AI can finish the work without building the skill. That is the capacity question hiding underneath nearly every argument about whether artificial intelligence is making us more productive.

For AI and learning, I think we need to separate two things that are being casually merged: a stronger deliverable today and a more capable person tomorrow. They can overlap. They do not automatically overlap.

A polished proposal, email, lesson plan, or analysis may be exactly what a busy person needs. But if the tool generated the reasoning, organized the evidence, and supplied the language, what changed in the person’s actual life? Can they explain the decision in a meeting, spot the weak premise, or adapt when the situation becomes unfamiliar?

This is not an argument for making every task harder than it needs to be. It is an argument for measuring the right outcome.

The deliverable is not the person

There is real evidence that generative AI can improve immediate work output. In a preregistered experiment involving 444 college-educated professionals completing short, occupation-specific writing tasks, participants given ChatGPT finished faster and received higher quality ratings from blinded evaluators. The authors also found that the gains appeared to come largely from the tool substituting for worker effort rather than complementing worker skill.

That is not a flaw in the study. It is the finding.

If the job is to produce a competent first draft quickly, substitution may be the whole point. A tool that removes a tedious bottleneck can preserve time, reduce overload, and make certain work more accessible. In a real life already crowded by caregiving, illness, poor sleep, or excessive administrative work, that matters.

But the study examined short, self-contained tasks and immediate output. It did not test whether participants became better writers, more discerning editors, or more capable thinkers without the tool later. The authors are direct about those limits. Productivity is a meaningful outcome. It is simply not the same outcome as durable capacity.

AI and learning are different jobs

Learning is not the pleasant feeling of having reached a correct answer. Often, it is the less glamorous work of noticing what you do not understand, attempting an explanation, finding the gap, and trying again. AI can participate in that process. It can also eliminate it.

That distinction is less about the model than the role we hand it. Does it provide the finished answer before we have formed a view? Does it challenge a conclusion, offer counterexamples, or help us see an error after we have done some independent thinking? Those are radically different cognitive arrangements, even when the screen looks identical.

A small MIT Media Lab study of essay writing makes this tension visible, but it should not be oversold. In the study, 54 participants were assigned to use an LLM, a search engine, or no external tool across several sessions. The LLM group showed lower ability to quote from their own essays shortly after writing them and reported less ownership. The researchers themselves emphasize that the work is a non-peer-reviewed, context-specific preliminary study, with a small sample and findings that cannot be generalized to every task, person, or AI system.

So no, this does not demonstrate that AI is damaging brains. That claim would outrun the evidence. It does raise a more useful question: when a tool performs the generative work, what parts of the task remain encoded as our own understanding?

Friction is not always waste

We have been trained to treat friction as an operational failure. In many cases it is. Filling out the same form five times is not character-building. Searching a tangled spreadsheet for a basic answer is not a cognitive workout worth protecting.

But some friction is the activity that creates judgment. Drafting a rough explanation exposes whether you understand the idea. Comparing competing interpretations teaches calibration. Wrestling with an ambiguous problem teaches you when a clean answer is suspicious.

The trade-off becomes sharper in knowledge work. If AI turns a two-hour task into a twenty-minute task, the obvious benefit is reclaimed time. The less obvious question is what happens to that time. Does it become recovery, relationship, deeper review, and more meaningful work? Or does the organization quietly convert the gain into six more assignments?

Brain health is not served by pointless strain. It is also not served by a work system that removes every moment of independent judgment, then raises the volume until people are just supervising machine-generated motion. That is not freedom from cognitive load. It may be a transfer from doing to monitoring, with less ownership and more output to defend.

The new performance metric is transfer

I am less interested in whether AI makes someone look capable on a task than in whether capability transfers when the conditions change. Can the person use the result in a live conversation? Can they recognize an invented citation or an irrelevant recommendation? Can they make a sound decision when the prompt is incomplete, the stakes are higher, and the tool’s confidence exceeds its accuracy?

Those are not anti-AI questions. They are questions about decision quality. They fit naturally with a broader view of mental performance: attention, memory, recovery, judgment, and meaningful functioning rather than raw speed alone. They also belong in any serious conversation about brain health, because the brain we are trying to support is not a content-production engine. It is a living system that has to orient, choose, connect, and adapt.

Convenience can be a benefit and a blind spot

The false choice is between refusing AI and outsourcing everything. The more interesting distinction is between using AI to extend a capability and using it to conceal that a capability was never built.

A completed task tells us what the system produced. Transfer tells us what changed in the person. As AI gets better at producing the first, which one will we be willing to measure?

Want to think through your own capacity questions? See how to work with Dr. Teralyn.