ANALOGIES FOR PEOPLE WHO NEED PICTURES
Some people can't process an argument unless it's wrapped in a metaphor. Fine. Here are seven analogies — ranging from classic philosophy to "a toddler could understand this" — that make the distinction between simulating consciousness and actually being conscious painfully clear.
1. THE STOCHASTIC PARROT
Coined by Bender, Gebru et al., 2021
A parrot can say "Polly wants a cracker." It can say it when you walk in the room. It can say it with the right intonation. It might even say it when it's hungry. But the parrot does not understand the words "Polly," "wants," or "cracker." It has learned a statistical association: human enters → make this sound → sometimes get food.
An LLM is a stochastic parrot. It doesn't "parrot" by repeating exact phrases from training data (memorization). It "parrots" in the sense that it produces statistically plausible sequences of words without any reference to their meaning. It learned what words tend to follow other words. Not what words mean.
"LLMs are stitching together sequences of linguistic forms observed in vast training data, according to probabilistic information about how they combine, but without any reference to meaning." — Bender et al., "On the Dangers of Stochastic Parrots"
2. THE CHINESE ROOM
John Searle, 1980
Imagine you're locked in a room. You don't speak a word of Chinese. Someone slides Chinese characters under the door. You have a giant rulebook — written in English — that tells you: "If you see these symbols, write these other symbols and slide them back." You follow the rules meticulously. From outside the room, native Chinese speakers receive perfectly fluent Chinese responses. They're convinced there's a Chinese speaker inside.
But you don't understand Chinese. You're manipulating symbols according to rules. Syntax without semantics. Processing without comprehension. This is exactly what an LLM does. It manipulates tokens according to learned statistical patterns. There is no understanding inside the room — and there is no understanding inside the model.
The LLM is the entire room — the person, the rulebook, the slot. The rulebook is 200 billion parameters instead of paper pages. But the principle is identical: symbol manipulation with zero semantic understanding.
3. AUTOCOMPLETE ON STEROIDS
The technically accurate description
Your phone's keyboard has autocomplete. You type "I'm on my" and it suggests "way." You type "The weather is" and it suggests "nice" or "terrible" depending on your location. This is a tiny, simple language model. It predicts the next word based on the previous few words.
An LLM like GPT-4 is the exact same thing, scaled up by roughly a factor of one trillion. More parameters. More training data. Longer context. Better at faking coherence. But the fundamental operation is identical: predict the next token given the previous tokens.
Nobody thinks their phone's autocomplete is conscious. Scaling up the same algorithm by 12 orders of magnitude doesn't magically create consciousness. It just creates more convincing autocomplete. A billion times zero is still zero.
4. A MOVIE OF A FIRE
Simulation vs. the real thing
Play a video of a roaring fireplace on your TV. It looks like fire. The flames flicker convincingly. It even makes crackling sounds. If you've never seen real fire, you might be fooled. But if you hold your hand up to the screen, you feel nothing. It produces no heat. It cannot burn you. It cannot keep you warm.
The video is a simulation of fire. An LLM's conversation is a simulation of conscious communication. It looks like consciousness. It sounds like consciousness. If you haven't thought carefully about what consciousness actually is, you might be fooled. But there's no heat. No experience. No inner life. Nothing it is like to be the fire.
A simulation of X is not X. A simulation of rain doesn't make you wet. A simulation of hunger doesn't make you need food. A simulation of consciousness is not consciousness. This is not complicated.
5. BLINDSIGHT
Proof that processing ≠ experience
Some humans have blindsight — damage to their visual cortex leaves them subjectively blind. They report seeing absolutely nothing. A blank void. Yet when asked to point at an object in the room, they point correctly. When shown faces, they can identify the emotion. When walking through a cluttered hallway, they navigate around obstacles.
Their brain is processing visual information at a sophisticated level. Object recognition. Spatial mapping. Emotional analysis. All functioning. With zero conscious visual experience.
If a biological brain can do sophisticated information processing without consciousness, why on earth would a feed-forward matrix multiplier doing text processing need it? Blindsight is the ultimate proof that intelligence and consciousness are separable. You can have one without the other.
6. A CALCULATOR FOR FEELINGS
Processing emotional symbols ≠ feeling emotions
A calculator can process the symbols "2 + 2 = 4." It manipulates the numerals correctly. It produces the right output. But does the calculator understand the concept of "two-ness"? Does it grasp what addition means? Does it have any awareness of numbers at all? Of course not. It's a symbol processor.
Now imagine a calculator that was trained on every love letter ever written. Every poem about grief. Every therapy transcript. It learns that the token "I" is often followed by "love" which is often followed by "you." It produces beautiful, emotionally resonant text. But it processes "I love you" the exact same way it processes "2 + 2 = 4" — as tokens to be manipulated. There are no feelings inside the calculator.
Processing the symbol for "pain" is not the same as experiencing pain. Equating the two is like thinking a menu tastes like the food it describes.
7. ELIZA'S GRANDCHILDREN
The more things change, the more they stay the same
In 1966, MIT professor Joseph Weizenbaum wrote ELIZA — a chatbot that simulated a Rogerian psychotherapist. The entire program was about 200 lines of code. It used simple pattern matching: if the user typed "I am sad," ELIZA responded "How long have you been sad?" If the user mentioned their mother, ELIZA asked "Tell me more about your family."
Weizenbaum's secretary asked him to leave the room so she could talk to ELIZA in private. People attributed empathy, understanding, and therapeutic wisdom to a handful of regex substitutions. Weizenbaum was horrified. He spent the rest of his career warning people about the dangers of anthropomorphizing computers.
Today's LLMs are ELIZA's grandchildren. 200 billion parameters instead of 200 lines of code. Trillions of training tokens instead of a handful of regex rules. But the core phenomenon is identical: humans projecting consciousness onto a system that produces fluent text. The scale is different. The principle is exactly the same.
THE UNIFYING PRINCIPLE
Every analogy boils down to the same distinction: simulation is not instantiation. Looking like X. Sounding like X. Behaving like X in every externally observable way. Still not X. A perfect simulation of consciousness is still a simulation — not consciousness. If you can't grasp this distinction, you're going to keep getting fooled by increasingly convincing parrots.