Before we talk about what AI is, let's name what most people think it is. There are two dominant wrong models, and they each cause predictable problems.
The Oracle
You ask it a question. It knows the answer. It's essentially a smarter Google — a vast database of truth that retrieves the correct answer when prompted correctly.
What this produces: You trust outputs without verifying. You get confused when AI is confidently wrong. You treat every response as fact.
The Robot
It follows instructions mechanically, like sophisticated autocomplete. Very fast, very literal, executing commands without anything like intelligence.
What this produces: You dramatically underuse AI's capabilities. You don't understand why careful prompting matters. You give up quickly when results disappoint.
Both models produce systematic mistakes. The Oracle produces over-trust. The Robot produces under-use. The accurate model produces neither.
What an LLM Actually Does
A Large Language Model is, fundamentally, a prediction engine.
It was trained on a massive dataset of text — books, websites, code, conversations, articles — and learned one thing with extraordinary precision:
Given this sequence of text, what word (technically, what "token") is most likely to come next?
That's not a metaphor. That's the literal mechanism. Every word in every AI response is selected through a probabilistic process of predicting the most contextually appropriate continuation of what you wrote.
What emerged from doing this at massive scale on virtually all of human-written language is something genuinely remarkable: a system that can reason through problems step by step, write code, translate languages, analyze documents, and explain concepts at multiple levels of complexity.
But the mechanism is prediction, not understanding. This distinction has real consequences.
The Alien Translator
Here's the mental model that actually works: imagine an entity that has read essentially everything humans have ever written, but has never experienced anything.
It has never seen the color red, felt cold, made a decision with real stakes, or cared about an outcome. It has no preferences, no fear, no curiosity. What it has is an extraordinarily rich statistical model of how humans write about all of these things.
It doesn't know things the way a person knows things. It knows them the way a language knows them — as patterns in how things are expressed. This model predicts your actual experience with AI far better than either wrong model:
- Why it writes beautifully about grief without feeling it: pattern-matching on how grief is written about, not experiencing it
- Why it hallucinates: when there's no strong pattern match for a specific fact, it generates something that looks like a correct answer — the shape of truth without the substance
- Why prompting matters: you're feeding context to a prediction engine that generates the most plausible continuation of what you wrote
- Why it's wrong with total confidence: confidence is a tone pattern, not a verification signal
What AI Is Good At — and Where It Fails
Once you hold the right model, AI's capabilities and limitations become predictable rather than mysterious.
- Generating well-structured text on any topic
- Drafting, rewriting, summarizing, reformatting
- Explaining concepts at different complexity levels
- Writing code that follows common patterns
- Synthesizing multiple sources into a summary
- Generating options, variations, and alternatives
- Following complex multi-part instructions
- Specific factual claims (events, statistics, citations)
- Precise mathematical calculation
- Information after the model's training cutoff
- Anything where being precisely right matters
- Genuinely original ideas (it generates novelty that looks original)
- Knowing what it doesn't know — rarely signals uncertainty accurately