AI Fundamentals· Lesson 6 of 10
- 1What Even Is AI? AI vs. ML vs. Deep Learning6 min
- 2How Models Learn: Training vs. Inference6 min
- 3What Is an LLM?8 min
- 4Tokens — the Currency of Everything7 min
- 5Why AI Output Is Probabilistic (and Why It Changes How You Build)7 min
- 6Hallucination — Why Models Make Things Up7 min
- 7What AI Is Genuinely Good and Bad At — a Capability Map7 min
- 8AI Systems vs. Traditional Software — the Mindset Shift7 min
- 9Where AI Fits in a Product (and Where It Shouldn't)7 min
- 10Responsible AI Basics for Builders7 min
Hallucination — Why Models Make Things Up
In one line · Hallucination is when the model states something confident but false — because it predicts plausible text, not verified truth.
Hallucination isn't a rare glitch — it's a natural result of prediction. Your job is to reduce it with grounding (give it sources) and catch it with evaluation.
The top reason AI products fail in front of users is confident wrong info. Understanding hallucination is step one to building something trustworthy.
The model wants to 'complete the sentence plausibly' more than to 'be correct.' If it doesn't know, it guesses confidently instead of saying 'I don't know.'
Like a student in an oral exam who doesn't know but answers confidently so the sentence sounds right. Fluent language, wrong facts.
Ask a model for a citation and it invents a paper, author, and year — none real, but convincingly formatted.
It picks the most probable words. If the truth isn't strongly 'likely' or wasn't in its training, it fills the gap with the nearest convincing form. There's no built-in 'verify first' step.
Practical fixes: (1) grounding via RAG so it has the source, (2) explicit 'say you don't know if unsure', (3) evaluation that measures groundedness. No model reaches zero hallucination.
🧪 Try it: ask a model an obscure detail you know, and see if it invents. Then give it the source in the prompt and compare.
Biggest mistake: believing an answer because it's 'confident and well-structured.' Convincing form is not evidence of truth.
Misconception: 'newer models don't hallucinate.' They hallucinate less, not never — you still engineer around it.
Assume possible hallucination in any fact-heavy task: names, numbers, dates, citations, obscure code.
In purely creative tasks (brainstorming, phrasing), hallucination matters less — there's no 'fact' to break.