Skip to content
POLARISAI ACADEMY
F03 · First Light

When the AI makes things up: check the answer

You are in the staff room after a late shift in home care, and a relative or a colleague is waiting for a reply before you go home. You ask an AI chat for help, and what comes back sounds finished, with a name, a year, and a link. Before you send it on, you look to see whether those details exist.

Level
First Light
Time
1 h (estimate)
Prerequisites
F01
Course 3 of 7
Jump to

After this course you can

After the course you can spot a detail an AI chat invented, open the source yourself, and ask a person when the matter is your health, a contract, or money.

Names, dates, and links

The chat can write a name that sounds believable, a date that seems reasonable, and figures that give an air of precision, as well as a book title, a clause of a law, or a link you can tap. The detail may come from the text you pasted in, and it may just as easily have been invented in the same reply.

A link that leads nowhere is obvious once you open it. Harder to spot is the name of someone who never worked with you, a book the library does not have, or a visiting time that was right last year and wrong this Saturday. A clause can look real on the page and still point to the wrong place.

You recognise the pattern when the reply gives you something you never supplied. You wrote that a relative asked about visiting time on the ward, and the chat filled in a clock time, a year, and a web address, details you have to check yourself in a source you trust.

The same calm tone when it is wrong

The chat writes in the same calm tone whether the detail was in your notes or the chat invented it. The text runs on without hesitation and sounds as reassuring as an email from someone who knows the job, which makes it easy to be taken in.

The words follow one another so the sentence sounds finished, even when the chat has not looked up your question. It has the text you pasted in and the phrases that usually appear in replies of that kind. A colleague who was there that day has the whole situation clear, while the chat has no memory of the person you help, and no grasp of that day.

So you read names, dates, figures, and links once more, even when the sentence flows. A reply to a relative, an office, or a bank can rarely stand a name or a sum that nobody can show you the source of.

Three checks you have time for

First you ask the chat to name the source, and then you open that page yourself. If the detail is not there, you take it out of your reply. If the page is only mentioned in the chat, you have not seen the detail yourself.

Then you compare it with the page of the people who handle the matter. You read visiting times at the clinic or on the council page, opening hours on the library page, and the terms for a product on the maker's page or on the receipt. Anything about your own situation you read on the office's own page.

Last, you ask a person who knows the subject when it concerns your body, a contract you are about to sign, or money that will leave the account. The chat can help you write the questions in plain words before you call your doctor's advice line, the lawyer, or your bank. You take the answer from them, and you send nothing you cannot stand behind.

Details about clients, customers and colleagues never go into an AI chat, not even with the name swapped out. If you want to use AI at work, ask your manager what the rules are.

The card for your aunt and the letter to the tax office

A birthday card for your aunt is something you can draft with the chat on a coffee break. If a line feels stiff, you cross it out and put in your own words, because you can hear whether the greeting sounds like you. Nobody expects a date, a sum, or a clause of a law in a card like that.

A letter to the tax office, a reply to the landlord, or an email about a bill is a different matter. An invented date, a wrong sum, or a clause that does not exist can cause you trouble for weeks. You read every detail against the letter you received and against the office's own page before you send.

Use the chat when you want help finding the words, but always read the reply through before it goes any further. You trust a detail when you have seen it on a page you opened yourself, or when a person who knows the subject has told you the same thing.

Take aches and medicines to the clinic, and let the chat help you write what you want to ask. Read a contract, go through what binds you with a lawyer, and let the chat phrase the questions. Check sums and dates against the letter and against your bank, and send nothing where the figure exists only in the chat.

An example

The book that was not there

You want to suggest something on sleep to a colleague who works nights in home care, and you ask the chat for a title and an author in your own words. The reply comes at once and sounds settled, with a name, a year, and a link.

You search the library catalogue on your phone. No book has that title, no author has that name, and the link does not lead to any such book.

You take out the title, the name, the year, and the link before you pass anything on. Then you ask at the desk, or you search yourself for the word sleep.

The title and the name were not real

You wrote

Suggest a book about sleep for someone who works nights in care. Give the title and the author.

The chat replied

Read Rest Between Shifts by Helen Ward, published in 2017. There is more about the book at example.com/rest-between-shifts, and most libraries have a copy.

The chat made up Rest Between Shifts by Helen Ward. It is not in your text, so you take it out.

Exercise

Five minutes on your break

Open the free version of an AI chat on your phone. Ask for the library's opening hours next Saturday, something you can check in five minutes.

Then ask the chat to name the source and open the link yourself. Go to the library's own page and read the same thing there. If the two do not say the same thing, stay with the page the library runs.

Finally, write one sentence you could send to a friend. Cross out the time or the link that was not there when you looked, and keep only what matches the page you opened yourself.

Check questions

Answer on your own first. Then open a question to see the answer.

The chat puts a name from work in the reply. What do you do?

You leave the details out and ask your manager what the rules are for AI at work.

A link looks real. What do you do before you send it?

You open the link yourself and see whether the page is about what the chat claimed.

The chat states a sum in a letter to the tax office. What do you trust?

You trust the letter and the office's own page, or a person who handles the matter.

How careful are you with a birthday card compared with a bill?

You rewrite the card in your own words, and on the bill you check every figure against the letter.

Your back hurts and the chat sounds sure. What do you do tomorrow?

You call your doctor's advice line, and the chat only helps you phrase the questions.

Take with you

Keep details from work out of the chat, ask for the source and open it yourself, and send no sum that appears only in the chat.

Final task

Ask an AI chat for something you can check yourself, such as opening hours or a book at the library. Mark one thing it invented and note where you saw that it did not hold.

The final task is marked against fixed answer keys and automated tests once the tests open.

Going deeper (optional)

Want to know how the technology behind the chat works? There is an optional deeper lesson. It has worked examples with numbers and more technical words. Read the deeper lesson →

Tests and diplomas coming

This lesson is open to read. Tests, diplomas and accounts come later; the waitlist is for those. Join the waitlist