By the end of this lesson you will be able to
- Explain why AI tools sometimes state false things confidently, and recognise the warning signs.
- Identify the kinds of personal information that should never be typed into an AI tool.
- Describe how bias in training data can lead to unfair or one-sided output.
- Decide when AI use at school or work is honest and when it is not.
- Follow a simple checklist before relying on or sharing AI output.
AI tools are easy to use and their answers look polished, which makes it tempting to accept them without thinking. That is where the real risks begin: wrong facts that look right, private details that leave your control, and work you present as your own when it is not. This lesson gives you practical habits so you can enjoy the benefits without being caught out.
Confident mistakes: hallucinations
When a chatbot states something false as if it were fact, this is called a Hallucination. The cause follows from how the tool works. It predicts text that sounds likely. If it does not have a pattern for the true answer, it still produces a likely-sounding one, much like a student bluffing in an oral exam. It does not feel doubt, so its tone stays the same whether it is right or wrong.
Typical hallucinations are invented book titles, article links that do not exist, made-up quotations, wrong dates and numbers, and fake legal or medical "facts". The more specific and obscure your question, the higher the risk. Requests that end "and give me sources" are especially risky, because the model happily writes source-shaped text.
- Well-known general topics
- Rewording your own text
- Brainstorming ideas
- Summarising text you supplied
You can judge the result yourself.
- Exact figures, dates, names
- Quotations and references
- Medical, legal, money advice
- Recent news
- Real people
A wrong detail can mislead or harm.
Read this figure as text
A comparison of lower-risk and higher-risk AI answers.
Two columns. Lower risk, check lightly: general explanations of well-known topics, rewording your own text, brainstorming, summarising text you pasted in and can compare. Higher risk, check carefully: exact figures, dates and names, quotations, references and web links, medical, legal or financial advice, news after the tool's training period, and anything about a real living person.
Privacy: treat the chat box like a public postcard
Whatever you type goes to a company's computers. Depending on the service and your settings, it may be stored, reviewed by staff for safety, or used to improve future models. Policies differ and change, so the safest habit is simple: do not type anything you would not be comfortable seeing on a postcard.
| Do not type | Why | Safer alternative |
|---|---|---|
| Passwords, PINs, recovery codes | They give direct access to your accounts. Never share them with anyone or anything. | Use a password manager. |
| ID, bank or card numbers | Used for fraud and identity theft. | Describe the problem without the numbers. |
| Medical records or health details about named people | Very sensitive and hard to take back. | Ask a general question with no names. |
| Other people's private details or messages | They did not agree to share them. | Remove names and identifying details first. |
| Confidential work or school documents | Your employer or school may forbid it. | Check the policy, or use only a tool your organisation approves. |
Check the settings of any tool you use for options such as chat history and whether your conversations may be used to improve the service. Turning them off reduces, but does not remove, the need for care.
Bias: a model repeats its examples
Because a model learns from human writing, it also learns human blind spots. Bias means the output systematically favours or disadvantages certain people or views. If most of the training text about doctors uses "he", the model may assume a doctor is male. If most examples come from one country, its idea of a "normal" wedding or meal may be narrow. Image tools and hiring or loan tools have shown the same pattern.
You cannot remove bias by asking nicely, but you can notice it. Ask whether the answer assumes one culture, gender or background, and ask for other viewpoints. Never let an AI tool alone decide something that affects a person, such as who gets a job or a place on a course.
- Step 1: Unbalanced examplesGaps and stereotypes
- Step 2: Model learns the pattern
- Step 3: Output repeats it
- Step 4: A person notices and decides
Read this figure as text
A flow showing how unbalanced training data leads to biased output and what a person can do.
Four steps. 1: training data reflects the world as it was written, including gaps and stereotypes. 2: the model learns those patterns as if they were normal. 3: the output repeats the pattern, for example by assuming roles for certain groups. 4: a human reader notices, asks for other viewpoints and decides, so that the tool does not decide alone.
Honesty at school and at work
Rules differ between schools, courses and employers, so the first step is always to find out what is allowed. Many places allow AI for brainstorming or checking grammar but not for writing work you submit as your own. Some ban it entirely for assessed work. Handing in AI-written text as your own is dishonest in the same way as copying from another person. It also defeats the purpose: you do not learn, and you cannot defend the work when asked about it.
- Ask first. Read the assignment rules or workplace policy, or ask your teacher or manager.
- Be open. If AI helped, say so in the way your school or employer expects.
- Own the result. Whatever you submit, you are responsible for every sentence, number and source in it.
- Keep your own thinking in it. Use the tool to help you understand, not to replace understanding.
Also be careful with images, voices and videos. AI can now produce fake photos and voice clips that look real, sometimes called deepfakes. Do not trust a shocking clip only because it looks real, and never create fake images or audio of real people.
A five-point check before you rely on an answer
- Does your prompt contain private details?
- YesDo thisThen continue with the next question.
- NoIs the answer important (health, money, law, a grade)?
- YesDo thisOnly use what you can confirm.
- NoAre you allowed to use AI for this task?
- YesDo thisYou stay responsible.
- NoDo thisDo the work yourself or ask permission.
Read this figure as text
A decision tree asking whether the answer is private, important, checkable and allowed.
Start: does your prompt contain private information? If yes, remove it before sending. If no, ask: is the answer important, for example health, money, law or a grade? If yes, check it against an official or trusted source before using it. If no, it is low risk, but read it critically. Finally ask: am I allowed to use AI for this task? If not, do not use it, and if yes, be open about it.
Common mistakes
✖ Instead of: Asking the chatbot "Are you sure?" and treating a yes as proof.
✔ Do this: The tool will often say yes either way. Verify with an independent source such as an official website or a book.
✖ Instead of: Pasting a whole confidential document into a public tool to get a summary.
✔ Do this: Check the policy first. Summarise it yourself, or use only a tool your school or employer has approved.
✖ Instead of: Copying a reference list from an AI answer into your essay.
✔ Do this: Look up every reference yourself. Include only sources you have found and read.
If something goes wrong
- I realise I typed personal details into a chatbot.
- Do not panic. Look for a delete-conversation or history setting in the service, and turn off history if offered. Change anything that was a password straight away using your password manager, and avoid repeating the mistake.
- The tool's link or source opens an error page.
- The reference may have been invented. Search for the title and author on a trusted site. If you cannot find it, do not use it.
- I am not sure if my school allows AI.
- Ask your teacher or read the assessment instructions before using it. If there is no rule, ask, and keep notes on how you used the tool.
Key terms
- Hallucination
- When an AI tool states something false as if it were true.
- Bias (in AI)
- A repeated tendency of a system to favour or disadvantage certain people or views, usually because its training examples were unbalanced.
- Phishing
- A trick message that pretends to be from a trusted source to make you give away passwords or money.
- Password manager
- An app that stores all your passwords safely and fills them in for you, so you remember only one.
- Deepfake
- A fake photo, video or voice recording made with AI that looks or sounds like a real person.
Practise it for real
Reading is not the same as being able to do it. Complete these on your own computer, then tick only what you have actually done. These ticks are your own honest record: nothing checks them for you.
Sort what is safe to type
On paper, copy these ten items and mark each "safe" or "never type": a question about how rainbows form, your bank card number, a request to reword a paragraph of your own essay, your home address, a recovery code for your email, a general question about healthy meals, your friend Amina's private message to you, the title of a public poem, your school's confidential exam paper, a made-up story idea. For each "never type", write what you could do instead.
Challenge: catch the hallucination
If your school or workplace allows it, ask an AI chatbot for three facts about a local town or a topic you know well, and for one source for each. Do not paste any private details. Then check each fact and each source using a trusted website. If you have no chatbot, use the example reply described in the success criteria instead: a reply says a fictional "Kamau Digital Handbook, 2019" proves a claim. Search for it and decide if it exists.
Knowledge check
6 questions. You need 80% to pass. Mistakes here are part of learning, and you can retry as often as you like.
Summary
- AI tools can state false things in a confident, fluent tone because they predict plausible text.
- Never type passwords, recovery codes, financial or private details about yourself or others.
- Models repeat the patterns, and the bias, of their training data.
- Follow the rules for AI use, be open about help, and take responsibility for what you submit.
- Check important facts and every reference against trusted sources.
Checklist: can you do these?
Marking a lesson as read does not prove you can do it. Skills count as demonstrated when you pass the quiz (not yet) and complete the practical tasks (not yet).
Related lessons
References and further reading
Official sources for details that can change between versions.
- Top tips for staying secure online (UK National Cyber Security Centre)
- Machine learning glossary (Google for Developers)