What Will Humans Lose? The Cognitive Crisis in the Age of AI

AI is fast. It does not get tired. And now it really can do a lot of language-based work: writing, conversation, translation, summarizing information, writing code, revising code.

There is no point denying that. Denial changes nothing.

What I want to ask is something else: if AI is this powerful, what is that power built on? How far does it actually go? Where is it genuinely strong, and where does it only look strong? If we do not think these questions through, both our optimism and our anxiety about AI will become shallow.

Why AI Became So Strong at Language First

In the current wave of AI, the main line is still the large language model.

That is not surprising. Language is the bridge of human communication, and text is the form of information that modern society can most easily digitize, standardize, store, and circulate. A huge share of human knowledge, rules, experience, methods, and arguments eventually becomes text, images, code, tables, documents, webpages, and papers.

In other words, what AI can most easily access is not the world itself, but the world already written down by humans.

That matters, because it defines both AI’s strengths and its limits.

What Is Reality

When people talk about AI, they often jump straight to the word “truth.” I would rather use a more ordinary word: reality.

Human beings are not born understanding the world. We first run into the world, and only then do we slowly grow an understanding of it.

Before a child has ever seen an ant, it is actually very hard to truly teach that child what an ant is. You can say ants are small, that they crawl, that they carry things, that they often appear in groups. But those are still just sentences. What really starts to give the child an understanding of “ant” is usually a concrete experience: crouching on the ground, seeing a line of ants come out of a crack in the bricks, noticing that they really do move around obstacles, really do carry food, and really are not quite like the other bugs the child has seen before.

The same is true at a higher level. If someone has no educational background and no similar prior experience, it is usually very hard to make them immediately understand black holes, quantum mechanics, or probability distributions. That is not because they are stupid. It is because they do not yet have enough nearby concepts in their head to catch these new ones.

So our understanding of the world does not appear out of nowhere. It is built on sensation, experience, comparison, trial and error, and memory. Light enters the eyes, sound enters the ears, smells are perceived, heat and cold are felt, pain is remembered, action produces feedback, failure brings costs. On that basis, little by little, human beings come to know reality.

What Is a Concept

I think the word “concept” matters a lot.

Concepts do not appear first as dictionary definitions that people later go on to understand. Most of the time, a concept is a stable structure that a person slowly extracts from repeated contact with reality.

As a child, you see many different dogs, and only later realize that the word “dog” gathers them together. You fall, bump into things, get hurt again and again, and only later understand that the word “danger” is not empty. You do work, fail, redo it, deal with consequences, and only later understand what “risk control” really means.

You could also say that a concept is what remains after experience has been compressed.

What Is Language

If concepts are structures extracted from experience, then language is what carries those structures.

The biggest role of language is to compress complex experience into symbols that can be shared.

Take the word “ant.” It is only one word, but it can immediately bring up a lot: size, shape, the way it moves, the feeling of seeing many of them together, even your own memory of watching ants as a child. Language is powerful because it allows people to activate understanding without having to relive the entire experience from scratch.

But language also has limits.

Because once you compress something, you lose information.

You say “pain,” but another person does not know what kind of pain you mean. You say “love,” and the other person may understand something completely different from your experience. You say “I’m under a lot of pressure,” and another person may hear only “busy,” not insomnia, agitation, or fear of making a mistake.

So language itself is not reality. It is an abstraction and compression of reality. To understand a word is not just to remember it, but to reconnect it to experience.

From the perspective of cognitive science, there are good reasons to think this way. The embodied cognition tradition argues that human concepts are not just abstract symbols; they are deeply tied to perception, action, and environment. Barsalou’s work on grounded cognition, Noe’s discussions of perception and action, and Hauk, Johnsrude, and Pulvermuller’s 2004 study on action words and motor cortex activation all point in the same direction: at least for humans, understanding is not just pure symbol manipulation.

Put simply, people do not understand words only because there is a dictionary in the brain. They understand them because the body and experience have actually stored something there.

What Is AI

If you follow the line above, AI’s position becomes easier to see.

Humans first come into contact with reality.

In reality, humans form experience.

Experience is extracted into concepts.

Concepts are compressed into language, images, code, and documents.

And what AI mainly learns from are those symbolic materials that humans have already processed.

So the way I prefer to understand today’s AI is this: it is built mainly on human language and symbol systems, and what it faces first is not reality itself, but human descriptions of reality.

This is not a put-down. On the contrary, it is exactly what explains why AI is so strong.

Modern society already contains a huge amount of work that operates on the symbolic level. Writing emails, writing contracts, writing reports, writing code, looking up information, summarizing information, making copy, doing preliminary analysis: these tasks are already built on language and structural relationships. It is natural that AI performs well there.

The Boundary of AI’s Ability

And that is also where the problem begins.

Because AI learns from descriptions, records, abstractions, and compressed materials, it is especially good at finding relationships among those materials. It is especially good at producing answers that look right, expressions that look clear, and structures that look complete.

But that does not mean it has come to know reality itself in the same way a human being does.

It can write an accurate description of an ant, but it has never actually seen one. It can summarize large amounts of text about love, responsibility, death, and risk, but it has never experienced those words in the first person.

So I think it goes too far to say that AI has already “understood the world.” A more careful way to put it would be this: AI has become very good at learning human descriptions of the world, but that is still not the same thing as directly understanding the world itself.

That distinction matters. If language is already one layer of abstraction away from reality, then AI is, to a large extent, built on a further layer of abstraction and fitting on top of that.

The more layers of abstraction you add, the farther you move from reality, and the easier distortion becomes.

Why Distortion Happens

Distortion is not a problem unique to AI. Humans have always distorted things too.

We read reports instead of seeing the scene. We read meeting notes instead of hearing the tone of the room. We read retrospectives instead of living through the confusion. We read conclusions someone else has written instead of walking through the process of arriving at them ourselves.

Abstraction obviously has value. Without it, humans could hardly cooperate at scale, and knowledge could hardly accumulate.

But abstraction always throws something away.

First it loses detail, then background, then cost, and what remains is a version that is clear, clean, and easy to circulate.

But reality is usually not like that. Reality contains local conditions, unclear context, and costs that are invisible at the time and obvious only afterward.

If a person spends too long dealing only with abstracted results and no longer touching reality itself, a problem slowly appears: they get better and better at talking about problems, and worse and worse at actually handling them.

That may be one of the real cognitive crises of the AI age.

What AI Still Needs Humans to Do

If AI is built on human descriptions of reality, then one thing follows from that: AI still depends on human beings continuing to connect with reality.

That includes at least a few layers.

First, infrastructure. Energy, compute, networks, chips, data centers: none of that grows by itself out of AI.

Second, modeling reality. The real world does not automatically turn itself into training data. Someone has to observe, record, measure, define, label, and validate it. If you want AI to do better in medicine, transportation, materials, manufacturing, robotics, or agriculture, then someone still has to stay in the real world and keep working there.

Third, calibration. Whether AI outputs are reliable, whether they can enter decision-making, and who is responsible when things go wrong all have to be tested against reality.

Fourth, value judgment. What is worth doing, what should not be done, what risks are acceptable, what costs are not: those are not answers a model can give by itself.

That is why I do not really agree with the easy claim that the stronger AI gets, the more humans only need to “specify the prompt.”

What seems closer to reality is the opposite. The stronger AI gets, the more humans need to know what reality they are dealing with, what problem they actually want to solve, and what consequences they are willing to bear.

Otherwise “specifying the prompt” easily turns into a kind of wish-making with no grip on reality.

How AI Changes People

What worries me most here is not whether AI will take away some jobs. It is whether it will slowly change the way people think.

One obvious change is that AI makes answers too easy to get.

That has clear advantages. It improves efficiency, lowers barriers, and saves time.

But the cost may be that people spend less and less time going through the process of figuring things out, understanding things for themselves, making mistakes, and finally thinking a problem through.

And a lot of real understanding grows precisely inside that process.

A person learns to cook not by reading ten recipes, but by knowing when the heat has gone too far, when the salt is too much, when the pan should come off the stove. A person learns management not by reading thirty articles about leadership, but by slowly realizing what went wrong through failed conversations, bad judgment calls, and unclear lines of responsibility. Writing is the same: what builds ability is not the final paragraph that sounds polished, but the stretch from confusion to clarity.

If AI increasingly handles that stretch for people, then yes, people will become faster. But they may also slowly lose something.

I think the first things to loosen are these: patience for exploring reality, the ability to form understanding firsthand, and the sense of responsibility for one’s own judgment.

Because when a sentence is written by you, you usually know which part you really understand and which part only sounds like you do. But when the paragraph is generated by a model, it becomes easy to think you are making a judgment when in fact you are only selecting among options.

What Humans Still Need to Do

If those worries are real, then what matters most for humans in the age of AI may not be competing with models over who can produce smoother language.

What matters more may be continuing to do the things abstract systems cannot do well, or cannot do reliably yet.

That means staying in contact with reality.

Doing firsthand observation.

Defining the real problem.

Building the interface that turns reality into models, and model outputs back into reality.

Bearing consequences.

Making value judgments.

Doing the things that cannot be completed by a merely plausible answer.

In other words, AI can become more and more like a powerful abstract system, while humans still have to make sure abstraction does not completely detach itself from reality.

The Human-AI Relationship I Find Most Convincing

The view I find most convincing is a plain one: AI is not reality itself, and it is not a substitute for human beings. It is better understood as a powerful tool built on top of human expression.

It can help humans handle large amounts of symbolic work, amplify efficiency, and extend our capabilities.

But that only works on the condition that humans are still in touch with reality, still providing experience, still doing modeling, still calibrating, still making value judgments.

If that line breaks, AI will become more and more like a system that only circulates within abstraction. It may still generate many beautiful answers, but those answers will drift farther and farther from reality.

So I do not like to understand the relationship between AI and humans as a matter of “who replaces whom.” I prefer to think of it this way: AI handles abstraction efficiently; humans are still responsible for connecting abstraction back to reality.

In the End

So let us return to the first question: what will humans lose?

I think what humans are most likely to lose is not any single skill, but the ability to grow understanding slowly through contact with reality, lived process, and responsibility for consequences.

If that really happens, people may look more efficient, more articulate, and quicker to arrive at answers, while the judgment underneath quietly becomes hollow.

So I am not against AI. I use it too.

I just increasingly feel that in the age of AI, people need to consciously preserve a few things: the ability to think a little more slowly, the ability to stay in contact with reality, and the ability to form judgment for themselves.

Because once those things are lost, what disappears may not just be a way of working. It may be a way of knowing the world.

References

1. Barsalou, L. W. (2008). Grounded cognition. Annual Review of Psychology, 59, 617–645.
2. Noë, A. (2004). Action in Perception. MIT Press.
3. Hauk, O., Johnsrude, I., & Pulvermüller, F. (2004). Somatotopic representation of action words in human motor and premotor cortex. Neuron, 41(2), 301–307. https://doi.org/10.1016/S0896-6273(03)00838-9
4. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745
5. Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310. https://doi.org/10.1038/s41598-020-62877-0
6. Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.