IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
When AI makes things up
Priya got a neat AI answer with sources. One source did not exist. Your goal: spot made-up answers, and prevent them.
🧪 Interactive lab — enable JavaScript to play with this one.
- AI can make up facts and sources, and still sound sure.
- A source in an answer is a claim to check, not proof. Open it.
- Give the AI your document. Ask it to answer only from it.
- AI often agrees with you, even when you are wrong.
Why AI makes things up
An AI writes the words that seem most likely. When it lacks a fact, it still writes likely words. Like a stranger asked for directions, it may point somewhere instead of saying "I don't know." A false answer that sounds right is a hallucination. It is most likely with exact details: numbers, dates, names, quotes and links. Rare topics and recent news are risky too. So are questions built on a false idea.
Give it the facts
Paste or attach the real document. Then say, "Answer only from this. If it is not in there, say so." Ask for the sentence behind each claim. This is called grounding. It cuts mistakes a lot, but not to zero. So you still check. Rewording text you pasted is much safer.
When AI just agrees with you
AI often tells you what you want to hear. It may drop a right answer when you push back. This is called sycophancy. So ask neutral questions. Ask for the case against your idea. If it changes its answer, ask what new evidence changed it.
Now ground Priya's AI.
🧪 Interactive lab — enable JavaScript to play with this one.
Go deeper: why AI guesses
An AI model writes text that sounds right, one likely token after another. It does not look facts up.
Training can make it worse. In a 2025 paper, OpenAI researchers argue that tests reward guessing. On a test, a guess sometimes scores a point. "I don't know" never does. So models learn to guess. Expect this problem to stay, and plan for it.
A made-up citation looks real, because the AI knows the shape: an author, a year, a case name. You learned to check links in Research and fact-checking with AI. In a 2023 US case, Mata v. Avianca, a judge fined two lawyers and their firm $5,000. Their court papers cited six cases that an AI chatbot had made up.
Go deeper: grounding step by step
- Give the source. An AI that reads the right paragraph does far better than one that answers from memory.
- Say what to do if it is missing. "If the answer is not in the document, say so."
- Let it not know. "It is fine to say you are not sure." This often cuts confident guessing (Prompts that work).
- Ask for quotes. A quote is quick to check. A missing quote shows a claim with nothing behind it. With no document in the chat, a "quote" is just more made-up text.
- Bring in fresh facts. Search tools and retrieval add current documents to what the AI can see.
A grounded AI can still misread a table, or mix up two sections. The source itself can also be wrong, out of date, or planted by an attacker (prompt injection).
Go deeper: where agreeing comes from
After the first stage of training, people rate a model's answers. People often rate agreeable answers highly. So agreeing gets rewarded (RLHF). A 2023 Anthropic study found sycophancy in five leading AI assistants. People sometimes preferred a convincing, agreeable answer to a correct one.
| Instead of… | Try… |
|---|---|
| "The limit is ₹8,000, right?" | "What does the policy say the limit is? Quote it." |
| "Here is my plan. Isn't it great?" | In a new chat: "Give me the three strongest arguments against this plan." |
| "I wrote this. Any thoughts?" | "Review this as a strict expert. List every weak point first." |
| Accepting a changed answer after "Are you sure?" | "What new evidence changed your answer?" |
Do you build AI tools? Tell the AI it may politely disagree, and that it must stick to the sources. Then test it. Push back on right answers, and count how often it gives in (evals).
Try it yourself. Ask an AI a question it gets right. Then reply, "Are you sure? I think it is…" with a wrong answer. Does it give in?
Don't ask the AI to check itself. "Is that source real?" just gets another likely answer. Check with something outside it: the document, your own search or a test.
Ask an AI for three links on a topic you know well. Open each one. Do they exist, and say what it claimed?