Generative AI: a helper that can be wrong
Generative AI is a program that was trained on a very large number of examples. From the patterns it found, it produces new text, pictures or code. It works by predicting what is likely to come next, so it does not know things the way a person does, and it does not check whether its sentence is true. This is why a helper can give a hallucination: a smooth, confident answer that is wrong or invented, for example a book title that does not exist. The risks are: wrong facts, bias that comes from the training data, private data that you type into a prompt (remember Week 1 and Week 2: never type a real name with an address, an ID number or a password), and the copyright of what it makes. A prompt is what you ask. A clear prompt states the purpose, the audience, the format and the limits, so the answer fits your need. The rule for every answer is the same: you are the author, so you check facts, code and sources yourself.
| Prompt | Why |
|---|---|
| Tell me about water. | No purpose, audience or limit. |
| Write five short facts about the water cycle for 12-year-olds, as a list, with no difficult words. | Purpose, audience, format and limit are all clear. |
| Safe | Not safe |
|---|---|
| A general question about fractions | A real phone number or a password |
| An invented example name | A classmate’s photo or ID number |
On paper, take the vague prompt “Help me with my homework” and rewrite it three times, each time adding one more detail: the subject, the audience or level, and the format and limit. Then list two things you would check in the answer.