122 (26): e2422633122.
  • Free: PubMed Central
  • Publisher: PNAS
• Risko, Evan F., and Sam J. Gilbert. 2016. "Cognitive Offloading." Trends in Cognitive Sciences 20 (9): 676–688.
  • Free: UCL Discovery (accepted manuscript)
  • Publisher: Trends in Cognitive Sciences