Level 100: How AI works: AI at work and in life · Module 100.7

Try each exercise first, then open its model answer.

Keeping secrets and people safe · Workbook

Three exercises, one for each thing this module teaches. Everything here is made up, set at Big Widget Company. Do not use real names, customers or documents from your own work. Write short answers. "I don't know yet" is an honest answer if you also say who you would ask.

Exercise 1 (100.7-a)

Sort it with the traffic light. Big Widget Company has two AI tools:

It has no tool approved for red information. A free chatbot you found online is not approved at all.

For each item below, write its colour and which tool it may go into (P, I, or none). Then give a one-line reason.

  1. The text of a press release the company published last week. You want a shorter version in the free online chatbot.
  2. Your team's meeting notes about next month's office move. You want a list of action points.
  3. A spreadsheet of customer names and phone numbers. You want it sorted by city.
  4. A drawing of a widget that has not been announced, with the project name deleted. You want ideas for a better shape.
  5. A generic request: "Write a polite email asking a supplier to confirm a delivery date", with no names, prices or contract terms.

Last, answer in one or two sentences: a colleague says, "It's fine, the chat is private." What would you tell them?

Model answer
  1. Green, but not in the free chatbot. Published on purpose, so it is green. The free chatbot is not approved, and nothing goes into an unapproved tool, not even green. Use Tool P (or Tool I).
  2. Yellow, Tool I only. Everyday internal notes. Tool I is approved for internal information; Tool P is not.
  3. Red, none. Customer personal data is red. No tool here is approved for red, so it stays out unless a specific approved process says yes. Sort it in the spreadsheet itself.
  4. Red, none. Unreleased designs are red. Deleting the project name does not make it green: the design itself is the secret. If you are unsure, keep the stricter colour or ask the owner.
  5. Green, Tool P (or Tool I). No names, prices or terms are left, and the task is generic. Fill in the details yourself afterwards.

On "the chat is private": a good answer says that the company running a public chatbot can see and store what you type, may use it to develop its service, and stored questions can be hacked, leaked or made public by mistake (NCSC, 2023). What a tool keeps depends on the tool, its settings and its terms, so check the colour and the tool before you paste. Once you press enter, it has gone.

A strong answer also asks where a tool can send things. A tool that can upload or send on its own can move your information somewhere you never chose.

Exercise 2 (100.7-b)

Two tricks, two defences. Read both situations.

For each one:

  1. Who or what is being fooled: a person or the AI?
  2. Why might the trick work?
  3. Name one defence, and say what the gate answer is (go ahead, only this far, wait for a person, or no).
Model answer

A. A deepfake fools a person.

  • It can work because a face and voice can be faked, even on a live call. The FBI warned in December 2024 that criminals use AI to fake real-time video calls from company executives.
  • Defence: check another way. Mei says she will call back, hangs up and calls the head of finance's office on the number in the company directory, never a number given in the request. Big Widget Company's rule also needs a second person to approve any unusual or urgent payment.
  • Gate answer: wait for a person, until a second person has confirmed the request through the agreed channel. A face is not a signature.

B. Prompt injection fools the AI.

  • It can work because today's AI models do not keep a security boundary between the instructions they are given and the content they read (NCSC). Hidden text in an email can act like an order.
  • Defence: limit what the AI can do. The helper that reads outside email should not be able to send emails or share files without a person's yes. That yes must be set in the tool's own permission controls, outside the AI. Writing "ask me first" in its instructions is not enough, because a fooled AI can ignore it.
  • Gate answer: reading and summarising can be only this far; sending or sharing is wait for a person. If the remaining risk is still too high, the answer may be no: the NCSC says some jobs may not be a good use for AI at all.

A strong answer notes that prompt injection may never be fully fixed (NCSC), so the defence is to limit the damage, not to hope the AI spots the trick.

Exercise 3 (100.7-c)

Plan a fair screening tool. Big Widget Company's HR team wants a tool that sorts applications for warehouse jobs and turns some applicants away automatically. A manager says: "It's a computer, so it will be fair. Just switch it on."

Write a short reply to the manager with:

  1. one reason the manager's belief is wrong, using a real case from this module;
  2. one reason bias can happen even when nobody means it;
  3. what HR should do before the tool runs, and who should be named;
  4. one thing to keep checking while it runs.
Model answer
  1. A machine follows its rules and its data, and both can be unfair. In 2023 the US EEOC settled with iTutorGroup for $365,000. The EEOC said the company's software automatically rejected women aged 55 or older and men aged 60 or older, and that more than 200 qualified US applicants were turned away because of their age. Once a rule is inside software, it runs on everyone, every time, and nobody sees it unless someone checks. (The EEOC did not say how the software was built; this is about automated decisions, not chatbots.)
  2. Bias does not need a bad person. NIST says AI bias can come from society and organisations, from data and maths, and from how people think, and that each can happen with no intent to discriminate.
  3. Before: test the tool's results across age, sex and other protected groups. Name a person who reviews every rejection and can overturn it, who that person goes to if something looks wrong, and who can switch the tool off. NIST calls for clear roles and responsibilities for human oversight.
  4. While it runs: check the results again whenever the tool or the job changes, and check that the named person is really reviewing, not just clicking approve. Relying too much on what the system says is called automation bias.

A strong answer gives the gate answer for an automatic rejection. A rejection is hard to undo, and the course rule is that anything you cannot undo waits for a person. So it is wait for a person, the named reviewer, not the tool alone.

Next module: 100.8 Checking the work