The Biggest AI Mistake I Ever Made (So You Don't Have To)

Everyone who uses AI makes this mistake eventually. I made it when I was ten years old, and it was embarrassing enough that I still think about it. Here's exactly what happened.

I'm Parikshet. I teach AI to kids, I hold a prompting certification, and I made a fundamental AI mistake that I should have known to avoid. I'm telling you about it so you can avoid it too.

What Happened

I was preparing content for KidsFunLearnClub — a lesson about how AI is used in a specific field. I used ChatGPT to research some background facts. The AI gave me a specific statistic with a clear, credible-sounding source: a named organisation, a specific year, a precise number. It looked exactly like the kind of cited fact you'd find in a legitimate article.

I didn't verify it. I was in a hurry. The source sounded real. The number fit the point I was making. I included it in the lesson.

A few weeks later, someone reading the content looked up the source. It didn't exist. The organisation was real, but the specific report the AI cited — and therefore the specific statistic — didn't exist anywhere. ChatGPT had invented a plausible-sounding citation complete with an accurate-sounding title and year.

I had to go back and fix the lesson. Embarrassing. Easily preventable. My fault.

Why This Happens (The Technical Reason)

This type of error — called a citation hallucination — is one of the most common and dangerous AI failures. Large language models don't retrieve information from a database; they generate text that is statistically likely given the context. When a model is producing a fact-heavy paragraph, it generates what citations typically look like — a credible-sounding organisation, a specific year, a plausible title — because that's what appears in academic and journalistic writing in its training data.

The model isn't lying. It doesn't know the citation doesn't exist. It produces what a citation should look like in that context, and that can be entirely fabricated.

This is why never trust AI citations is one of the most important rules in AI use. Not "usually verify" — never trust without verification.

The Rule I Follow Now

After that mistake, I created a simple rule that I've followed ever since: if the AI gives me a specific named source, I look it up before I use it. Not "I check if it sounds right." I look it up — open a browser, search the organisation's name plus the report title, verify it exists, verify the statistic appears in it.

This adds maybe two minutes to anything I'd use a citation for. The alternative is the risk of including a fictional fact in content that other people are going to read and trust. That risk is not worth two minutes.

What This Taught Me About AI More Broadly

The mistake was a useful lesson about a principle I'd intellectually understood but hadn't felt the consequences of: AI's confidence is entirely disconnected from its accuracy. I knew this. I'd read about it. I'd taught it. But I'd never felt the cost of ignoring it.

Now I feel it every time I'm tempted to skip verification on something that sounds right. The memory of having to go back and fix a lesson because I trusted a fabricated citation is a better teacher than any explanation of what hallucination is.

Make the mistake once. Learn the lesson. Don't make it twice in a context that matters more.

Frequently Asked Questions

What is a citation hallucination?

When AI generates a plausible-sounding but entirely fabricated reference — a real-sounding organisation name, a credible-sounding report title, a specific year — that doesn't actually exist.

How do you verify AI citations?

Search the organisation's name and report title directly. Confirm the organisation exists and has published the specific report. Find the actual statistic in the actual document. If you can't find it, don't use it.

Does AI know it's making up citations?

No. AI has no internal truth-checker. It generates what citations statistically look like in its training data. The fabrication is not intentional — the model simply cannot distinguish between things that exist and plausible-sounding things that don't.