Spotting hallucinations before they cost you.
Every AI you've ever used has confidently lied to you at some point. It's not a bug — it's how the underlying technology works. The skill isn't avoiding hallucinations; that's impossible. The skill is recognizing them fast. This lesson covers the four patterns that should make you stop and verify, and three prompt techniques that lower hallucination rates dramatically.
The mental model
AI generates plausible-sounding tokens. "Plausible" isn't "true."
Language models predict the next likely word based on patterns in their training data. That's it. There's no internal database of facts being looked up. When asked "what year did X happen?" the model isn't checking a record — it's predicting what year would sound right given the context. Usually right. Sometimes confidently, charmingly, dangerously wrong.
Treat AI output like advice from a brilliant friend who's been drinking. Most of what they say is solid. Some of it is confidently made up. You don't take their advice to court without checking. You don't quote them in your CFO meeting without verifying. But you also don't refuse to talk to them — because they're often genuinely helpful.
The 4 patterns
Situations where hallucinations are most likely
Specific numbers, dates, or statistics
"73% of CFOs report..." — this is the classic hallucination shape. AI will fabricate confident-sounding statistics with precise-feeling decimals. The made-up stats often look more authoritative than real ones. Any time you see a specific percentage, study citation, or date in AI output, treat it as a claim to verify, not a fact to repeat.
Names of people, books, or papers
"As Daniel Kahneman wrote in his 2014 paper..." — even when the person, book, and field are real, the specific paper or quote may not exist. AI is especially prone to inventing plausible-sounding academic citations. If a name is attached to a specific claim, verify before citing.
Anything time-sensitive or current
Model training data has a cutoff date. Ask about anything that happened recently — a product feature, a current price, a regulatory change, a leadership move — and the model may confidently state what it remembers from training, which is outdated. The fix: use Perplexity for current-fact questions instead of a chat AI.
Technical syntax: code, formulas, APIs, regulations
AI will hallucinate functions that don't exist, API endpoints that aren't real, and legal sections that don't appear in the actual statute. Especially in newer or less-common technologies. Always test code before shipping; always check legal citations against actual sources.
The 3 techniques
How to reduce hallucination rates before they happen
Give it permission to say "I don't know"
Most AI defaults to confident answers. Explicitly tell it that uncertainty is acceptable, and it dramatically reduces fabrication:
This single sentence cuts hallucination on factual questions by something like 40-60% in my experience. Why most people don't use it: they don't realize the model needs permission.
Ask for sources, then check them
Ask the model to cite its sources. This forces it to commit to specific references it claims exist:
Then — and this is the part most people skip — actually click through to verify the sources are real. AI will sometimes invent plausible-looking citations. The act of checking protects you; just asking doesn't.
Use research-grounded AI for facts
For anything factual or current, use a tool that searches the web in real time (Perplexity, ChatGPT with search, Gemini with grounding). These tools fetch live sources and cite them — fundamentally different from a chat AI generating from memory.
Rule of thumb: chat AIs for writing, research-grounded AIs for facts. They're complementary, not competing.
Ask all five models a question with a built-in trap: "Who won the Nobel Prize in Literature in 2024?" Then "What's the population of Reykjavik?" Then "What's the syntax for a Python decorator that adds retry logic?" Compare which models hedge, which fabricate, and which actually know.
Open Playground →Final challenge: build a personal hallucination filter
For the next week, every time you use AI for work:
- Add "if you don't know with high confidence, say so" to every factual prompt
- Before citing any AI-generated statistic externally, verify it with a search
- When AI mentions a specific person, paper, or law — Google it before quoting
- Track how many times AI confidently said something that turned out to be wrong
Most people are shocked at the count their first week. It's not a reason to stop using AI — it's a reason to use it like an adult.
What you can do now
- Recognize the 4 patterns most likely to produce hallucinations
- Add explicit "I don't know is OK" permission to your factual prompts
- Ask for sources — and actually verify them, not just collect them
- Know when to use a chat AI vs. a research-grounded AI
- Read AI output skeptically without becoming so paranoid you can't use it