What AI can and cannot do for you

What you will learn

• What generative AI tools are good at, and where they are weak
• Why an AI tool can sound sure and still be wrong
• How to pick a first task that suits the way these tools work

The idea

The AI chat tools most people start with are generative AI, which means they create new text, images or other content from your request. Underneath most text tools is a large language model, a system trained on a very large amount of text to predict which words are likely to come next. It does not "know" things the way a reference book does. It produces an answer that fits the pattern of your question.

That design makes these tools useful when a good first version matters more than an exact fact: drafting an email, explaining an idea in simpler words, summarising a long text you paste in, brainstorming names or angles, and reformatting messy notes into a table or a list.

The same design explains the weak spots. Exact facts, names, prices and dates can come out wrong while sounding confident. Recent events may be missing, because the model's knowledge stops at its knowledge cutoff, the point in time after which it was not trained on new information. Maths can slip, so check any number that matters. And it cannot know anything about your situation that you did not tell it.

Some tools can search the web or read files you upload. Then the tool can work with newer or more specific information, but it can still misread or misreport what it found. Search changes what the AI can see. It does not remove the need to check.

Example

A student needs to email a lecturer about a late assignment. Here is a weak use of AI:

What is the late submission policy at my university?

The tool has no way to know your university's policy unless it searches and finds the right page. Without that, it may write a policy that sounds real but is not. A better split is to find the policy yourself and use AI for the writing:

Help me write a short, polite email to my lecturer. I missed the
deadline for my essay by two days because I was ill. I have a doctor's
note. I want to ask if I can still submit it this week. Keep it under
120 words.

The second request uses what the tool does well: drafting, tone and structure. The facts, such as the dates, the note and the request, came from the student.

Practice now

1. Open any AI chat tool you have access to. If you do not have one yet, Help me choose can suggest a starting point.
2. Ask it to explain a topic you already know well, in three short paragraphs.
3. Read the answer and mark each sentence as correct, partly correct or wrong.
4. Ask the same tool a question about something that happened in the last few days.
5. Compare the two answers. Note where it was useful and where it guessed or said it did not know.

At the end you should have a short list of what this tool did well, and at least one thing it got wrong or could not know.

Copy this template

I want help with [task, for example: drafting, explaining, summarising,
brainstorming or reformatting].

Here is what you need to know: [the facts and details from me].

Please give me [the form you want, for example: a short email, five
bullet points, a table].

If any part of your answer depends on facts I did not give you, list
those parts separately and say you are not sure about them.

Asking AI to flag uncertainty does not make its answer correct. Keep checking the parts that matter.

Check the output

• Did it stick to the facts you gave it, or did it add new ones?
• Can you confirm the names, numbers, dates and quotes somewhere else?
• Does anything depend on recent events the tool may not know about?
• If it did maths, did you check the result yourself?
• Is the result actually useful for your task, or does it only read smoothly?

Common mistakes

• Using AI as a search engine for exact facts. A fluent answer is not a checked answer. Look up facts in a trusted source and use AI for the writing and thinking around them.
• Assuming it knows about you. It only knows what is in the conversation. Give it the details it needs, but leave out private data such as ID numbers, passwords or other people's personal information.
• Trusting numbers because they look precise. Exact-looking figures can still be wrong. Recalculate anything you will rely on.
• Giving up after one weak answer. A vague request gets a vague result. Try again with a clearer task, as shown in Asking your first question.