📎 This article includes 1 downloadable practice file ↓
A “hallucination” is a fluent, confident answer that’s false: an invented statistic, a non-existent section of a law, a fake reference. It happens because the model generates plausible text, not because it looks things up.
High-risk questions
- Exact numbers, dates, rates and thresholds (tax rates change).
- Legal sections and case names.
- References, links and quotes.
- Niche facts about small companies or people.
- Anything after the model’s training cut-off.
Prompts that reduce it
- Give the source: “Answer only from the text below. If it isn’t there, say so.”
- Allow uncertainty: “If you’re not sure, say you don’t know.”
- Ask for evidence: quotes, section numbers, links you can open.
- Use search-enabled tools for current facts, and still click the sources.
Verification routine
- Is this fact important to the decision? If yes, verify.
- Check it against an official source (government site, the actual document, your system).
- Open every reference; fake ones look real.
⚠️ Never file returns, sign contracts or give advice based only on an AI answer about laws, tax rates or deadlines. Confirm with the official source or a professional.
💡 AI is most reliable when it transforms information you give it (summarise, rewrite, extract, classify) and least reliable when it recalls facts on its own.
Practice
Do the four experiments; check whether the references in step 4 really exist.
📎 Practice files for this article
- 📄Lesson 7 prompt worksheetCopy-paste prompts and exercises to try in ChatGPT, Copilot, Gemini or Claude.⬇ TXT · 383 B
Free to use for learning. Files with macros (.bas) are plain text — import them with Alt+F11 → File → Import File, and always test on a copy.
✨ Ask AI about this article
Stuck on a step? Ask a question and the AI answers using this article.
Free · AI can be wrong