AI and the End of Productive Friction

Every math teacher on the planet, and nearly every parent, will, at some point in their life, be asked some variation of the following questions: “Why do I have to learn math? When will I ever use this?” It is a good question. Most adults have no practical reason to perform long division by hand. Calculators are faster, more accurate, and always nearby. Forcing an accountant to manually divide large numbers before approving a financial statement would not make the work more rigorous. It would make the accountant slower.

And yet, we still teach children division. The reason points toward a problem we are about to encounter far beyond arithmetic. Learning a process and performing that process are not the same thing. Long division may be an inefficient way to calculate a quotient, but learning division helps build an understanding of magnitude, proportion, place value, estimation, and the relationships among numbers. Once that understanding exists, the calculator can take over much of the execution.

A skill can become unnecessary to perform long before it becomes unnecessary to learn.

Generative AI is forcing us to decide how far that principle extends. A calculator automates a narrow operation. AI can write the report, construct the argument, generate the code, summarize the evidence, propose the diagnosis, and explain why its own answer makes sense. The immediate result can be better work produced in less time. That is an extraordinary benefit.

It also creates a problem we have not had to confront at this scale before: The quality of our output can improve faster than our own capability.

For most of history, intellectual work did two jobs at once. It produced something useful, and it developed the person doing it. Writing the analysis also taught the analyst how to structure an argument. Debugging the program taught the programmer how software usually fails. Working through thousands of diagnoses helped the physician recognize when a familiar pattern did not quite fit. AI gives us the ability to separate those outcomes. And before we universally celebrate that efficiency, we should understand what it might be costing us.

The Work Was Teaching Us

Educational psychology has been demonstrating some version of this distinction for decades. Roediger and Karpicke demonstrated this with a counterintuitive finding about memory. In their study, students who repeatedly reviewed material looked stronger immediately afterward. Students forced to retrieve the material from memory struggled more in the moment but remembered substantially more of it days later. The experience that produced better short-term performance was not the one that produced better learning.

The easier learning experience created stronger immediate performance and a stronger feeling of competence. The more demanding one produced more durable knowledge.

Manu Kapur found a similar pattern in mathematics. Students allowed to struggle with unfamiliar problems before being taught how to solve them often performed poorly at first. But after receiving instruction, they developed stronger conceptual understanding and were better able to apply what they had learned to new problems than students who were given the correct method from the beginning.

A broader body of learning research points in the same direction. From Vygotsky's work on learning just beyond one's current independent ability to Anders Ericsson's research on deliberate practice, improvement tends to happen when people are pushed beyond what they can already do, given feedback or support, and forced to adapt. Expertise develops not simply by accumulating information but by repeatedly confronting problems that expose the limits of current understanding.¹

None of this means difficulty is inherently valuable. Plenty of friction is pointless. Filling out redundant forms does not build character. Waiting three days for information that could arrive instantly does not improve judgment. Performing arithmetic by hand every time you need an answer would be absurd.

Productive friction is effort that develops, calibrates, or maintains a capability we still need.

The difficulty matters only because of the abilities it helps develop. That makes the question surrounding AI much harder than whether a machine can perform a task faster than we can. Of course it can. The real question is whether performing that task ourselves was building something we will still need after the machine takes over.

Borrowed Competence

A 2025 study involving nearly a thousand high school mathematics students illustrates the problem remarkably well. Researchers gave students access to GPT-4 while they worked on practice problems. One version behaved much like an ordinary chatbot. Another was deliberately designed to protect learning by providing structured guidance rather than making answers too easy to obtain.

While the AI was available, both groups improved dramatically. Students using the ordinary GPT interface improved their practice performance by 48 percent relative to students without any AI. Those using the safeguarded tutor improved by 127 percent.

Then the AI was removed. Students who had used the ordinary GPT system performed 17 percent worse on the subsequent unassisted exam than students who had never received AI access. The more carefully designed tutor largely avoided that penalty.²

The students using unrestricted AI had produced better work while developing less capacity to produce it independently. I think of this as borrowed competence.

There is nothing inherently wrong with borrowed competence. Modern life depends on it. I cannot manufacture the processor inside my computer, repair my car’s engine, or reproduce the infrastructure that delivers electricity to my house. Human progress depends on specialization and tools.

The problem begins when you are responsible for the quality of an output whose underlying process you do not understand. A calculator usually does not create much confusion about where its competence ends and yours begins. You decide what to calculate. It returns the result.

Generative AI can blur that boundary. It can choose an approach, carry it out, explain the result, defend the explanation, and present the whole thing in polished language. The better it gets, the easier it becomes to confuse the capability of the combined human-machine system with capability that actually resides in the human.

Borrowed competence works remarkably well…until it doesn’t.

The Programmer Who Cannot Read the Code

Programming makes the problem easy to see. A novice programmer can now describe what a program should do, receive code in seconds, run it, copy an error message back into the model, receive a correction, and repeat the process until the software works.

Depending on the task, that may be an excellent way to build software. It may also produce someone who can create functioning software without understanding very much about why it functions.

The declining value of manual code generation is not itself concerning. Programmers already rely on libraries, frameworks, autocomplete, documentation, compilers, and countless abstractions. There is no particular virtue in memorizing syntax that a machine can supply instantly.

But someone still needs to know whether generated code is secure, whether it handles unusual conditions, whether the architecture will survive expansion, whether a fix creates another problem somewhere else, and whether an apparently functional solution is doing what everyone thinks it is doing. A 2025 study of undergraduate programming students focused on exactly this kind of task. In one part of the research, students generally viewed generative AI as useful for coursework. In a separate part, 147 students were asked to find and correct errors embedded in code produced by a large language model. They had significantly more difficulty with those debugging tasks than with conventional instructor-designed programming problems.³

That does not show that AI use caused weaker programming ability. The study was not designed to establish that. What it does show is that evaluating generated code poses its own demands, and those demands require substantive underlying knowledge.

As writing code becomes easier, understanding code may become more important. If AI writes most routine functions, debugging may become one of the clearest tests of whether a programmer actually understands the system. Future programmers may rarely need to write every function from scratch. But if they skip the years of writing, breaking, and fixing code themselves, they may reach the point where they are expected to supervise systems whose logic they never fully learned. Medicine raises the stakes further.

What Does the Doctor Still Need to Know?

Imagine a physician working with an AI system that can review a patient's symptoms, laboratory values, imaging, medications, history, and prior records, then generate a differential diagnosis and recommend treatment. As these systems become more capable, an obvious question follows: How much underlying biology does the physician still need to know?

Medical education requires enormous amounts of foundational science. Anatomy. Physiology. Pathology. Pharmacology. Microbiology. Much of that knowledge is rarely recited explicitly during an ordinary clinical encounter.

If an AI system can retrieve it instantly and make increasingly accurate recommendations, reducing the amount physicians must learn will become tempting, especially in communities where access to quality healthcare is limited. Some of it probably should be reduced. But there is a difference between knowledge we no longer need to retrieve from memory and knowledge that has become part of how we understand a problem.

Nicole Woods, Lee Brooks, Geoffrey Norman, and colleagues have studied this question in clinical reasoning. In one experiment, students learned medical conditions either through underlying basic-science mechanisms or through direct associations between symptoms and diagnoses. Immediately after training, the two approaches produced similar diagnostic performance. One week later, the group taught the causal mechanisms retained its performance while the comparison group's accuracy declined. Other experiments found that learners with knowledge of underlying mechanisms performed better when cases included unfamiliar terminology or irrelevant information.⁴

The value of the biology was not that physicians needed to consciously reconstruct every diagnosis from first principles. The causal knowledge helped organize what they were seeing and gave them something to fall back on when the usual pattern became unreliable.

That sounds a lot like long division. An experienced physician no more needs to reason from cell biology through every routine diagnosis than I need to perform long division every time I calculate a percentage. The underlying knowledge can still matter when something does not fit.

Now add AI. In a multicenter study published in JAMA, 457 clinicians evaluated clinical cases with different forms of AI assistance. Helpful AI improved their diagnostic accuracy modestly. When researchers systematically biased the AI recommendations, clinician accuracy fell by 11.3 percentage points. Providing explanations for the incorrect recommendations did not significantly eliminate the effect.⁵

That does not prove that more knowledge of biology would have prevented the errors. It does establish something more limited and still important: Keeping a human in the loop does not automatically create an effective safeguard against a convincing machine.

If the physician's role increasingly becomes supervising machine-generated diagnoses, what exactly must the physician know in order to disagree? That question will eventually matter in almost every knowledge profession.

There Is Nothing Noble About Doing Things the Hard Way

There is an obvious counterargument to everything I have written so far. Removing cognitive work is precisely what tools are for.

Writing allows us to stop memorizing everything. Search engines allow us to retrieve information without knowing where it is stored. Spreadsheets perform calculations that once required enormous amounts of manual labor. GPS prevents us from learning every road in an unfamiliar city. None of this represents civilizational decline.

AI is already producing real gains. In an experiment involving 453 college-educated professionals performing writing tasks, Shakked Noy and Whitney Zhang found that ChatGPT reduced the average time required by 40 percent while increasing evaluated output quality by 18 percent.

In another study, researchers examined the introduction of a generative AI assistant to 5,172 customer-service agents. Productivity increased by 15 percent on average, with particularly large gains among less experienced and lower-skilled workers. The evidence also suggested that AI was helping disseminate some of the practices used by stronger employees.⁶

That matters because it complicates the easy story. AI does not necessarily prevent people from learning. Sometimes it helps them learn faster. The high-school mathematics study makes the same point. AI designed as a tutor largely avoided the learning penalty created by unrestricted answer generation.

So the answer cannot be that we should preserve friction. We should remove as much useless friction as possible. The difficulty is figuring out which friction was useless.

The Expertise Ladder

There is another consequence that may ultimately matter more than whether any individual becomes too dependent on AI. A great deal of entry-level professional work serves two purposes. Junior programmers fix simple bugs before they design complicated systems. Young physicians see routine patients before they manage strange cases independently. Analysts gather data, reconcile numbers, construct models, research details, and write drafts that more senior people revise.

From the organization's perspective, this can be inefficient. Senior employees often could complete the immediate task faster and better. But the organization has historically received two outputs from the junior employee's work. It gets today's spreadsheet, memo, diagnosis, or code. It also gets a slightly more experienced employee. A beginner becomes useful partly by doing work that an expert no longer needs to do.

AI changes that bargain. If a machine can produce the first draft, clean the data, write the routine code, summarize the evidence, or analyze the straightforward case faster and cheaper, removing that work makes economic sense. What if that work was also the apprenticeship?

We do not yet have evidence showing that AI will hollow out the expertise pipeline across professions, and it would be premature to pretend otherwise. But the possibility follows directly from the way expertise has traditionally developed.

How do we produce senior judgment after automating junior experience?

That may become one of the more difficult organizational problems created by AI. For generations, much apprenticeship has been hidden inside production. People accumulated thousands of small experiences while doing economically useful work. If machines absorb those tasks, organizations may need to recreate the developmental function deliberately.

A software company might have little economic reason for a junior programmer to write basic code that AI can produce instantly. It may have a very good developmental reason. A hospital may eventually have little operational reason for a trainee to work through a straightforward diagnostic problem without AI assistance. It may still need physicians who have experienced enough cases unaided to recognize when an automated diagnosis is implausible. A consulting firm may discover that AI can produce a competent first analysis in minutes. It still needs some method of turning twenty-five-year-olds into people whose judgment it will trust at forty-five.

The irony is hard to miss. As production becomes more efficient, developing capable people may require deliberately preserving experiences that are no longer efficient production.

Execution, Understanding, Judgment

I find it useful to separate three forms of capability that are often treated as one. Execution is the ability to perform the procedure yourself. Understanding is knowing what the procedure is doing, how its pieces relate, and roughly what a sensible result should look like. Judgment is knowing when the result should be trusted, questioned, modified, or rejected.

AI will make human execution unnecessary in an expanding number of tasks. There is no reason to mourn most of them. Understanding is different. It may remain useful long after manual execution disappears because it allows us to interpret and interrogate what the machine produces. Judgment is harder still.

In research involving Boston Consulting Group employees performing realistic professional tasks, GPT-4 substantially improved speed and performance on tasks that fell within the model's capabilities. But on a task deliberately chosen to fall outside that capability frontier, participants using AI were 19 percentage points less likely to reach the correct solution.⁷

AI was useful often enough to deserve trust. Then it encountered a problem where that trust became a liability. This is probably closer to the world we are entering than either the utopian or dystopian version of AI.

The machines will not always be wrong. That would be easy to manage. They will be right constantly. And because they are right constantly, the human being responsible for recognizing the unusual moment when they are wrong will need judgment that cannot be summoned on demand.

Judgment is usually the residue of experience. You solve enough problems, make enough mistakes, encounter enough exceptions, and receive enough correction that eventually something about a bad answer bothers you before you can completely explain why. We should be careful about assuming that we can remove the experiences and preserve the residue.

What We Still Need to Learn

The calculator eventually forced education to distinguish between calculation and mathematical understanding. We decided, imperfectly but reasonably, that adults did not need to execute every calculation manually. We did not conclude that because calculators exist, children no longer need to understand numbers. Generative AI is asking us to make the same decision across far less settled domains.

What must a programmer understand if AI writes most of the code? What biology must a physician internalize if AI proposes most diagnoses? What mathematical processes should a child struggle through if an AI tutor can provide every answer immediately? Which parts of writing actually teach someone to think? And how much junior work can an organization eliminate before it accidentally eliminates the process through which its senior people used to become good?

There will not be one answer. Different domains will tolerate different levels of borrowed competence. As AI becomes more reliable, those boundaries will move. But reliability alone cannot settle the question. A skill can become unnecessary to perform long before it becomes unnecessary to learn. The challenge of artificial intelligence is therefore larger than deciding which tasks machines should perform for us. We also have to decide which processes humans still need to learn after machines make performing those processes unnecessary. A useful test is simple:

Can someone who never learned the underlying process reliably recognize when AI has done it badly?

If the answer is yes, remove the friction. If the answer is no, then the friction may still be doing important work. The danger is not that work becomes easier. It is that we remove the part of the work that was teaching us how to know when the easy answer is wrong.

Notes

1. Henry L. Roediger III and Jeffrey D. Karpicke, “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention,” Psychological Science 17, no. 3 (2006): 249–55, https://doi.org/10.1111/j.1467-9280.2006.01693.x; Manu Kapur, “Productive Failure in Learning Math,” Cognitive Science 38, no. 5 (2014): 1008–22, https://doi.org/10.1111/cogs.12107. For the broader learning frameworks, see L. S. Vygotsky, Mind in Society (Harvard University Press, 1978); and K. Anders Ericsson et al., “The Role of Deliberate Practice in the Acquisition of Expert Performance,” Psychological Review 100, no. 3 (1993): 363–406, https://doi.org/10.1037/0033-295X.100.3.363. ↑

2. Hamsa Bastani et al., “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics,” Proceedings of the National Academy of Sciences 122, no. 26 (2025): e2422633122, https://doi.org/10.1073/pnas.2422633122. ↑

3. Dan Kohen-Vacs et al., “Integrating Generative AI into Programming Education: Student Perceptions and the Challenge of Correcting AI Errors,” International Journal of Artificial Intelligence in Education 35, no. 5 (2025): 3166–84, https://doi.org/10.1007/s40593-025-00496-4. ↑

4. Nicole N. Woods, Lee R. Brooks, and Geoffrey R. Norman, “The Value of Basic Science in Clinical Diagnosis: Creating Coherence among Signs and Symptoms,” Medical Education 39, no. 1 (2005): 107–12, https://doi.org/10.1111/j.1365-2929.2004.02036.x; and Woods, Brooks, and Norman, “The Role of Biomedical Knowledge in Diagnosis of Difficult Clinical Cases,” Advances in Health Sciences Education 12, no. 4 (2007): 417–26, https://doi.org/10.1007/s10459-006-9054-y. ↑

5. Sarah Jabbour et al., “Measuring the Impact of AI in the Diagnosis of Hospitalized Patients: A Randomized Clinical Vignette Survey Study,” JAMA 330, no. 23 (2023): 2275–84, https://doi.org/10.1001/jama.2023.22295. ↑

6. Shakked Noy and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science 381, no. 6654 (2023): 187–92, https://doi.org/10.1126/science.adh2586; Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics 140, no. 2 (2025): 889–942, https://doi.org/10.1093/qje/qjae044. ↑

7. Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality,” Organization Science 37, no. 2 (2026): 403–23, https://doi.org/10.1287/orsc.2025.21838. ↑