Practice test

Answer first, then open the reasoning

AI fundamentals 20%

1. 1. Which best describes how a large language model produces an answer?

   • AIt looks the answer up in a database of verified facts
   • BIt generates text by predicting likely continuations from patterns learned in training
   • CIt follows rules a programmer wrote for each type of question
   • DIt searches the live internet and quotes the top result
   Answer and reasoning

   B. Language models generate text statistically. They do not consult a store of verified facts, which is exactly why a fluent answer can still be false. Some products add search or document retrieval on top, but that is a layer around the model rather than how the model works.

2. 2. Which of these is an example of AI that classifies rather than generates?

   • AA tool that drafts a reply to a customer email
   • BA filter that labels an incoming message as spam or not spam
   • CA tool that creates a product image from a description
   • DA tool that writes a first draft of meeting minutes
   Answer and reasoning

   B. A spam filter sorts inputs into categories. The other three produce new content, which is what generative means. A good deal of workplace AI is still classification and prediction: fraud flags, credit scoring, document routing. Those systems make decisions about people and attract the most governance attention.

How AI works 20%

1. 3. An assistant cites a regulation, with a section number, that does not exist. What is this?

   • AOverfitting
   • BLatency
   • CA hallucination
   • DAn encryption error
   Answer and reasoning

   C. A hallucination is confident, plausible output that is false. Citations, figures, names and dates are where it does the most damage, which is why they are the first things to check and the last things to take on trust.

2. 4. You ask about a rule change announced last month and the assistant describes the old rule. What is the most likely reason?

   • AIts training data stops at a cutoff and it was given nothing current
   • BIt is concealing the new rule
   • CThe prompt was too short to be understood
   • DThese systems cannot process regulatory text
   Answer and reasoning

   A. Models learn from data collected up to a cutoff date. Unless the tool retrieves current sources, it cannot know what changed afterwards, and it will not say so. The fix is to supply the current document yourself or use a tool that searches and cites.

Generative AI and everyday use 20%

1. 5. You need a one-page summary of a twenty-page supplier contract for a finance manager. Which instruction is most likely to produce something usable?

   • A"Summarise this."
   • B"Is this a good contract?"
   • C"Make this shorter and simpler."
   • D"Summarise this contract for a finance manager in under 300 words. List payment terms, price-change clauses, termination rights and penalties, with the clause number for each."
   Answer and reasoning

   D. The fourth states the audience, the length, the format and what matters. Asking for clause numbers does something the others do not: it makes the summary checkable against the source, which is where most of the value sits.

2. 6. You paste 400 invoice lines into an assistant and ask for the total. What is the sound practice?

   • ATrust the total, since arithmetic is straightforward
   • BAsk twice and use whichever answer appears more often
   • CCompute the total in a spreadsheet or ledger, and use the assistant at most to build or explain the formula
   • DSplit the list into smaller batches so the tool cannot make a mistake
   Answer and reasoning

   C. Language models mis-add and drop rows while sounding entirely certain. Use a deterministic tool for arithmetic and tie the figure back to the source. Smaller batches reduce the error rate without removing it, and asking twice measures consistency rather than correctness.

Risks, limits and responsible use 25%

1. 7. A colleague pastes a customer's full name, identification number and bank details into a free public assistant to draft a letter. What is the main problem?

   • AThe letter will be poorly written
   • BPersonal data has been disclosed to a third party that may retain it, which is likely to breach both policy and data protection law
   • CThese tools cannot write formal letters
   • DNothing, provided the letter is deleted afterwards
   Answer and reasoning

   B. Once pasted, the data has left the organisation’s control. Deleting the output does not delete the input. The answer is an approved enterprise deployment, or removing the identifying details before drafting.

2. 8. A screening tool consistently ranks applicants from one group lower, though their qualifications are comparable. What is the most likely cause?

   • AThe model learned patterns from historical hiring data that reflected past bias
   • BThose applicants used the wrong file format
   • CThese tools are neutral, so the ranking is probably accurate
   • DThe tool is running too slowly
   Answer and reasoning

   A. Models reproduce the patterns in their training data, including the unfair ones. "The system decided" is not a defence. Somebody has to test outcomes across groups and act on what the test shows, and that obligation sits with the organisation deploying the tool.

AI governance awareness 15%

1. 9. Under the EU AI Act, which of these uses is prohibited?

   • ADrafting marketing copy with an assistant
   • BTranslating internal documents
   • CInferring employees' emotions at work, other than for medical or safety reasons
   • DSummarising meeting notes
   Answer and reasoning

   C. Emotion inference in the workplace sits among the practices banned by Article 5, with narrow medical and safety exceptions. The other three are ordinary low-risk uses that the Act does not restrict.

2. 10. Your team drafts a credit recommendation with an AI tool and a manager approves it. The recommendation is wrong. Who is accountable?

   • AThe tool
   • BThe vendor that built it, and nobody else
   • CNobody, because the error came from software
   • DThe organisation, and the people who relied on the output and approved it
   Answer and reasoning

   D. Software cannot hold accountability. The organisation that chose the tool and the person who signed own the decision. A vendor contract may shift some of the cost; it does not shift responsibility towards customers or a regulator.

How the assessment differs from this

Frequently asked