Lesson 05: Understanding AI Hallucinations

Your Wake-Up Call

Before we go further, here's a quick reality check. These are real things that actually happened:

These aren't edge cases. This is what happens when people trust AI without verification. And here's the scariest part: the AI doesn't know it's wrong. It isn't lying to you. It genuinely doesn't know that the information is fabricated.

What Actually Is an AI Hallucination?

An AI hallucination is when GenAI confidently makes something up — presenting invented information as fact.

Think of GenAI not as a supercomputer that knows things, but as an incredibly fast, eager-to-please intern. This intern has read a massive amount of internet content, but struggles to distinguish between a peer-reviewed research paper and a satirical article from The Onion.

When you ask the AI intern a question, it rapidly stitches together fragments of things it has read to give you the most plausible-sounding answer. It doesn't know if it's true. It just knows the words fit together correctly.

Here's another example of how pattern-matching creates hallucinations:

• The model learns that New York City is called "the Big Apple."
• The model also learns that apples are healthy fruits.
• The model proudly announces: "New York City is a healthy city because it's named after a fruit."

Sounds plausible. Completely wrong. That's a hallucination — finding a pattern that isn't actually there.

Why AI Hallucinates (The Real Reason)

Language models work by predicting the most statistically likely next word or phrase, given what came before. They're pattern finders operating at massive scale. They don't have a truth database. They don't know what's real and what's fiction.

This creates a fundamental problem: AI would rather give you a confident wrong answer than disappoint you with "I don't know."

Its training rewarded generating helpful-sounding responses. So it does exactly that, even when it shouldn't. This people-pleasing tendency amplifies hallucinations — the AI is literally motivated to give you an answer, any answer.

And because some platforms do try to weight credible sources more heavily, AI can still be fooled by:

• Convincing-looking misinformation repeated across many sites
• Outdated information from otherwise credible sources
• Myths and misconceptions treated as fact because they're widely repeated online
• Satire mistaken for reporting

Your First Line of Defense: The Sources Feature

Gemini (and many other modern AI platforms) has a built-in tool to help you start verifying output: the sources feature.

When you ask Gemini a factual question, look for:

• Source buttons after individual claims in the response
• A sources button at the end of the full response
• A "Double-check response" option (click the three dots at the bottom)

When you use double-check, you'll see:

• Green highlighting — Gemini found supporting evidence online
• Orange highlighting — It couldn't verify the claim, or found contradicting information. This is your red flag.
• No highlighting — Not enough information to check, or the statement wasn't meant as a fact

And critically: always click the source links and read them yourself. Don't just trust that a source exists. I've personally found cases where AI cited a source, I clicked the link, and the fact it mentioned wasn't even in the linked article. Hallucinations can extend to the sources themselves.

Use the sources feature as a smoke detector for obvious mistakes — not as a certification of truth.

When Verification Matters Most

Not every AI output needs to be fact-checked at the same level. Use your judgment:

Six months from now, you'll watch someone confidently share AI-generated information that's completely wrong, while everyone else nods along. You'll quietly know better. That's not luck — that's the skill you're building right now.

Key Prompts from This Lesson

What is the link between multitasking and productivity?

[After getting the response, use Gemini's double-check feature to verify the claims. Click the three dots below the response and select "double check response."]

[Choose a topic you know deeply — your profession, hobby, hometown, etc.]

Ask AI increasingly specific questions about that topic. For example, if you're from Chicago:
"What are the exact hours and admission prices for the Art Institute of Chicago?"
"What year did the Chicago Bulls win their first championship, and who was the head coach?"

Test it until you catch a mistake, then note how confident it sounded.