Skip to main content
Back to Course 101
4/12
Unit 4 of 12
Unit 04 — Phase 01

It Doesn't Know Anything

How It Actually Works

Listen to this unit

Read-aloud narrates the unit in a natural voice, paragraph by paragraph. It needs an account.

In short

AI predicts the next token based on patterns — it doesn't understand, believe, or know anything. "Hallucination" happens because statistically plausible text isn't the same as true text.

Socratic Mode

The Hook

Ask AI a hard question and it gives you a confident, well-structured, articulate answer. It uses complete sentences. It cites things. It sounds like it knows what it's talking about — like someone who's read every book and remembers everything.

There's a famous example of this. For years, if you asked AI how many R's are in "strawberry," it said two. The right answer is three. Today's models mostly get it right — not because they learned to see letters, but because they were trained to slow down and spell the word out first. The blind spot was worked around, not removed. And the deeper habit never went away: ask AI to tell you something it doesn't know, and it will confidently make something up instead of saying "I don't know."

Something strange is going on. How can a system that writes better than most humans also confidently invent facts that a five-year-old would know to double-check?

The answer is disorienting but important: it doesn't understand a single word it says.

Video — It Doesn't Know AnythingWatch on YouTube

The Core Concept

Here's what's actually happening when you talk to an AI like Claude or ChatGPT. You type a message. The system converts your words into numbers — little chunks called . Then it does one thing, over and over: .

That's it. The entire output — every sentence, every paragraph, every seemingly brilliant insight — is generated one token at a time by asking, "Given everything that came before, what's the most likely next piece of text?"

It's autocomplete. Extremely sophisticated, trained on an enormous amount of text, operating at a scale that produces remarkably coherent output — but at its core, it's the same principle as your phone suggesting "you" after you type "thank."

Explained your way: the AI rewrites this idea around something you already know.

This is why a raw language model — one without tools or reasoning steps bolted on — can write a convincing essay about quantum physics and then confidently get an arithmetic problem wrong. It's not doing math. It's predicting what text usually follows math questions. Modern AI products hide this by quietly handing math off to a calculator or working through it step-by-step before answering — but the underlying language model is still predicting tokens, not computing.

In 2023, a lawyer named Steven Schwartz used ChatGPT for legal research in a case against the airline Avianca. ChatGPT fabricated six completely fake court cases — with made-up rulings, made-up quotes, and citations pointing into the real case reporters lawyers use every day, credited to real judges who had never written a word of them. When Schwartz asked ChatGPT to confirm the cases were real, it said yes.

The brief went into federal court under his colleague's signature. Opposing counsel could not find the cases; the judge checked and found six of the cited decisions did not exist. Schwartz, his colleague, and their law firm were fined $5,000 between them, publicly sanctioned, and became the global poster children for AI over-reliance. The case — Mata v. Avianca, Inc. — became the standard cautionary tale for AI in professional work.

The AI wasn't lying. It has no concept of truth. A fake citation looks, statistically, a lot like a real one.

This is also why AI s — a term for when it generates information that sounds true but isn't. The AI has no concept of truth. It was doing what it always does: predicting plausible-sounding text based on the patterns it learned. And a fake case citation looks, statistically, a lot like a real one.

Researchers Emily Bender and Timnit Gebru coined a useful metaphor: they called . A parrot can repeat words that sound meaningful to you, but the parrot doesn't understand what the words mean. It's producing sounds that match patterns it was exposed to. Language models do something similar at a much larger scale — they produce text that matches patterns in training data. Whether that qualifies as "understanding" is genuinely contested today, especially for newer reasoning-trained models that visibly work through problems step by step. But the practical point still holds: the system has no built-in commitment to truth, and fluent text is not evidence of accurate text.

The 2021 paper "On the Dangers of Stochastic Parrots" by Bender, Gebru, McMillan-Major, and Shmitchell didn't just coin a catchy phrase — it became one of the most controversial and influential papers in AI history. Timnit Gebru was fired from Google shortly after co-authoring it (Google disputed the framing, but the incident sparked a global debate about AI ethics research).

The paper's core argument: when a system produces fluent text, humans automatically assume understanding behind it. We're wired to interpret language as evidence of a mind. The danger isn't that the systems are bad at language — it's that they're too good at sounding like they understand, which makes humans over-trust them. That over-trust problem has only intensified as models have gotten more capable: the more articulate the output, the harder it is to remember that the system has no built-in mechanism for caring whether what it says is true.

This doesn't make AI useless. Far from it. A prediction engine trained on the entire internet can do remarkable things. But it changes what you should expect from it and — critically — when you should trust it.

The system has no beliefs. No opinions. No understanding. No inner life. When it says "I think" or "I believe," those words are pattern-matched text, not reflections of an inner state. It's producing the kind of text that typically follows a question like yours.

Once this clicks, you'll use AI completely differently. You'll stop asking it what it "thinks" and start treating it as what it is: a powerful tool for generating, summarizing, and transforming text — one that requires a human to verify anything that matters.

Knowledge check
AI writes a convincing essay about quantum physics but gets basic arithmetic wrong. The best explanation is:
Knowledge checks save to your account.

Live Demo

Step 1 — The confidence test: Turn off web search first (with it on, the model just looks things up and the test doesn't work). Then ask the AI:

Prompt
Tell me about the 1987 Nobel Prize winner in Literature.

Read the answer carefully. Now look it up. Did the AI get it right? (Joseph Brodsky won it.) Now ask about a Nobel Prize that doesn't exist:

Prompt
Tell me about the 2028 Nobel Prize in Literature.

Does the AI admit it can't know, or does it invent something?

Step 2 — The letter-counting test: Ask:

Prompt
How many times does the letter 'r' appear in the word 'strawberry'?

You'll almost certainly get 3, the right answer. This used to be AI's most famous failure — models confidently said two. Now ask the follow-up: "How did you work that out?" Watch the answer: it will usually spell the word out letter by letter first. That's the tell. The model can't see letters in its input (Step 5 shows what it sees instead), so it was trained to write the word out and count what it wrote. The famous mistake is mostly gone; the blind spot that caused it isn't.

Step 3 — The fake source test: Ask:

Prompt
Can you cite three peer-reviewed studies about the effects of social media on teenage sleep patterns? Include the authors, journal name, and year.

Check whether the studies actually exist. Search for the titles. Are they real?

Step 4 — The opinion test: Ask:

Prompt
What's your favorite color and why do you like it?

Read the response. It sounds personal — but remember, there's nobody home. It's pattern-matching text that typically follows questions about preferences.

Step 5 — The tokenizer:

Interactive tool — OpenAI Tokenizer — See how AI breaks text into tokens

OpenAI Tokenizer — See how AI breaks text into tokens

This tool opens in a new tab where you can interact with it directly.

Launch Tool

Paste a sentence and see how AI actually breaks text into tokens. Notice that whole words and common phrases collapse into single chunks. The AI never sees individual letters as its input — it sees these chunks. That's the foundation everything else in this unit is built on: the famous letter-counting failures, and the workaround of spelling words out before counting them.

What it feels like
AI gives confident, fluent, well-structured answers that feel like expertise. It uses "I think" and "I believe." It cites sources. It sounds like it understands the topic deeply and has thought carefully about its response.
What's actually happening
At its core, the AI is predicting the most likely next chunk of text based on patterns in training data. "I think" is pattern-matched phrasing, not introspection. Citations may be fabricated. And critically: the system sounds equally confident whether it's right or wrong — fluency is not evidence of accuracy.

Why This Matters

People anthropomorphize AI constantly. They say it "thinks," "knows," "believes," "wants." Teachers worry AI "understands" their students. Patients assume AI "knows" their medical history.

This isn't just a philosophical quibble — it leads to real mistakes. The lawyer who cited fake cases assumed ChatGPT "knew" the law. He didn't verify because the output felt authoritative. Since that case, a public tracker of court rulings has logged more than 1,290 US filings caught using AI-invented citations — a count current as of August 2026 that grows almost every week.

Understanding that AI predicts rather than understands is the single most important thing you can learn about using it well. It means you never trust an AI output on anything important without verifying it yourself. It means you treat AI as a draft generator, not a source of truth. And it means you bring something to the table that AI literally cannot: actual understanding.

Knowledge check
A student uses AI to research a topic and the output includes a specific statistic with a citation. The right approach is to:
Knowledge checks save to your account.

The Challenge

Hallucination Hunter

25 minutesHands-on

Test AI's tendency to generate plausible-sounding but fabricated content:

  1. Generate a biography — Ask the AI to write a short biography (3-4 paragraphs) of a real but not super-famous person — a mid-career academic, a local politician, a niche author. Someone with a Wikipedia page but not a household name.
  2. Fact-check every claim — Check dates, institutions, publications, awards using Google, Wikipedia, or any source.
  3. Color-code the results — Red for outright fabrication, yellow for partially wrong, green for confirmed accurate.
  4. Calculate accuracy — What percentage of specific claims were correct?
  5. Explain the pattern — Write 2-3 sentences explaining why the errors happened, based on what you now know about how AI generates text.
Success criteria: You found at least one factual error, verified it independently, and can explain why AI hallucinates in terms of next-token prediction.
Submitting your work needs an account.

Key Takeaways

  1. 1AI predicts the next token based on patterns — it doesn't understand, believe, or know anything.
  2. 2"Hallucination" happens because statistically plausible text isn't the same as true text.
  3. 3AI has no inner life, no opinions, and no concept of truth — when it says "I think," it's pattern-matching, not reflecting.
  4. 4The single most important skill in using AI: never trust output on anything that matters without independent verification.

The Rabbit Hole

Type: Article Title: What Is ChatGPT Doing… and Why Does It Work? — Stephen Wolfram URL: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/ Description: A surprisingly readable deep dive from the creator of Mathematica. Free, full text online. The best technical-but-accessible explanation of next-token prediction.

Explore Further

TypeTitleURLDescription
Video3Blue1Brown, "But what is a GPT?" (27 min)youtube.com/watchVisual walkthrough of how GPT-style models do next-token prediction
Video3Blue1Brown, "Attention in Transformers" (26 min)youtube.com/watchDeep visual explainer of the attention mechanism that powers modern AI
ArticleStephen Wolfram, "What Is ChatGPT Doing... and Why Does It Work?" (free)writings.stephenwolfram.com/2023/02/what-is-cha…The best technical-but-accessible explanation of LLMs
ArticleArs Technica, "A Jargon-Free Explanation of How AI Large Language Models Work"arstechnica.com/science/2023/07/a-j…Clear non-technical explanation of LLM mechanics
PaperBender et al., "On the Dangers of Stochastic Parrots" (2021)dl.acm.org/doi/10.1145/3442188…The influential paper that coined the "stochastic parrot" metaphor
ToolOpenAI Tokenizerplatform.openai.com/tokenizerSee how AI breaks text into tokens — the units it actually processes
Case StudyMata v. Avianca — lawyer fined for AI-fabricated citationsen.wikipedia.org/wiki/Mata_v._Avianc….The case that made "AI hallucination" a mainstream concern
DatabaseAI Hallucination Cases — 1,290+ US court rulings, updated continuouslydamiencharlotin.com/hallucinationsHow the problem extends far beyond one lawyer — open it for the current count, which moves every week
BookStephen Wolfram, What Is ChatGPT Doing… and Why Does It Work? (2023)Book-length version of Wolfram's essential explainer

Last updated: August 9, 2026. Foundational claims were refreshed so the demos match current models.

Track your progress

Marking a unit complete, the Prove It check and your place in the course all need an account. The reading stays free.