We Are Teaching Machines to Think. Are We Forgetting How?

There is something mildly absurd about the fact that we now have machines capable of explaining Kant, summarizing Nietzsche, solving fairly complicated mathematical problems, and writing surprisingly convincing arguments about consciousness, yet we are increasingly using them to avoid thinking about relatively simple things.

Ask someone to draft an email, and they open ChatGPT.

Ask them to interpret a paragraph, and they open ChatGPT.

Ask them what they think about a particular issue, and there is a reasonable possibility they will first ask ChatGPT what they should think.

I am not entirely sure this is progress.

Now, before anyone concludes that I have suddenly become anti-AI, let me clarify something. I work in this industry. I believe artificial intelligence is among the most consequential technologies we have developed. I am fascinated by what it can do, particularly when it extends human capability beyond its biological limitations.

But extending capability and replacing cognitive effort are two very different things.

And I think we are beginning to confuse them.

The strange comfort of not having to decide

Human beings have always looked for ways to reduce cognitive effort.

We invented writing because remembering everything was exhausting. Calculators because arithmetic is repetitive. Search engines because finding information manually is inefficient.

None of these inventions destroyed human intelligence.

But generative AI is different in one important respect.

It does not simply retrieve information or perform calculations. It can produce an interpretation, construct an argument, recommend a decision, and explain why that decision supposedly makes sense.

In other words, it can occupy the space where human judgment normally operates.

That space matters.

There is an old philosophical question about whether a person who consistently performs good actions without understanding why they are good can really be considered morally wise.

The distinction is subtle but important.

Doing the right thing and understanding the right thing are not identical achievements.

Something similar is happening with AI.

A person can now produce an excellent strategic proposal without understanding the assumptions behind it. A student can submit a sophisticated essay without having developed the argument. A manager can approve an AI-generated recommendation without recognizing the trade-offs it conceals.

The output may be impressive.

The intellectual development that normally accompanies producing that output may never happen.

And this is where things become interesting.

Intelligence is not merely the production of answers

We have become extraordinarily good at measuring intelligence through outputs.

Correct answers. Completed tasks. Benchmarks. Productivity gains.

But intelligence also involves the ability to recognize when a question has been framed incorrectly.

It involves tolerating uncertainty long enough to discover something unexpected.

Sometimes it involves sitting with an uncomfortable problem for three days, only to realize that the assumption you started with was nonsense.

A machine that immediately produces a plausible answer can interrupt that process.

Not because the machine is malicious.

Because convenience is seductive.

Consider GPS navigation.

Most people can travel through unfamiliar cities with extraordinary efficiency today. Yet many barely remember the route afterward.

The technology has improved navigation while reducing the need to construct an internal map.

This is not necessarily a problem.

Unless we assume that arriving at a destination and understanding the geography are the same thing.

AI creates a similar distinction between obtaining an answer and developing an understanding.

The real danger is intellectual dependency

I am less concerned about AI becoming more intelligent than humans than I am about humans becoming less willing to exercise their own intelligence.

That is not the same as predicting widespread cognitive decline. We do not yet have enough evidence to make such sweeping claims.

But the dependency mechanism deserves attention.

When a system repeatedly offers coherent explanations, confident recommendations, and conveniently organized conclusions, it becomes tempting to treat plausibility as authority.

Eventually, the question changes.

Instead of asking, "Does this argument make sense?"

We begin asking, "Does the AI agree?"

That is an extraordinary transfer of intellectual responsibility.

And it can happen without any formal decision to surrender it.

Just through habit.

So what should we actually do?

I do not believe the answer is to use less AI indiscriminately.

That would be a remarkably unimaginative response to an incredibly powerful technology.

Instead, we should become more deliberate about the cognitive work we delegate.

Use AI to challenge your reasoning, not merely confirm it.

Ask it for counterarguments rather than only conclusions.

Let it expose assumptions you missed.

Use it to investigate ideas you could not explore efficiently on your own.

But occasionally, formulate your own position before asking the machine.

The difference may sound trivial.

I suspect it is not.

Because the future of intelligence should not be a competition in which humans gradually stop thinking while machines get better at producing answers.

It should be a collaboration in which machines increase what human beings can understand, question, and ultimately decide.

We are investing enormous resources into making artificial intelligence more capable.

Perhaps we should devote a little more attention to keeping human intelligence intellectually active.

After all, the point of building a more intelligent tool was never to make its operator less thoughtful.