The knowledge check: what has really stuck

All questions with answers

1. What does a language model do at its core?

   • It understands text roughly the way a person does.
   • Building on vast amounts of text, it predicts the most plausible next word.
   • It looks the answer up in a verified database.
   Show the answer

   Correct answer: Building on vast amounts of text, it predicts the most plausible next word.

   That is why even a wrong answer sounds fluent and confident: the model works with plausibility, not with truth. The rules in the next questions follow from this.

2. Which building block is missing from most instructions, and does the most damage when it is?

   • A polite greeting.
   • The quality rule: invent nothing, and ask when information is missing.
   • A request for an especially long answer.
   Show the answer

   Correct answer: The quality rule: invent nothing, and ask when information is missing.

   Without this line the tool answers knowledge gaps with a confident invention. With it, you get honest questions back far more often. It is the most important sentence in the whole course.

3. An angry customer email needs answering. It contains the customer's name and contract details, and you are using a free private account. What applies?

   • Fine, as long as I delete the chat afterwards.
   • Not fine: take names and details out, placeholders are enough, or use a tool your business has approved.
   • Fine, free accounts are anonymous anyway.
   Show the answer

   Correct answer: Not fine: take names and details out, placeholders are enough, or use a tool your business has approved.

   Deleting the chat does not get the data back from the provider, and no account is anonymous. Placeholders such as “Customer A” almost never change the result. The data traffic light practises exactly this.

4. The tool delivers a statistic with a source reference for a customer proposal. What do you do?

   • Use it, a source reference means it is substantiated.
   • Open the source yourself and check that it exists and that the figure is really in it.
   • Ask the tool to confirm the statistic once more.
   Show the answer

   Correct answer: Open the source yourself and check that it exists and that the figure is really in it.

   Language models also invent source references, formally correct-looking ones with plausible titles. And a tool confirming its own answer checks nothing. Figures that leave the business are checked by a person against the primary source.

5. What does Art. 4 of the AI Act require from a business that uses AI?

   • Measures that support the AI literacy of its staff. Nobody has to guarantee a particular level.
   • Reporting the use of AI to the authority.
   • Nothing, the article only concerns large companies.
   Show the answer

   Correct answer: Measures that support the AI literacy of its staff. Nobody has to guarantee a particular level.

   The duty applies regardless of size and sector as soon as AI is used. Since the July 2026 rewording, in force since 27 July, no particular level of literacy is owed. A business that records its measures can evidence them later.

6. You suspect that AI tools are being used privately in your team. What is most likely to bring clarity?

   • A short anonymous survey: who uses what, and for which tasks?
   • A clear ban until a policy is in place.
   • Waiting until somebody mentions it on their own.
   Show the answer

   Correct answer: A short anonymous survey: who uses what, and for which tasks?

   A ban only pushes the use deeper into the shadows, and nobody mentions it on their own. An anonymous survey shows the real picture, because nobody has to fear consequences.

7. You want a contract passage rephrased in plain language. How do you keep the input lean?

   • Replace names, amounts and recognisable details with placeholders before the text goes in.
   • Delete the chat after the answer.
   • Enter an invented company name in your profile.
   Show the answer

   Correct answer: Replace names, amounts and recognisable details with placeholders before the text goes in.

   Deleting takes effect only after the transfer, and a profile name changes nothing in the content. What matters is what is in the text: placeholders such as “Customer A” are almost always enough. Strictly speaking that is pseudonymisation, not anonymisation – the data traffic light explains the difference.

8. What is separating the material in an instruction for – with three quotation marks, for example?

   • The tool can tell instruction and text apart, and does not treat pasted content as an order.
   • It makes the answer shorter.
   • It is a formality without any effect.
   Show the answer

   Correct answer: The tool can tell instruction and text apart, and does not treat pasted content as an order.

   Without a clear separation the model mixes order and material. If a pasted email itself contains a request, in the worst case it gets carried out.

9. For which task is a language model the wrong choice?

   • The binding calculation of a proposal with quantities and prices.
   • The first draft of a reply to a complaint.
   • The summary of a long set of minutes.
   Show the answer

   Correct answer: The binding calculation of a proposal with quantities and prices.

   A draft and a summary are checked by a person afterwards. In a calculation the figure itself is the result, and a language model invents figures in the same confident tone as everything else.

10. What belongs on the single page of a company AI policy first?

    • What is allowed, what is off limits and who to ask when in doubt.
    • A list of every AI tool available worldwide.
    • The full text of the AI Act.
    Show the answer

    Correct answer: What is allowed, what is off limits and who to ask when in doubt.

    A rule only works if it gets read. Three answers cover the everyday cases: allowed, off limits, contact person. Everything else can grow later.