When Intelligence Becomes Cheap, What Should Remain Difficult?

When intelligence becomes cheap, the question is not whether to use AI, but which difficulties are worth keeping.

There is an obvious irony in writing this article with AI. I used AI to read and compare several recent papers, to challenge some of their arguments, and to help organize the thoughts that follow. The process was faster than it would have been otherwise. In some respects, it was probably better.

But it also raises a question I increasingly ask myself when working this way:

Which part of the thinking did I just stop doing?

Once you become accustomed to having an intelligent system available almost continuously, the threshold for asking it becomes very low. A difficult paragraph. A comparison that does not quite resolve. Something you cannot immediately remember. An argument whose structure is still unclear.

Instead of staying with the problem, you can hand it over. That is enormously useful.

It may also be one of the most important problems education now has to solve. Not whether students should use AI. That question is rapidly becoming obsolete.

The harder question is: Which cognitive tasks should we happily delegate to AI, and which ones do we need to keep doing ourselves precisely because doing them is how we develop the ability to think?

When output stops proving ability

Economist Jason Potts approaches the problem from an unusual direction in his recent paper The University after AI. His argument is that universities have never primarily sold access to information. Printing, radio, computers, search engines and online courses all reduced the cost of distributing knowledge, yet the university survived.

What universities really provide, in his economic model, is matching and verification: bringing students, academics, employers and institutions together, and credibly certifying that somebody possesses a certain capability.

AI affects that mechanism differently from earlier technologies. Consider an essay. Historically, a strong essay told us something about the person who produced it because writing one required a mixture of knowledge, reasoning, synthesis, language and effort. The same was true, to different degrees, of code, analyses, problem sets, presentations and research papers.

Generative AI weakens that connection. If people with very different levels of underlying capability can produce increasingly similar outputs, then the quality of the output becomes a less reliable indicator of the capability of the person behind it.

Potts describes this as signal collapse. Crucially, it is not simply a problem of cheating. The relationship between assessed output and underlying competence can weaken even when AI use is entirely permitted.

That creates an obvious problem for universities: how do you certify someone’s ability when the artifact they submit tells you less and less about what they can actually do?

But there is a second problem that may be deeper.

Better performance is not necessarily better learning

MIT’s recently released report on AI in teaching and learning approaches the same transformation from the educational side. The committee argues that MIT needs to rethink assessment, reinforce hands-on learning and decide explicitly where AI should and should not be part of a course. More fundamentally, it frames the challenge as preserving an education that develops human powers of discovery, problem solving and invention rather than simply optimizing production.

The important distinction is between performance and learning.

AI can make us perform better without necessarily making us better at performing.

We can write a clearer paragraph without becoming clearer writers. We can produce functioning code without understanding all of it. We can summarize an argument without having worked through the argument ourselves. We can reach an answer without experiencing the uncertainty that previously came before it.

MIT educators have described this missing element as cognitive friction: the difficult part of learning in which we struggle with a proof, fail at an experiment, argue with another person, or simply remain stuck long enough to work something out. MIT faculty involved in the report have explicitly warned that immediately turning to AI when students encounter that difficulty can bypass the process through which the underlying skill is acquired.

That leads to a strange conclusion. Education may need to preserve some inefficiency deliberately.

Some friction is productive

We have spent decades using technology to eliminate friction. Usually, that is sensible. If a machine can calculate faster, search better, translate instantly or automate repetitive work, there is little reason to romanticize the slower method.

But education is unusual because some of the friction may be part of the production process: trying and failing, searching memory before searching externally, formulating an argument before seeing a polished version, sitting with ambiguity, explaining something badly and trying again, discovering that the approach you spent twenty minutes pursuing was wrong.

From the perspective of efficiency, these are delays. From the perspective of learning, they may be doing much of the work. And AI changes their economics radically because the cost of avoiding them is approaching zero.

The temptation is therefore understandable. Why struggle for half an hour with a problem if an excellent explanation is one prompt away?

The answer cannot simply be: because struggling is good. Sometimes it is not. Confusion can remain confusion. Repetition can waste time. AI can be an extraordinary tutor, critic, simulator and intellectual partner.

The question is more precise: when is AI removing unnecessary friction, and when is it removing the friction through which capability is formed?

Scaffolding or substitution?

A recent arXiv paper with the deliberately provocative title Large-Language Models as a Cognitive Virus offers a useful distinction. The viral metaphor is a mathematical model of how dependency could spread through a population; it should not be mistaken for empirical proof that widespread cognitive decline is occurring. In fact, the authors explicitly state that the cognitive losses used in part of their model are illustrative assumptions, and that augmentative AI use could leave unaided capability unchanged or even improve it.

The more interesting idea for me is their distinction between scaffolding and substitution. Scaffolding makes a difficult cognitive activity easier while preserving, perhaps even strengthening, the person’s later capacity to perform it independently. Substitution removes the need to perform the cognitive operation at all.

An AI system can do either. It can ask questions that help me work through an argument, or it can write the argument. It can explain why my code fails, or it can replace the code. It can challenge my interpretation of a paper, or it can read the paper and tell me what to think about it.

None of these choices is automatically right or wrong. But they produce different kinds of humans.

The paper therefore arrives at a deceptively simple test:

What remains when the tool is removed?

Two kinds of capability

This distinction may become increasingly important because we now have to think about two different forms of competence. There is the capability of the human-AI system, and there is the capability of the human alone. They are not the same.

With AI, I can already work across areas and at speeds that would previously have required considerably more time or additional expertise. That is genuine augmentation. There is little reason to pretend otherwise.

But it is also perfectly possible for our combined capability to rise while some of our unaided capabilities decline. Those developments are not contradictory.

A person using GPS may navigate more effectively while becoming worse at navigating without it. A photographer using autofocus can reliably make sharper images without becoming better at manually focusing a lens. In many cases, that trade is entirely sensible.

The stakes become different when the outsourced ability is reasoning, judgment, memory, synthesis or the capacity to formulate a problem.

Education cannot optimize only for what a student can produce while connected to a powerful AI system. But neither can it pretend that the AI system does not exist. It may have to cultivate both capabilities deliberately.

The return of difficulty

This is where Potts’s institutional argument and MIT’s educational argument unexpectedly meet.

If written output becomes cheap and therefore less useful as proof of competence, Potts suggests moving assessment toward things that remain difficult to outsource: oral defense, live problem solving, practical demonstrations and other forms in which a person’s reasoning becomes observable again.

At first glance, that sounds like a technological step backwards. A viva is inefficient. It requires an experienced person to sit with another person and determine whether that person really understands something. It is far harder to scale than collecting hundreds of written submissions.

But perhaps that inefficiency takes on new value precisely because everything around it has become cheap.

And there is a subtle but important overlap here. Live discussion, studio critique, oral defense and practical work do not merely verify capability. They can also help produce it.

The same friction can therefore serve two functions: it makes competence harder to fake, and it creates situations in which competence is more likely to develop.

That strikes me as a much more interesting response to AI than trying to build ever more sophisticated systems for detecting whether somebody used it.

The wrong question

The debate is still often framed as a single question: should students use AI? I think that is already the wrong question. AI will be part of the environments in which these students eventually work. Refusing to teach them how to use it intelligently would make little sense.

The better question is: what are we trying to develop in the person, and does AI help develop it, or does it perform that part for them?

The answer will change depending on the task. A student who has never learned to construct an argument may need to struggle through one. An experienced researcher using AI to compare fifty papers is doing something quite different. A beginner asking an AI to produce a finished solution may be outsourcing the exact operation they need to learn. An expert using the same system to explore alternatives may be extending abilities they already possess.

The technology is the same. The cognitive relationship is not.

What should remain difficult?

AI is compressing the distance between question and answer, between intention and execution, between idea and representation. That compression is enormously powerful.

But some of the time being removed may not have been wasted time. It was the time in which we searched, doubted, remembered, failed, reconstructed and eventually understood.

If so, the task ahead is not to resist cognitive acceleration. Nor is it to preserve difficulty for its own sake. It is to become much better at distinguishing between friction that merely slows us down and friction that forms us.

Because when intelligence becomes cheap and answers become immediate, perhaps one of the most important educational decisions will be surprisingly simple:

What should we deliberately keep difficult?

Sources

1. Jason Potts. The University after AI. SSRN.
2. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training. Final Report.
3. MIT News. The crucial human component in computing and AI.
4. Ricard Solé et al. Large-Language Models as a Cognitive Virus. arXiv.