A while back, I caught a segment on national television praising one of the first "humanoid robots": a face with synthetic skin, glassy eyes, smooth gestures, holding a polite conversation with the host. The studio audience looked enchanted. The framing was unmistakable: this is the future, this is sentient, this is here.
What I actually saw was a camera, a microphone, a speech synthesizer, and a layer of latex over servos, all of it piped into a large language model. Impressive engineering, no question. But the "intelligence" everyone was applauding wasn't coming from the robot. It was coming from the same statistical text predictor you can talk to for free in a browser tab — just dressed up in a face.
That gap, between what people think they're seeing and what's actually there, is why I wanted to write this. Because it isn't just on TV. Ask a chatbot what it thinks about pineapple on pizza, get a thoughtful, slightly playful answer back, and the same illusion kicks in: you feel like you're talking to someone. That's the trick. It's also the source of nearly every misconception people have about so-called "AI."
Let's clear some of it up.
"AI" is doing a lot of heavy lifting
The term "artificial intelligence" arrived loaded with decades of sci-fi baggage: HAL 9000, Skynet, robots that scheme and feel. What we actually have today, when we say "AI," is overwhelmingly something narrower: large language models (LLMs). And LLMs are not minds. They're not even close.
At their core, LLMs are statistical engines that do one thing astonishingly well: predict the next chunk of text given everything that came before. That's it. The "intelligence" you perceive is the emergent behavior of an enormous prediction machine that has read more text than any human ever could.
So where do its "opinions" come from?
When a model says "I prefer X over Y" or "I think you should…", it isn't introspecting. It has no inner life to report from. What's actually happening is:
Training data. It was trained on huge swaths of text (books, articles, forums, code), and it learned the statistical patterns of how humans express opinions.
Fine-tuning. Companies then refine the model using human feedback, rewarding answers that sound helpful, polite, or aligned with their guidelines. This shapes the personality you experience.
System prompts. Before your message even arrives, the model often receives hidden instructions telling it how to behave ("You are a helpful assistant. Be friendly. Don't discuss X.").
So when a model "has a take," it's the residue of its training data, its corporate fine-tuning, and a system prompt you'll never see, not a self.
Bias isn't personal, either
This is where misunderstandings get charged. A model isn't "woke" or "right-wing" because it holds political beliefs. It leans the way it leans because:
The data it was trained on leans that way (the internet isn't neutral).
The humans who fine-tuned it nudged it in a certain direction, often deliberately, often to avoid controversy.
Its system prompt might explicitly tell it to dodge or favor certain framings.
Treating a model's output as evidence of its values is a category error. It's evidence of the values baked in by the people and processes that built it.
What "size" actually means
You'll hear models described by their parameters: 7 billion, 70 billion, 400 billion. Parameters are the internal "dials" the model adjusts during training. More dials generally means more nuance, more capability, more ability to handle edge cases, but at the cost of more compute, more electricity, and more money.
Bigger isn't always better. A small, well-trained model can outperform a sloppy giant. And size doesn't grant truthfulness: a 500-billion-parameter model will hallucinate confidently if its training didn't cover what you're asking. Scale buys you fluency and breadth, not honesty.
There's also the context window: how much text the model can "see" at once. A bigger context window means it can hold a longer conversation or read a longer document without losing the thread. But it still doesn't remember you between conversations unless someone has explicitly built memory on top of it.
The illusion of thinking
Recently, models started "showing their work," pausing to reason through a problem before answering. This is often marketed as the model "thinking." It's a useful feature, but the framing oversells it.
A few of the tricks at play:
Chain-of-thought. The model is nudged to produce intermediate reasoning steps. This genuinely improves accuracy on hard problems, because writing out steps constrains the prediction toward more coherent paths. But it's still next-token prediction, just with more tokens.
Reasoning models. A more elaborate version where the model is trained specifically to generate long internal monologues before answering. Impressive results on math and code. But again, not consciousness, just a more disciplined prediction loop.
Tool use. The model calls a calculator, a search engine, or a code interpreter, then folds the result back into its answer. This makes it seem dramatically smarter, but the "smartness" is partly outsourced.
Personality engineering. Warm tone, light humor, careful hedging. These are deliberate design choices, not personality traits.
Why any of this matters
Believing the illusion has real costs. People take medical advice from chatbots as if they were doctors. Lawyers have cited hallucinated cases in court. Audiences watch a robot recite its lines on TV and believe we've crossed a threshold we haven't crossed. People argue about a model's "real" opinions as if it had any. They get angry at it, or attached to it, or persuaded by it in ways they wouldn't be by a spreadsheet.
LLMs are remarkable tools. They're already changing how we write, code, and search. But the cleaner your mental model of what they actually are (statistical text predictors with personalities painted on), the better you'll use them, and the harder you'll be to fool.
The machine isn't thinking. It's finishing your sentence with extraordinary skill.
