智能的幻觉：当我们高估了自己的造物

Just a few days ago, the journal Nature published a striking article introducing a new book, "The Illusion of Artificial Intelligence," by AI pioneer Luc Julia.

Julia's core argument is very sharp — he says today's most advanced AI models are essentially nothing more than "glorified pocket calculators."

This statement sounds a bit harsh, and you might even think he's just trying to grab attention.

But if you think carefully, behind this judgment lies a fundamental question we've been avoiding for a long time.

Let's turn the clock back to 1956.

That year, at the Dartmouth Conference, the term "artificial intelligence" was formally proposed for the first time.

From that moment on, these two words were destined to be misunderstood.

Why?

Because the word "intelligence" itself carries a strong human-centered connotation.

When we hear "intelligence," what comes to mind is thinking, judgment, creativity, and insight — these are the core abilities of human consciousness.

So when we attach the word "intelligence" to machines, we're already unconsciously engaging in a kind of metaphorical projection.

This projection, in more academic terms, is called "anthropomorphism."

We humans are naturally inclined this way — we see clouds and think they look like dragons, we talk to pets thinking they understand, we give our cars names.

This psychological mechanism is amplified to the extreme when facing AI.

You chat with ChatGPT, it responds coherently and logically, and you feel it "understood" your question; you ask AI to paint a picture, it produces something exquisite, and you feel it has "creativity."

But there's a fatal confusion here — we equate "the quality of the output" with "the process that produced the result being intelligent."

Let me give you an analogy.

A precision loom can weave magnificent brocade with complex patterns and brilliant colors.

But you wouldn't say the loom "understands" aesthetics, nor would you say it "created" art.

It's simply executing a preset program precisely.

Today's large language models, fundamentally speaking, are doing something no different from a loom — they search for statistical patterns in massive amounts of data, then generate the next most likely word based on probability distributions.

This process is precise, efficient, and awe-inspiring, but it is not "thinking."

Julia's insight lies in pointing out a collective cognitive bias we all share.

Humans have a deep-rooted tendency to overestimate the capabilities of what they create.

This is nothing new.

Every major technological breakthrough in history has been accompanied by similar illusions.

In the 19th century, people saw steam locomotives racing past and thought machines would soon replace all human labor; when electricity became widespread in the early 20th century, some predicted that humanity would eliminate all diseases within fifty years.

The power of technology is real, but our imagination about technology always runs faster than the technology itself.

The current AI boom is, to some extent, the latest episode of this historical cycle.

Major tech companies are racing to announce that their AI "possesses reasoning ability," "can creatively solve problems," and "is about to achieve artificial general intelligence."

Capital markets add fuel to the fire, media embellishes the story, and the public oscillates between amazement and fear.

But if you strip away the fancy rhetoric and look at the technology itself, you'll find a less glamorous truth: these systems excel at pattern matching and statistical inference, and their performance on specific tasks can far exceed humans,

but they have no idea what they're doing.

This leads to a deeper philosophical question: what is "intelligence"?

Or rather, what is "understanding"?

The "Chinese Room" thought experiment proposed by John Searle in 1980 remains profoundly resonant to this day.

A person who doesn't understand Chinese at all is locked in a room with a comprehensive rule book.

People outside pass in Chinese notes, and he looks up the corresponding answers in the book and passes Chinese responses back out.

From the outside, the room "speaks Chinese"; but the person inside has no clue what the Chinese means.

Today's AI is the digital version of that room.

It processes symbols but doesn't understand meaning.

Some might argue: who cares whether it understands?

As long as the results are good, that's all that matters.

This pragmatic attitude certainly has its merits.

A good hammer doesn't need to "understand" what a nail is.

But the problem is, when we slap the label of "intelligence" on the hammer, we start having unrealistic expectations — we expect it to judge on its own where to strike,

we expect it to understand the blueprint of an entire building, we expect it to make moral judgments at critical moments.

And these are precisely the things a hammer cannot do.

What's more worrying is that this "illusion of intelligence" is influencing our social decisions.

When judges start relying on AI systems to assist with sentencing, when hospitals use AI for preliminary diagnosis, when financial institutions let AI decide who gets a loan — if we mistakenly believe these systems truly "understand" the case,

the condition, or the credit situation, we'll relax our oversight and questioning of them.

In reality, these systems have merely found certain correlations in historical data — they don't know what justice is, what health is, or what risk is.

Their biases are hidden in the training data, and their errors are masked by the cloak of statistics.

Julia is not denying the value of AI.

Quite the opposite — he's calling on us to approach this technology with a more clear-headed and honest attitude.

AI is a powerful tool — possibly one of the most powerful tools humanity has ever invented.

But a tool is a tool.

A scalpel in a surgeon's hands can save lives; in the hands of someone with no medical knowledge, it's just a dangerous piece of metal.

The key isn't how "intelligent" the tool is, but how wise the person using it is.

I think Julia's most valuable contribution is that he reframed the relationship between us and AI.

We shouldn't view AI as some "new intelligence" that's closing in on humanity, but rather understand it as a precise yet narrow information processing system.

It can beat the world champion at Go, but it doesn't know it's playing chess.

It can write smooth and fluent articles, but it doesn't know what meaning the words carry.

It can pass almost every standardized test, but it has no intellectual curiosity whatsoever.

This distinction isn't about belittling AI — it's about protecting ourselves.

Because when we overestimate AI, we're simultaneously doing two dangerous things: first, we abdicate the judgment responsibilities that humans should bear; second, we underestimate the uniqueness and irreplaceability of human intelligence itself.

Human intelligence isn't just the ability to process information — it encompasses emotions, intuition, moral judgment, the pursuit of meaning, the perception of beauty, and empathy for suffering.

None of these things can be captured by any statistical model, nor simulated by any algorithm.

Nearly seventy years have passed, and perhaps we should seriously consider whether it's time to give this technology a more accurate name.

The term "artificial intelligence" has caused too deep a misunderstanding — it has led countless people to believe we're actually creating some form of "intelligence."

Perhaps "advanced automated information processing" doesn't sound as exciting, but at least it's honest.

At the end of the day, true wisdom isn't about the speed and precision of processing information, but about maintaining a clear understanding of the world — including a clear understanding of what we ourselves have created.

In this age where AI is everywhere, maintaining this clarity is itself a remarkable form of wisdom.