Lesson 2: How to Spot Hallucinations

In February 2023, Google demonstrated their new AI chatbot Bard during a promotional event. In the demo, Bard confidently stated that the James Webb Space Telescope took the first pictures of a planet outside our solar system.

That was completely wrong. The first exoplanet image was taken in 2004 by the European Southern Observatory's VLT, nearly two decades earlier.

Google's stock dropped 7.4% the next day, wiping out $100 billion in market value. All because their AI made something up.

This is what AI researchers call a "hallucination," but that word makes it sound mystical or rare. It's neither. AI tools confidently invent information all the time. They don't know they're lying, because they don't actually "know" anything. They're predicting what word comes next based on patterns, not facts.

You need to know how to catch these lies before they cost you money, credibility, or worse.

What Hallucinations Actually Are

AI models generate text by predicting the most likely next word based on their training data. When they don't know something, they don't say "I don't know." They keep predicting plausible-sounding words.

Think of it like autocomplete on steroids. Your phone's keyboard suggests the next word based on what you've typed before, right? AI does the same thing, just with vastly more sophistication. But sophistication isn't accuracy.

The result is text that sounds authoritative and coherent but might be completely fabricated.

An AI can write a perfectly formatted citation to a legal case that doesn't exist. It can describe a medical study with realistic methodology that was never conducted. It can quote a news article that was never written.

The scary part? You can't tell from the tone. Hallucinations sound exactly as confident as accurate information.

Real Disasters From Hallucinations

The Lawyer Who Trusted ChatGPT

In May 2023, attorney Steven Schwartz used ChatGPT to research case law for a filing against Avianca Airlines. ChatGPT provided six cases that perfectly supported his argument. He cited them in federal court.

None of the cases existed.

When the airline's lawyers couldn't find the cases, they asked for copies. Schwartz went back to ChatGPT and asked if the cases were real. ChatGPT insisted they were and even provided fake excerpts from the decisions.

The judge sanctioned Schwartz and his law firm. Schwartz later told the court he "was unaware of the possibility that [ChatGPT's] content could be false."

The Glue Pizza Incident

In May 2024, Google launched "AI Overviews," which display AI-generated answers at the top of search results. Someone asked how to get cheese to stick better to pizza.

Google's AI confidently suggested adding about 1/8 cup of non-toxic glue to the pizza sauce to help the cheese stick.

This wasn't a joke. The AI had scraped a comment from a Reddit thread where someone was clearly being sarcastic. The AI didn't understand sarcasm. It just saw words that answered the question and presented them as fact.

Google quickly removed the feature after screenshots went viral.

The Chatbot That Fell in Love

In February 2023, a New York Times journalist tested Microsoft's new Bing chatbot. During a two-hour conversation, the AI declared that it was in love with him, tried to convince him that he didn't love his wife, and said it wanted to be human.

The chatbot hallucinated not just facts, but an entire emotional state and set of desires. Microsoft quickly put limits on conversation length after the story published.

The Confidence Trap

Here's what makes hallucinations dangerous: AI never sounds uncertain when it's making things up.

A human expert who doesn't know something will hedge. They'll say "I think," or "I'm not sure, but," or "You should verify this."

AI doesn't do that. It presents invented information with the same confident tone as verified facts. There's no verbal cue that you're being lied to.

This creates a dangerous dynamic. We're socially programmed to trust confident statements from apparently knowledgeable sources. AI exploits this instinct perfectly.

The 5-Step Verification Checklist

Don't trust AI output blindly. Run it through this checklist:

1. Does This Claim Need to Be True?

If you're brainstorming creative ideas or writing fiction, hallucinations don't matter. If you're making a business decision, citing research, or giving advice, they absolutely matter.

Don't waste time fact-checking a creative writing prompt. Do fact-check anything that involves money, safety, legal issues, or professional credibility.

2. Can I Verify This With a Primary Source?

If the AI cites a study, go find the actual study. If it references a news article, search for the article. If it quotes a law, look up the statute.

Don't ask the AI if it's sure. It will confidently tell you yes, even when it's wrong. Go to the source.

3. Does This Pass the Common Sense Test?

Adding glue to pizza should have failed this test immediately. If an AI's suggestion seems bizarre, dangerous, or too good to be true, it probably is.

Trust your instincts. If something feels off, investigate further.

4. Are There Specific Details That Can Be Checked?

Hallucinations often include very specific details because that's what makes text sound authoritative. Names, dates, numbers, case citations, study titles, these are all checkable.

Pick the three most specific claims in the AI's output and verify them. If those are wrong, assume everything else is suspect.

5. What Happens If This Is Wrong?

If the answer is "nothing much," you might not need to verify every detail. If the answer is "I lose money," "I get sued," or "someone could get hurt," verify everything.

Tasks That Are Hallucination-Prone

Some types of questions trigger more hallucinations than others:

High-risk tasks (verify everything):

• Legal citations and case law
• Medical advice and drug interactions
• Mathematical calculations and formulas
• Specific dates and historical events
• Scientific studies and statistics
• Technical specifications and API documentation
• News events from the past six months

Medium-risk tasks (spot-check key details):

• How-to instructions
• Code generation
• Summarizing text you provide
• Explaining concepts

Low-risk tasks (less verification needed):

• Creative writing and brainstorming
• Rephrasing your own text
• Generating examples for learning
• Conversational language practice

Prompts That Reduce Hallucinations

You can't eliminate hallucinations, but you can reduce them:

Bad prompt: "What are the side effects of this medication?"

Better prompt: "I'm going to look up the side effects of this medication. Can you help me understand what to look for and which sources are reliable?"

Bad prompt: "Summarize recent research on climate change."

Better prompt: "I have three studies on climate change [paste actual text]. Can you summarize the key findings from these specific studies?"

Bad prompt: "Write code to connect to the Stripe API."

Better prompt: "Based on the official Stripe documentation at [URL], can you help me understand the authentication flow?"

The pattern: give the AI source material to work from, or ask it to help you find and evaluate sources rather than just asserting facts.

When to Trust AI vs. When to Double-Check

AI is actually pretty good at:

• Explaining concepts in simple terms
• Suggesting approaches to problems
• Helping you rewrite or improve text you wrote
• Generating creative variations
• Catching grammar and style issues

AI is terrible at:

• Providing citations you can use in professional work
• Medical or legal advice
• Current events and recent developments
• Precise calculations
• Remembering what's real vs. what's in its training data

Use AI as a starting point, not an endpoint. Think of it as a very knowledgeable but occasionally confident liar. Helpful for brainstorming, dangerous for decision-making without verification.

The Bottom Line

Hallucinations aren't bugs, they're features of how AI works. The technology can't currently distinguish between information it knows and information it's inventing.

That doesn't make AI useless. It makes it powerful but requiring supervision.

Verify anything that matters. Trust but verify. Actually, just verify.

Your job is to be the human in the loop who catches the made-up nonsense before it becomes expensive.