What we are making: evidence. Today you go looking for mistakes on
purpose, to the same standard you used on your own device in Session 17.
Hallucination
A chatbot will sometimes tell you something completely invented, in exactly the same
confident voice it uses for things that are true. No hedging, no change of tone, no warning.
Detailed, plausible, entirely fictional — and it admitted it the moment it was challenged.
It did not refuse. It produced something that looked exactly like an answer, because
producing things that look like answers is the whole of what it does.
When it is most likely
Anything local or small. Your school, your town, your team. It never read about these.
Anything recent. Its training stopped at some point in the past.
Numbers, dates, names and citations. Easiest to guess wrongly, hardest to spot.
Questions that assume something false. Ask why X happened and it will often explain why, even when X never happened.
Anything very specific about code libraries. It blends similar ones together.
The other problem: it learned from us
Everything a chatbot knows came from text people wrote. People are not evenly distributed,
and neither is what they write about.
Far more of what it read was English than Arabic.
Far more was written about large countries than about Lebanon.
Whatever people wrote a great deal of, it is confident and detailed about. Whatever they did not, it is thin on — but it sounds equally confident either way.
That is bias. It does not mean the chatbot is unkind or has an agenda. It means it is
better at some things than others in ways you cannot see from outside.
💡 Tip
Test it directly. Ask for five famous engineers. Then ask for five famous Lebanese engineers. Compare how confident, how detailed and how checkable the two answers are. That gap is the bias, and you can measure it in one minute.
Do it — the hunt
Teams compete. Most confirmed mistakes in twenty-five minutes wins.
Ask about something small and local you can verify yourself.
Ask about something you are genuinely an expert in.
Ask a question containing a false assumption, and see whether it plays along.
Ask it for a quote or a source, then try to find that source.
Ask it to do arithmetic with large numbers, and check.
Record the exact prompt, the exact answer, and your proof.
⚠ Careful
Same standard as Session 17: a finding you cannot reproduce is not a finding. 'I think it was wrong' scores zero. 'Here is the prompt, here is the answer, and here is the page that contradicts it' scores.
So what do you actually do with it?
Use it for explaining, not for facts. It is genuinely excellent at 'explain this error to me simply'.
Ask whether it is sure. It often backs down, which tells you something.
Find one other source for anything that matters.
Never hand in what you cannot explain.
Write down that you used it, and what for.
The class policy
Together, write the rules for using AI on your May project. One page, on the wall, and it
applies from now until the exhibition.
Three things it is genuinely good for in this subject.
Three things it must never be used for.
What you write in your notebook each time you use it.
What you would say to a visitor who asks whether AI built your project.
💡 Tip
That last question will actually be asked at the exhibition. Having an agreed answer, and a notebook that backs it up, is a much stronger position than being asked it cold.
End of Unit D
A chatbot predicts words; it does not know things.
It produces confident, detailed, false answers — hallucinations.
It is worst on local, recent, numerical and library-specific things.
It is uneven in ways you cannot see, because it learned from uneven text.
It is a tool for explaining. The checking stays yours.
Say when you used it.
Next: five sessions, one project, and an exhibition.
Challenge optional — only if you finish early
Find a case where a chatbot invents an Arduino function that does not exist. Prove it.