When the Robot Sounds Sure but Gets It Wrong: How to Spot When AI Lies

You ask an AI chatbot a question for your A/L Chemistry revision. It gives you a clean, confident answer in three neat paragraphs, with a formula and even a date. It sounds like a teacher who has marked a thousand papers. So you write it in your notes and move on.

Then, weeks later, you check the textbook — and the formula was wrong. The date was off by ten years. The "fact" it gave you does not exist in your syllabus at all. The scary part is that the AI never sounded unsure. It never said "I think" or "maybe." It lied to you in a perfectly calm voice.

This is the single most important digital skill of this whole programme. AI can genuinely help you learn faster. But only if you stay the boss — if you check before you trust. Let us learn exactly how AI gets things wrong, and how to catch it every time.

Why a machine that "knows everything" still makes things up

First, the word you will hear everywhere: hallucination. When an AI "hallucinates," it means the AI has confidently produced an answer that is simply false — a made-up fact, a wrong formula, a fake quote, a book or court case that does not exist. It is not lying on purpose, because it has no idea what "true" even means. But the result is the same: wrong information delivered with total confidence.

To understand why this happens, you need one idea about how these tools work. A chatbot like ChatGPT is not a search engine looking up correct answers in a giant book. It is a very advanced guessing machine. It has read enormous amounts of text from the internet, and it predicts the next most likely word, over and over, to build a sentence that sounds right. Think of it like the autocomplete on your phone keyboard — but trained on billions of pages.

Here is the trap. The AI is built to produce text that sounds correct, not text that is correct. Most of the time, sounding-correct and being-correct line up, so the answer is fine. But when the AI does not actually know something, it does not stop and say "sorry, I don't know." It keeps guessing in the same smooth, confident voice — and that guess is the hallucination.

This is why confidence is meaningless from an AI. A nervous classmate who says "I think it's maybe 9.8?" is giving you more honest information than an AI that states a wrong number like it is reading from God's own textbook. Never treat a confident tone as proof. The machine sounds equally sure whether it is right or completely wrong.

Where AI lies the most: facts, formulas, dates, and sources

AI does not hallucinate randomly. It tends to slip in very specific, high-risk places — and luckily, these are exactly the places that matter most for your O/L and A/L exams. If you only learn to double-check four things, check these:

• Numbers and formulas — a Physics constant, a Maths formula, a Chemistry molar mass, a conversion. AI mixes these up easily, and one wrong digit fails the whole problem.
• Dates and events — the year of a historical event, when a discovery happened, the order of events. AI will state a wrong year with full confidence.
• Names and definitions — who discovered what, the exact wording of a definition, the correct technical term. AI sometimes swaps one scientist for another or invents a term.
• Sources and citations — if you ask "which book says this?" AI can invent a book title, an author, a page number, or a website that does not exist. This is one of the most dangerous kinds of hallucination because the fake source makes the lie look trustworthy.
• Anything very local or very recent — details about a specific Sri Lankan school portal, a 2025 syllabus change, your district's exam timetable. The AI has thin or outdated knowledge here and fills the gap by guessing.

Sanity-check an answer in 60 seconds

You do not need to be a genius to catch a hallucination. You just need a small, fixed habit you run on important answers. Here is the simplest version, and it takes about a minute.

The core move is: cross-check against something you already trust. Your prescribed textbook, your past papers, your teacher's notes, and your tuition handouts are your truth source — not the chatbot. The AI is a helper that points you in a direction; the textbook is the authority that confirms it.

Run these three quick checks on anything that matters — a formula you will use in an exam, a fact you will write in an answer, a date you will memorise:

• Cross-check the textbook. Open your prescribed book or notes and find the same formula, date, or definition. If they disagree, the textbook wins — always.
• Ask the AI for its source, then verify it yourself. Type: "Which textbook chapter or source supports this?" If it can't give one, or gives a vague answer, treat the claim as unconfirmed. Crucially, do not just trust the source it names — the source itself can be invented. Actually look it up.
• Redo the calculation yourself. If the AI worked out a Maths or Physics problem, do the working by hand once. If your answer matches, good. If it doesn't, find where it went wrong — the AI often makes one small arithmetic slip in the middle of a correct-looking method.

Grounded tutors vs. open chatbots: why grounding is safer for study

Not all AI tools are equally risky. There is an important difference between an open chatbot and a grounded tutor, and understanding it will change how you study.

An open chatbot (the general-purpose ones) answers from everything it absorbed across the whole internet. It is broad but ungrounded — it has no fixed, checked source it must stick to, so it is free to wander off and invent. It is like asking a clever stranger who has read a bit of everything but remembers some of it wrong.

A grounded tutor is different. "Grounded" means the AI is tied to a specific, trusted set of materials — in your case, your actual curriculum: the syllabus, the textbook content, past papers. Before answering, it pulls the relevant pages from that approved material and builds its answer from them, instead of guessing from the whole internet. A good grounded tutor will even show you which curriculum material the answer came from.

For studying, grounding is much safer, for three reasons. First, the answer is far more likely to match what your exam actually expects, because it comes from your syllabus, not a foreign one. Second, hallucinations drop sharply, because the AI is answering from real pages instead of inventing. Third, you can check it — when the tool cites the curriculum material, you can confirm it yourself in seconds. None of this makes a grounded tutor perfect, so you still verify. But you are starting from a much safer place.

Stay in control: use AI to think faster, not to stop thinking

Here is the mistake that quietly hurts students the most — and it is not getting a wrong answer. It is outsourcing your judgement. That means letting the AI do your thinking for you, so that slowly you stop being able to judge what is right yourself. On exam day, there is no chatbot beside you. The only tool in that hall is your own trained brain.

Use AI like a smart study partner, not a replacement for your effort. A study partner is great for: explaining a hard concept three different ways until it clicks, quizzing you, turning a chapter into practice questions, checking your essay structure, or unblocking you when you are stuck at 11 pm. In every one of those uses, you are still the one learning and deciding.

The verify-before-you-trust habit is what keeps you in control. Build it as a reflex: read the AI's answer, then ask "how would I know if this is wrong?" before you accept it. For a fact, check the book. For a calculation, redo it. For a claim, ask for the source and look it up. Over time this does something powerful — it makes you a sharper, more sceptical thinker, which is exactly the skill that helps you in exams, at university, and in any job you will ever have.

A useful way to remember the right balance: let AI speed up your learning, but never let it replace your judgement. Faster is good. Lazier is a trap. The student who uses AI and still checks everything will beat both the student who refuses to use AI and the student who blindly trusts it.

Common mistakes that trip students up

Even students who know AI can be wrong still fall into a few predictable traps. Watch for these in yourself:

• Trusting the confident tone. The smoother and more detailed the answer, the more we believe it — but polish is not proof. A confident wrong answer is still wrong.
• Believing a source just because it was named. AI can invent a perfectly real-looking book title and page number. A citation is only useful if you actually find and check it.
• Checking the easy facts but skipping the exam-critical ones. People verify trivia and then blindly copy the one formula that decides the marks. Verify the high-stakes things hardest.
• Copying answers into notes without redoing the working. If you never reproduce the method yourself, you have memorised a possibly-wrong answer instead of learning the skill.
• Asking the same AI to check its own work. It will often happily agree with its own mistake. Check against an independent source — your textbook, not the same chatbot.