Hallucination
In Malay: Halusinasi
An AI hallucination happens when an AI model produces an answer that sounds confident, fluent, and authoritative, but contains false or completely fabricated details — such as invented statistics, non-existent court cases, wrong dates, or broken web links. Understanding hallucination is the single most important safety habit for anyone using AI at work or school.
Example: The AI quotes a "2023 Universiti Malaya study" that never actually existed.
Why do AI models hallucinate in the first place?
People often wonder why a cutting-edge AI can invent a fake fact when a basic pocket calculator never gets a sum wrong. The reason comes down to how an LLM is built: at its core, it is trained to predict the most plausible-sounding sequence of words, not to query a verified database of truth.
When you ask about a narrow topic where the model's training memory is thin — such as a specific Malaysian regulation or a local university paper — the model still tries to generate a helpful-sounding completion. Because it masters the tone of formal writing so well, a guessed answer reads with the exact same confidence as a true one.
Real-world situations where hallucinations cause trouble
In Malaysian workplaces and classrooms, hallucinations tend to catch people out in four classic scenarios:
- Academic citations: A student asks an AI for five journal references on a local topic, and the model blends real authors' names with plausible-sounding paper titles that were never published.
- Legal and policy clauses: Asking a chatbot to quote specific Malaysian statutory sections or government grant rules from memory.
- Numbers across long documents: Summarising a dense report where the model accidentally mixes figures from two separate tables.
- Changing local facts: Asking about current pricing, office hours, or application steps without live web search.
4 practical habits to cut down hallucinations
You do not need to avoid AI because of hallucinations — you just need to anchor the model to real source material (a practice known as grounding):
- Feed the source text directly (RAG style): Instead of asking from memory, upload your PDF or paste the notes into the prompt and write "Answer only using the facts in the text below."
- Give the AI permission to say 'I don't know': Add this line to your prompt: "If the answer is not stated in the provided text, write 'Not stated in the source' instead of guessing."
- Use web search for current facts: Turn on web search mode and click the cited links yourself to confirm the numbers match.
- Ask for exact quotes: When summarising a long contract or report, tell the AI to include the exact short quote or section heading next to each claim.
The golden rule: when human verification is non-negotiable
My rule of thumb is simple: if you are brainstorming creative angles, polishing your own rough draft, or writing fiction, the model's imagination is a feature. The moment your output touches money, contracts, laws, health, real people's reputations, or official policy, treat the AI as a draft assistant only and verify every factual claim against the primary source.
Common questions
Do paid or newer AI models still hallucinate?
Yes. Newer models and reasoning models hallucinate significantly less often than older versions, but no LLM is 100% immune. Providing source documents and checking important facts remain essential habits.
Why doesn't the AI just admit when it doesn't know something?
By default, language models are trained to be helpful and complete the pattern of your request. Unless you explicitly instruct the model to flag uncertainty or stick strictly to an uploaded document, it may fill gaps with its best statistical guess.
How do RAG and grounding help stop hallucinations?
Grounding and RAG (Retrieval-Augmented Generation) force the AI to look up relevant passages from your real documents or live web search first, then write its answer based on those retrieved facts rather than relying on memory alone.
Want copy-ready prompt tips that make AI flag uncertain claims automatically? Check out /tips/ or read the practical guides at /belajar/.