Recently, AI chatbots like ChatGPT have become popular worldwide. They can write poetry, code, solve math problems, and even pass the bar exam. Many people are exclaiming: Is AI about to achieve true understanding?
But hold on. First, consider this question: A parrot can fluently say “I am hungry,” but does it really understand the meaning of that phrase?
The Dilemma of Passive Learning: The Super Achiever Who Can Only “Repeat”
Modern AI is like that parrot, only a supercharged version. Their learning method is very passive: they consume the entire internet’s text, images, and videos, and then desperately memorize the statistical patterns—what words often follow which words, and which images typically accompany which descriptions.
This process is akin to reading every recipe in the world without ever stepping into a kitchen. You know that the word “stir-fry” is often followed by “vegetables” and that “a pinch of salt” means something. But you have never felt the sting of hot oil splattering on your skin, nor tasted the unpleasantness of an unbalanced flavor. Your “knowledge” is flat and cold, like a manual that has never been opened.
The paper gives an example: AI can perfectly describe the instruction “walk 10 meters north, then 10 meters south,” even providing the correct answer. But it has never actually “walked” a step. For it, “north” is just a symbol, and “10 meters” is just a number; these words lack muscle memory, a sense of direction, and the solid feeling of “returning to the starting point.”
Active Exploration: This Is How We Understand the World
Living beings (including you and me) learn in a completely different way. We do not download data in the background; we learn through “trial and error” in the world.
Babies wave their limbs as soon as they are born, and toddlers learn to walk by stumbling. Every push, pull, grasp, and fall leaves a deep imprint in the brain. When you first lift a pot lid and get burned by steam, you not only learn the term “steam burn” but also permanently record the instinctive reaction of “danger! Get away!” in your body.
Scientists call this learning method active inference. Simply put: our brains are prediction machines, but they are also action machines.
You do not just passively receive information; you actively choose what to see, touch, and taste. If you want to know whether honey is sweet, you dip your finger in it; if you want to understand a cat’s personality, you go to pet it (and might get scratched). Every action validates or corrects your predictions: “Hmm, honey is indeed sweet,” “Oh, it turns out cats don’t like having their bellies rubbed.”
This hands-on feedback is the foundation of true understanding. You understand “table” not just because you have seen the word a thousand times, but because you have placed a cup on it countless times (it bears weight), crawled underneath to retrieve a ball (it has space), and lifted it during a move (it has weight). These concrete interactive experiences weave a rich web of meaning.
The Alien’s Confusion: Understanding the World Through Words Alone
The paper presents a clever thought experiment. Imagine an alien named Wordy, which has only one sensory channel: receiving streams of human text. It survives by predicting the next word, having read our novels, news, recipes, and social media.
Wordy might perfectly mimic human speech, but it lives entirely in a “vacuum.” When a human chef burns a dish, they taste the bitterness, smell the char, and see their family frowning—these real consequences immediately correct their understanding. But Wordy only sees the phrase “burned” and might follow it with the word “inedible.” It learns the associations between words, not the causal chains of the real world.
This is like learning about love by watching ten thousand romantic movies without ever holding anyone’s hand. You might be able to write a touching love letter, but the words lack the memory of a racing heartbeat or the tingling sensation of holding hands. Your “understanding” is second-hand, borrowed.
Why AI Does Not Understand “Falling Hurts”
Lacking bodily experience, AI has a fatal shortcoming: it does not know what is important.
For humans, “survival” is the lowest priority. You fear the dark, heights, and snakes; these are self-preservation programs etched into our genes through evolution. Babies crave their parents’ embrace because warmth, safety, and love are essential for survival. These feelings of “care” make certain information resonate loudly in the brain, while others fade into background noise.
AI lacks this “care.” It does not know that “errors” have consequences beyond affecting a numerical score. It generates dangerous suggestions without fear of injury; it spouts nonsense without feeling shame. Its “attention mechanism” is merely mathematical weight, not genuine concern.
This explains why AI can “seriously spout nonsense.” It lacks real-world anchors to verify its outputs. Humans feel guilty when lying because they anticipate being caught and punished. When AI generates false information, it feels no internal conflict—it has no inner self.
The Path Forward: AI Needs to Learn to “Live” First
So, how can AI achieve true understanding?
The paper’s author believes the answer is not to “feed more data” but to let AI first learn to “live.”
Just like in human evolution: our ancestors first learned to walk, run, and avoid danger, taking millions of years to develop language. Language is built on solid bodily experiences. Trying to make AI master language before acquiring bodily experience is like asking someone who has never seen the ocean to understand “nostalgia.”
A promising future path for AI should be:
1. Interaction First, Language Later: Let robots navigate the real world to establish basic bodily knowledge—gravity, inertia, causality. Just like babies learn to grasp a milk bottle before learning to say “milk.”
2. Curiosity-Driven: Not just completing tasks, but allowing AI to actively explore “what I do not know.” It should be fascinated by moving shoelaces and eager to explore unknown corners, like a kitten.
3. Facing Consequences: When AI’s suggestions lead to bad outcomes, it needs to “suffer the consequences.” This feedback can establish true causal understanding rather than mere statistical correlation.
Of course, this raises new ethical dilemmas: if an AI truly begins to care about its “survival,” how can we ensure its goals align with those of humanity? But that is another story.
Mirrors and the Outside World
Current generative AI is like a super high-definition mirror. It reflects human knowledge and can even stitch together patterns we have never thought of. But behind the mirror, there is no consciousness, no self, and no real concern for the world.
It is one of the most fascinating inventions of the 21st century, allowing us to glimpse the nature of our own thinking from the outside. But when we lean in to take a closer look—there is no one looking back in the mirror.
True understanding begins with a body that can cry and laugh, feel pain and fear, starting from those moments that make our hearts race and palms sweat. For AI to bridge the gap from “processing symbols” to “understanding meaning,” it may need to start by learning to “live.”
—
(Adapted from “Generating meaning: active inference and the scope and limits of passive AI”)