Hallucinations: confident nonsense and how to catch it | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Classroom / Module 01: What AI Actually Is / Chapter 3 Hallucinations: confident nonsense and how to catch it Watch first, then read. Same lesson, your pace. What you will learn – Why hallucination is not a bug but a side effect of prediction – Where hallucinations cluster (and where they’re rare) – Five checks that catch almost all of them Remember Chapter 1: an LLM predicts the most likely next words. Now the uncomfortable consequence: if you ask something the model doesn’t know, it doesn’t stay silent, it predicts what an answer would probably look like. The result is a fluent, confident, completely made-up statement. This is called a hallucination . It isn’t lying, there was no intent to deceive. It’s the machine’s version of bluffing in an exam: writing something , because a blank page feels worse than a wrong page. Where hallucinations live Hallucinations cluster around specific kinds of questions. Learn the map and you know where to walk carefully: High risk Why Example Specific numbers, dates, statistics Plausible digits are easy to generate “The market grew 23.4% in 2023” Citations, sources, URLs It predicts what a citation looks like Perfectly formatted papers that don’t exist Niche people, products, events Rarely in training data, so patterns are thin Biography of a semi-famous person Anything after its training cutoff It must guess recent events Last week’s news, prices, scores Math with many steps It predicts numbers, doesn’t calculate Long division gone subtly wrong Low risk Why Explaining concepts Well-covered patterns (photosynthesis, inflation, recursion) Rewriting, summarising text you provide It stays on the desk (your context window) Code for common problems Millions of examples trained it well Notice the pattern: the more specific and checkable the claim, the higher the risk. And the flip side is your biggest defence: give the model the source material yourself. Summarising a document you pasted is far safer than asking about the document from memory. The five-second verification habit You don’t need to distrust everything, just route high-risk outputs through quick checks: Numbers → find the original source. If AI says “India’s literacy rate is X”, search it once. Thirty seconds. Quotes/citations → click or search the exact phrase. No result = invented. Anything recent → ask two different models (e.g., Kimi and ChatGPT). Agreement isn’t proof, but disagreement is a red flag. Ask for zero confidence theatre: add to your prompt: “If you are not sure, say ‘I am not sure’ instead of guessing.” It helps more than you’d expect. High stakes = human eyes. Money, health, law, grades: never ship an unchecked number. One honest warning: AI is also excellent at sounding certain when wrong. Tone tells you nothing. Only checks tell you something. Try it yourself Ask any AI: What awards did the author [pick a lesser-known book you love] win? Give exact years. Then verify every claim with a web search. Count: correct, wrong, invented. Next, paste the actual Wikipedia page of that author into the chat and ask again. Compare answers, you’ve just proven that giving context beats trusting memory, which is a technique you’ll use for the rest of your life. Key takeaways – Hallucination = plausible-sounding invention, a side effect of next-word prediction. – Risk concentrates in specific numbers, citations, niche topics and post-cutoff facts. – Providing source material in the prompt dramatically reduces hallucination. – Verify numbers, quotes and anything recent; use a second model as cross-check; tone ≠ accuracy. Download the exercise sheet (PDF) Module workbook (PDF) ← Prev: Tokens, context windows and why AI forgets Next: Meet the models: your four-AI stack → Classroom / Module 01: What AI Actually Is / Chapter 3 Hallucinations: confide
Hallucinations: confident nonsense and how to catch it
Written by
in