Course A · Module A2 Prompting · Lesson 6 of 17
Confident doesn't mean correct. Learn to design prompts that catch wrong answers before they leave your desk.
The forgetful intern from Lesson 1 has one dangerous habit: when they don't know something, they don't say so. They produce a fluent, professional-sounding answer that's simply wrong. This is called hallucination, and it costs real money — a wrong statistic in a client proposal, a fake legal citation, a made-up competitor fact. Your single win today: by the end you'll have four concrete moves that stop confident-but-wrong answers before they cause damage.
Imagine your brilliant intern studied intensively by reading millions of documents — but they did it two years ago and haven't checked the internet since. When you ask for a specific number or a recent event, they don't say "I'm not sure." They pick the most plausible-sounding answer from memory and state it as fact. They genuinely don't know they're wrong.
That's not a bug the company forgot to fix. It's how the technology works — pattern completion, not fact lookup. Knowing that, you design around it.
Instead of asking the AI to recall a fact, paste the source into the prompt. "Here is our pricing table: [paste]. Now draft the proposal." It can only hallucinate facts it has to invent — give it the facts and there's nothing to invent.
Add to your prompt: "Quote directly from the document I provided. Do not add information from outside it." When you can see the quote, you can verify the quote. Unverifiable claims disappear.
Treat every specific claim — a statistic, a date, a law reference, a competitor detail — as a draft that needs one lookup. A quick search takes thirty seconds. Sending a wrong number to a client can cost much more.
For legal advice, medical guidance, financial projections, or anything where a wrong answer has real consequences: use AI to draft structure and language, then hand the substance to a qualified expert. Draft helper — not expert substitute.
Answer before opening. Recalling it now is what makes it stick.
Verify each one before publishing. Specific numbers the model generated from memory are high-risk hallucination targets. Find the original source for all three, or remove the claims you can't verify.
Paste the actual product facts into the prompt. The model can only make up facts it has to invent. Paste your spec sheet, pricing table, or feature list directly into the context — it then reflects what you gave it, not what it guessed.
No — use it for structure, not substance. AI can draft the language and format, but legal text is exactly where a wrong-but-confident answer is catastrophic. Have a lawyer review the substance. "Draft helper, not expert substitute" is the rule.
Read this next (primary source)
What Is Prompt Engineering? — Google Cloud. The section on grounding and reducing hallucinations maps directly to the "feed it the source" and "ask it to cite" moves above. ~8 min. See RESOURCES.md for more.
The words hallucination, context, and prompt are in the course glossary. Keep it open.