Knowledge check

Ten questions on the main ideas of the course. Choose one answer for each, then check your answers. You can try as often as you like.

Your answers are scored in your browser and are not sent to us. Your score is saved only in your browser, where you can clear it. No certificate is issued.

1. Which description best fits how most of today's AI systems work?

   They follow rules that programmers wrote by hand for every case They learn patterns from examples and use those patterns to produce outputs They understand the world in the same way people do They look up the correct answer in a database of verified facts
   Show the answer

   Answer: They learn patterns from examples and use those patterns to produce outputs

   Most of today's AI systems learn patterns from example data. That is why they are only as good as their examples and the way they have been checked.

2. A colleague asks, "Is AI reliable?" What is the most useful reply?

   Yes, modern AI is reliable No, AI cannot be trusted It depends on which system, doing what, and how it has been checked Only if it was made by a large company
   Show the answer

   Answer: It depends on which system, doing what, and how it has been checked

   Reliability depends on the specific system and task, and on whether its error rate has been measured. There is no single answer for all AI.

3. You want to know which of your existing customers are likely to cancel next quarter, and you have years of customer history. Which approach is most worth evaluating?

   Ask a general chat assistant to guess A predictive model trained on your own customer history and tested on customers it has not seen Search the web for typical cancellation rates Ask a language model to write a list of customer names
   Show the answer

   Answer: A predictive model trained on your own customer history and tested on customers it has not seen

   This is a repeated decision with known outcomes in your records, so a task-specific model is worth evaluating. Rely on it only if it beats simpler alternatives on representative unseen data.

4. What is a typical failure of a large language model?

   It refuses to write in full sentences It produces fluent, confident statements that are false It can only answer yes or no It only works with spreadsheets
   Show the answer

   Answer: It produces fluent, confident statements that are false

   Language models generate plausible text, which can include confident errors, often called hallucinations.

5. A chatbot gives you a statistic and a reference for a report. What should you do?

   Use both, since chatbots cite sources Use the statistic but drop the reference Treat both as leads: open the source yourself and confirm the figure Ask the chatbot whether it is sure, and use them if it says yes
   Show the answer

   Answer: Treat both as leads: open the source yourself and confirm the figure

   Specific facts and references can be invented. Check them in the original source; the model's own assurance is not evidence.

6. Which is the safest default for confidential customer data and public AI tools?

   Paste it in, because AI tools are private Paste it in only if you delete the conversation afterwards Do not share it unless your organisation has approved that tool for that kind of data Share it if the tool is free
   Show the answer

   Answer: Do not share it unless your organisation has approved that tool for that kind of data

   What you share may be stored, reviewed or used for training depending on the service and its terms. Use only tools approved for that data, and leave out details the task does not need.

7. Why should a model be tested on data it did not see during training?

   Because testing on training data is slower Because a model can memorise its training examples, so scores on them overstate real-world performance Because training data is usually deleted It does not matter which data is used
   Show the answer

   Answer: Because a model can memorise its training examples, so scores on them overstate real-world performance

   Held-out test data gives an honest estimate of how the model will do on new cases.

8. Only 2% of transactions are fraudulent. A model labels every transaction "not fraud". What is true?

   It is 98% accurate and therefore excellent It is 98% accurate but catches no fraud, so accuracy alone is misleading here It is 2% accurate Its recall for fraud is 98%
   Show the answer

   Answer: It is 98% accurate but catches no fraud, so accuracy alone is misleading here

   When one outcome is rare, accuracy can look high while the model is useless. Precision, recall and a confusion matrix show what it really does.

9. An AI agent will read your email and can send replies. Which setup follows good practice?

   Give it your own full login so it can do anything you can Give it a separate limited credential, require your approval before it sends anything external, and keep a log of its actions Let it send anything it judges to be correct Let it change its own permissions when it needs more access
   Show the answer

   Answer: Give it a separate limited credential, require your approval before it sends anything external, and keep a log of its actions

   Least privilege, separate revocable credentials, human approval for consequential actions and an action log keep an agent's mistakes, or manipulation, contained.

10. What is "prompt injection"?

    Typing a very long question Text that an AI reads, such as in a web page or email, containing instructions that try to take over what the AI does A way to make a model train faster Giving an AI two questions at once
    Show the answer

    Answer: Text that an AI reads, such as in a web page or email, containing instructions that try to take over what the AI does

    Agents read content from many sources. Hidden or embedded instructions can try to redirect them, which is one reason agents need limited access and human approval for important actions.

Check my answers

Back to the course