Big jackpot. €50 million — capped, pot absolutely full. And I, someone who works with AI systems professionally every day, type this: "Generate 8 lottery tickets for a 6-from-49 draw, no repeated numbers."
The AI delivers:
1 – 9 – 17 – 25 – 33 – 41
2 – 10 – 18 – 26 – 34 – 42
3 – 11 – 19 – 27 – 35 – 43First thought: great, done. Second thought, three seconds later: hold on.
That's wrong.
Shouldn't have happened to me. Did anyway. And that's exactly why I'm writing this: because this mistake doesn't mean you don't know the tool — it means you know it well enough to stop reading every prompt twice. Welcome to the club.
Why "technically correct" sometimes isn't enough
No number appears twice — true. All picks are within the valid range — true. The AI delivered exactly what you typed.
The problem lies elsewhere. Lotteries aren't just about probability — they're about payout. Every combination has the same jackpot odds: roughly 1 in 140 million (including the bonus digit). No prompt changes that. What differs: how many other players picked the same combination. Jackpots are split between everyone who matched — the more people share your numbers, the smaller your slice.
Systematic patterns are popular. And anyone who typed the same prompt as you got the exact same tickets. The more people play those tickets, the lower your expected payout if you win. The statistical expected value drops — not because of the win probability, but because of jackpot splitting.
The better prompt would have been: "Generate 8 random lottery tickets for a 6-from-49 draw, no repeated numbers and no recognisable patterns."
Four extra words. Completely different result.
Required disclaimer, which we as tech people feel slightly embarrassed about needing: lottery is gambling. Statistically, every euro spent returns about 50 cents in expected value — no matter how clever the prompt. We know this. We still play sometimes when the jackpot is high enough and the "what if" feeling seems worth the price.
Example two — when technically correct becomes dangerous
Slightly more technical, same pattern.
Prompt: "Build me a REST API that stores and retrieves customer data — names, email addresses, orders."
The AI delivers: clean endpoints, JSON responses, database connection, error handling. Everything there. It runs.
What's missing: authentication. No login required. Anyone who knows the URL — or just tries GET /customers — can pull all customer names, all email addresses, all orders. Completely. Without a password.
The AI built what was asked for. An API. Security wasn't mentioned in the prompt, so none was added. Why would it be?
That's not a mistake. That's obedience.
The pattern
Both examples follow the same logic: AI doesn't necessarily add what you take for granted. Sometimes you get lucky and it fills in the blanks sensibly. Often it doesn't — as these examples show.
You know lottery numbers should be random. The AI might know that too — but it doesn't have to act on it unless you ask. You know an API should be secured. It builds an open one anyway, because you didn't mention authentication.
This applies everywhere. Say "write me an email to my boss" and you get something — sometimes even something usable. Say "write me a short, polite email to my boss — he doesn't know the background, the tone should be direct but not pushy" — and you get something reliably usable.
AI models are very good at delivering what the prompt says. Relying on unspoken expectations is a gamble.
What helps
Name what you take for granted. That's usually where the gap is.
Three questions that almost always help:
- What should the result not do or contain?
- What constraints apply that aren't obvious from the task?
- What's the actual purpose behind the task — not just the task itself?
For the fundamentals of good prompting: What is a Prompt? — start there.
Then: just try it. Play the lottery? Better not start — the house always wins in the end. Problem gambling — Bundeszentrale für gesundheitliche Aufklärung