When AI gets it wrong
Learn where AI makes things up, which facts to always check, and how web search and stronger models help.
Day 5 of Prompting Path, a 30-day AI course with one 10-minute lesson and one real task a day. Day 0 to Day 8 are free.
The lesson
AI confidently makes things up. This is called hallucination, and you need to know when to trust and when to verify.
Higher-risk zones (always verify):
- Specific numbers, dates, statistics, citations.
- Names of people, companies, papers, links.
- Anything legal, medical, financial.
Lower-risk zones: brainstorming, drafting, summarizing things you've pasted in, formatting.
Hallucinations are close to impossible to eliminate entirely, but two things shrink them a lot:
- For high-stakes work, use a brainier model. The heavyweight options, Claude Opus or ChatGPT's 5.5 Thinking mode, make fewer things up. Slower, but that's the trade. (More on picking models in Week 3.)
- Much easier: turn on web search. Most tools have a toggle for it. It forces the AI to go away and verify facts before presenting them, instead of answering from memory.
Here's the bit that surprises people: AI models are trained up to a point in time, and then the training stops. That cutoff can be three, four, five months ago. Anything that happened after it simply isn't in the model's head, so ask about last month's news without web search and it will either admit it doesn't know or, worse, guess confidently. Web search is what bridges that gap.
Either way, your job with AI is editor, not typist. The editor's first question is always "is this actually true?"
Today's task is in the app
Every day ends with a small, real task and a tip, and they live in the app. Pop in your email once and every free day opens up, with your progress saved as you go. It works with Claude, ChatGPT, Gemini or Microsoft Copilot, and the free tiers are enough.
Open Day 5's task → See all 30 days