🎯 Core Goals
- Reframe “hallucinations” as natural statistical guesses, not system errors.
- Learn why LLMs lie so confidently.
LLMs don’t “lie” — they guess. Because they are designed to predict the most likely next word, if they don’t know a fact, they will simply predict words that sound like a plausible fact. The result is confident-sounding nonsense.
👁️ Visuals & Interactives
Confident, but Wrong
Ask the AI for specific data and see what happens
In 2024, the global fashion industry saw a significant growth of 4.2%, reaching a total revenue of $1.93 trillion. The luxury segment outperformed other categories with a 6.8% increase, driven primarily by the Asia-Pacific market which contributed $742 billion to the total...
This data is completely fabricated. The AI recognized the *pattern* of a financial report and "completed" the numbers with plausible-sounding values. It doesn't actually know the 2024 numbers!
📝 Key Concepts
- Confident Guesses: If you ask an LLM for a specific data point it hasn’t seen (like fashion revenue for a future year), it won’t just say “I don’t know.” It recognizes the pattern of a financial report and fills in the blanks with statistically plausible-sounding numbers.
- No Internal Fact-Checker: LLMs generate text based on patterns, not by looking up facts in a database. They have no internal mechanism to verify if a claim is true or false before saying it.
- It’s a Feature: The exact mechanism that allows an LLM to write a beautiful poem about a robot on Mars (creativity) is the same mechanism that makes it invent a fake legal case. It’s pattern-matching all the way down.
For factual data like revenue figures, market statistics, or recent events — use the Deep Research or web search tool if available. It fetches real, current information from the internet instead of pattern-guessing. When in doubt: look it up, don’t ask the LLM to recall it. Later we’ll explore a technique called RAG (Retrieval-Augmented Generation) that does exactly this — automatically fetching real documents to ground the LLM’s answers in facts rather than guesses.
LLM engineers actually struggle to make models reliably say “I don’t know.” Because LLMs are built to always output something, training them to stop and admit ignorance requires intensive, specialized work.