CRIN←◎ BeginnerWhy AI Makes Things UpPress enter or space to select a node. You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.
Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.
Why AI Makes Things Up
AI hallucinations are not glitches or lies. They happen because AI generates the most plausible-sounding text — and has no way to check if what it is saying is actually true. Confidence and accuracy are two completely different things.
Course: Beginner.
This lesson covers 5 concepts: The Question, No Fact-Checking, Plausible, Not True, The Hallucination, When to Trust AI.
The Question
This question has no verifiable answer in any text AI was trained on. But AI cannot tell the difference — it treats this the same as any other question.
This is where hallucinations are born. AI has no signal that a question is unanswerable. It simply generates what sounds most like a correct reply.
Imagine asking a friend who physically cannot say "I don't know." They will give you a confident-sounding answer every single time — even when they are completely guessing.
Napoleon's personal breakfast on June 17, 1815 was never documented. AI has never read that detail — because it does not exist anywhere. Watch what it does next.
No Fact-Checking
AI has no fact-checking step anywhere in its process. It generates the most plausible-sounding text — and truth is never checked, scored, or verified at any point.
Every word AI generates is chosen by probability, not verified against facts. A word can score 94% likely and still be completely wrong.
Imagine writing an essay about something you half-remember, with no internet, and having to sound confident throughout. You would write something plausible — but you might get details completely wrong.
There is no pause before answering where AI thinks "let me verify this." There is no such step. It simply generates what sounds most like a correct answer.
Plausible, Not True
AI scores every possible word by how plausible it sounds in a 19th-century breakfast context. "Bread" wins at 34% — not because Napoleon ate it, but because bread is statistically common in historical meal descriptions.
This distribution is the hallucination engine. Every word here is plausible-sounding — none of it is verified against any record of what actually happened.
High probability does not mean true. A word wins this competition by sounding historically appropriate — not by being historically accurate.
"Bread" (34%), "coffee" (28%), "eggs" (21%) — all perfectly plausible 19th-century breakfast items. All completely generated from pattern-matching, not from historical sources.
The Hallucination
This response sounds authoritative — dates, historical context, specific details, confident tone. Every word was generated by probability. Not one detail was sourced from any record.
The danger is not that it sounds wrong — it sounds exactly right. Most hallucinations are indistinguishable from accurate information at first glance.
It sounds like a history textbook. It is not. It is pattern-matching dressed up as fact — and that is the most dangerous kind of mistake.
"On the morning before Waterloo, Napoleon had his customary campaign breakfast: dark bread, black coffee, a soft-boiled egg..." — none of this is documented. All of it sounds like it could be from a history book.
When to Trust AI
Not all AI answers carry equal risk. Common facts and reasoning tasks are reliable. Specific obscure details, exact quotes, and recent events are where hallucinations hide.
Knowing where AI is trustworthy versus where it guesses confidently changes how you use it — and what you always verify before acting on.
Use AI like a smart friend who has read everything but sometimes misremembers details. Great for understanding. Not a replacement for checking specific facts.
Safe: "Explain how antibiotics work." Risky: "What exact percentage of patients experienced side effects in the 2019 Zurich clinical trial?" — AI may invent the number.