Who Will Educate Our Instincts?

AI can do more of our work. Are we still learning how to judge it?

Malcolm Gladwell made a passing comment on Bill Maher’s show that stayed with me longer than the rest of the conversation.

Asked whether we should trust our gut in an age of misinformation, he drew a distinction: informed intuition is powerful; uninformed intuition is dangerous. Then he asked the question that caught my attention:

“Are we doing a better or a worse job of educating our instincts?”

Gladwell sounded optimistic. AI, he suggested, puts credible information at our fingertips, potentially helping us see the world more clearly.

I share that optimism. But I’m less certain about the connection between having information available and developing the judgment to use it.

AI can give us an answer. It can increasingly do the work. But does using it make us better at knowing whether the answer—or the work—is any good?

The programmer who no longer writes much code

I had been thinking about something similar while listening to David Heinemeier Hansson, the creator of Ruby on Rails, discuss his enthusiasm for AI coding.

DHH has moved toward an agent-first workflow in which he barely writes code by hand. In a recent interview, he described how the tools have changed his process without changing his standards for the software he ships. Source

Good for him. There is nothing inherently virtuous about typing every line yourself.

But there is an important distinction between an experienced programmer deciding not to write code and a new programmer never learning much about how code works.

DHH brings years of experience to the conversation with an agent. He can question an architecture, reject an unnecessary abstraction, or sense that a proposed solution is more complicated than the problem warrants.

Someone starting today may be able to generate the same application. They may not yet know which questions to ask about it.

This is not an argument against beginners using AI. It is a question about how beginners become experts.

If we automate the activities through which people traditionally developed judgment, what replaces those activities?

Instinct is learned, not downloaded

“Trust your gut” can sound like permission to stop thinking. But good intuition is not simply a strong feeling.

Kahneman and Gary Klein, approaching expertise from different perspectives, identified important conditions for trustworthy intuitive judgment: an environment with learnable regularities and an opportunity to learn them. Confidence alone does not establish accuracy. Source

In practice, that learning might look like this:

You think something will work. You try it. Something unexpected happens. You investigate. The next time, you notice a warning sign earlier.

Eventually, what once required deliberate analysis can feel immediate. You cannot always explain your concern straight away, but experience tells you where to look.

The question is what happens when an AI agent intervenes throughout that process.

If it proposes the solution before I form an opinion, fixes the bug before I investigate it, and supplies the explanation before I struggle to understand it, I may finish sooner.

But what have I learned?

Perhaps a great deal—if I question the explanation, test the solution, and examine the failure.

Perhaps very little—if I accept the result and move on.

The distinction is not between using AI and refusing it. It is between two ways of using the same tool.

An apprenticeship accelerator—or an apprenticeship bypass?

There is a reasonable objection here: why assume the old way of learning was the best way?

Much of it was slow, repetitive, and unnecessarily frustrating. Spending an afternoon on a configuration error does not automatically make someone a better engineer. An AI tutor that explains the problem in five minutes might be a substantial improvement.

There is evidence that AI can help less-experienced people catch up. A study of more than 5,000 customer-support agents found that AI assistance produced its largest productivity gains among novice and lower-skilled workers, with suggestive evidence that the system helped transmit effective practices. Source

That matters. AI does not only amplify expertise people already possess. It can also make expertise more accessible.

But improved performance and improved understanding are not necessarily the same thing.

In a small randomized study of developers learning an unfamiliar Python library, AI assistance reduced immediate comprehension scores on average. Participants who used AI to ask conceptual questions and seek explanations tended to retain more understanding than those who delegated the work. Those usage-pattern findings were observational, and the study did not establish what happens to expertise over years. Still, it illustrates the distinction. Source

AI can accelerate an apprenticeship. It can also let us bypass one.

A completed task does not tell us which has happened.

What if AI supplies both fast and slow thinking?

It would be comforting to divide the work neatly: humans provide intuition; machines provide analysis.

I doubt that division will hold.

The System 1 and System 2 distinction popularized by Thinking, Fast and Slow is useful shorthand for fast, automatic judgments and slower, deliberate reasoning. But it is not a permanent allocation of responsibilities between people and computers.

AI systems are being developed for both kinds of work.

TypeSafe’s Jev, for example, is presented as a “System One” model for fast, structured decisions: classification, routing, scoring, and similar tasks. The company’s performance claims need scrutiny, but the ambition is relevant here. AI is not being built only to deliberate at length. It is also being built to decide quickly and cheaply. Source

So perhaps humans will not always be needed to make the initial judgment, select the tool, or inspect every result. Some of those decisions may become reliably automated too.

I do not think the argument for educating our instincts should depend on insisting otherwise.

The practical question is what we can responsibly delegate—and on what evidence.

A system being fast does not make it right. A system being elaborate does not make it right either. And neither does a human having a strong feeling about its answer.

AI is not quite a hammer

I sometimes think about AI becoming an ordinary tool.

We pick up a hammer when we need to drive a nail. We use a calculator when we need to do arithmetic. Perhaps we will learn to reach for different AI systems just as naturally.

But the analogy has limits.

A hammer does not usually suggest what I should build. A calculator does not persuade me that I have chosen the right business strategy.

AI can help frame the problem, propose the solution, execute it, and explain why the result is good. That is useful—and it makes independent judgment harder to locate.

The skill we need is therefore more than knowing which tool to pick up.

It includes asking:

• Am I solving the right problem?
• What would convince me that this result is correct?
• How costly would a mistake be?
• Am I delegating something I need to understand?
• Is this getting me closer to an outcome I actually want?

Those are not all questions of instinct. Some deserve slow, explicit thought. Part of good judgment is recognizing when the quick answer is not enough.

Learning while delegating

I do not want to preserve difficulty for its own sake. Nor do I think young people should have to repeat every frustrating experience their predecessors endured.

But I would like us to distinguish between friction that wastes time and friction that teaches us something.

When learning a new domain, that might mean forming a prediction before asking AI for an answer. Asking what would make its recommendation fail. Diagnosing a problem before accepting a repair. Occasionally attempting a task without assistance to find out what we actually understand.

When doing familiar, easily checked work, we might delegate much more aggressively.

The balance should depend on whether we are trying to finish a task, develop a skill, or do both.

That distinction matters for individuals. It also matters for organizations. If experienced people can now produce more while juniors encounter fewer opportunities to learn, we need to design new ways for that learning to happen.

We cannot simply celebrate the productivity gain and assume the apprenticeship will take care of itself.

The question that remains

Gladwell’s question is worth keeping close precisely because it does not require us to be pessimistic about AI.

These tools may help us learn faster, explore further, and accomplish things we could not previously attempt. I want those benefits.

But I also want to know what happens to the person using them.

After a year of working with AI, am I better at recognizing a weak argument, an unnecessary complication, or an answer that needs checking?

Or am I mainly better at obtaining a plausible result?

We do not need to do all the work ourselves. But we should care whether the work we delegate leaves us more capable of judging what comes back.

AI can help us finish the task. We still need to ask whether it is helping us educate our instincts.