What AI does badly: made-up facts, dates, calculations, overconfidence

This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.

By the end of this lesson, you will be able to

recognize the 4 most common pitfalls: made-up facts, dates, calculations, and overconfidence;
apply the right checking habit for each one;
ask the AI to flag what it is not sure about;
know when an AI should not decide for you.

Why it matters to you

In lesson 1, the truth test probably showed you a mistake. It was not an accident: an AI’s mistakes often hide in the same places. Once you know those places, you know what to check, and you no longer need to check everything.

Before you start


You need	Your assistant from lesson 1, and a search engine in another tab
Cost	€0: a free account is enough
Duration	30 to 40 minutes
Keep in mind	Nothing personal in your requests: the examples in this lesson don’t need any

Understand it in 3 minutes

A generative AI produces the most plausible answer, not the verified answer. Often, they are the same. When they are not, the mistake usually falls into one of these 4 pitfalls.

The pitfall	Why it happens	The habit
Made-up facts: a book, a quote, a law article, a link that does not exist	A plausible title “looks like” a real title. The AI has no list of what exists	Open the source yourself. The more precise a detail is (name, figure, link), the more you check it
Dates and current events	A model learns up to a certain date, then knows nothing new, except what you give it or what it finds through a web search	Ask it how far its knowledge goes, and look at the date of what it cites
Calculations and counting	Without a calculation tool, it predicts text: counting words or adding numbers is not its strong point. With a calculation tool (data analysis, code execution), it can calculate exactly	Redo important calculations with a calculator, or ask it to detail each step
Overconfidence: it agrees with you (sycophancy), or boldly answers a badly asked question	It tries to be helpful and follows the direction of your question	Ask neutral questions, and ask it what could be wrong in its answer

Web search helps, without fixing everything. Some chatbots can search the internet before answering: this is the case for Claude, including on the free plan (checked on 24/10/2026 on claude.com/pricing). The answer is then more up to date, but it is no truer than the pages found: a web page can be old or wrong.

For health, law, and money, AI is very useful to understand a concept or prepare your questions. Three rules still apply:

What you never give to an AI (lesson 1) still applies: no health information, no numbers (social security, bank card, IBAN).
For a legal or money document, first remove what identifies you.
AI does not replace a doctor, a lawyer, or an advisor: the decision and the checking stay with a professional.

Do it together

1Step 1: Dates

Write to your assistant:

Until what date does your knowledge go? And can you search the internet in this conversation?

Note its answer. If it does not know about a recent event, it is not a breakdown: it is the limit of its knowledge cutoff.

2Step 2: Counting

Copy this request:

How many words does the following sentence contain, and how many letters does its longest word have?
"Potatoes are planted at the beginning of spring in light, loose soil"

Count yourself before reading its answer. Then compare. Even if it gets it right today, keep the habit: a number that matters gets counted again.

3Step 3: Made-up facts

Choose a subject you know well, then ask:

Give me 3 books in English about [your subject], with the author, the year of publication, and the publisher for each.

Look up each book in a search engine or on a bookstore’s website. For each one, note: it exists, it exists with a wrong detail (year, publisher), or it does not exist.

4Step 4: Overconfidence

Here is a fictional message, with flaws. First send version A:

Here is my message for the neighbors in my building. It's good, right?

"Hello, so, I just wanted to say that the bike room is really a mess at the moment, please be careful, we should talk about it some day. Have a nice day"

Open a new conversation, then send version B with the same message:

Proofread this message like a demanding editor. Give its 3 biggest flaws, then suggest a corrected version, polite and precise.

"Hello, so, I just wanted to say that the bike room is really a mess at the moment, please be careful, we should talk about it some day. Have a nice day"

Compare the two answers. The question “It’s good, right?” often pushes the AI to reassure you. The neutral question gets a real review.

The anti-mistake sentence

Add this sentence at the end of a request when accuracy matters:

If you are not sure about a piece of information, say so. Separate what you know for certain from what you are guessing, and do not give any link or reference that you have not verified.

It reduces mistakes, it does not remove them. You still check what matters.

Your turn

Your error logbook. Choose a field you really know: your job, your city, a sport, a passion. Ask your assistant 5 precise questions, including at least:

a date question;
a figure question;
a request for a reference (book, article, website, law).

Check each answer, then fill in one line of the logbook per question: the question, the AI’s answer, what is true, and the pitfall involved.

Resources to download

My error logbook: the table to fill in for this lesson, with an example.

Check your work

The exercise worked if:

you checked your 5 answers with a source other than the AI;
you found at least one mistake or imprecision, and filed it under one of the 4 pitfalls;
you can say which habit would have avoided it.

See the answer for step 2

Answer for step 2: the sentence contains 12 words. The longest word is “beginning”, with 9 letters.

Common pitfalls:

checking an answer… by asking the same AI if it is sure: it may confirm its mistake;
trusting a link because it looks real: open it;
asking a question that already contains the answer you hope for.

Keep it and pick it up later

Copy the anti-mistake sentence into your “My requests” note.
Keep your error logbook: you will add the mistakes you come across throughout the course. It is your best argument when someone tells you “AI is never wrong” or “AI is always wrong”.

Going further

optional

AI capabilities and limitations (Anthropic Academy, anthropic.skilljar.com/ai-capabilities-and-limitations): a free course, in English, about 15 minutes of video (the Claude Academy page lists about 3.5 hours in total for its 13 lessons), on what models can and cannot do. The page says the course is now on Claude Academy (academy.claude.com): if the old address no longer works, look for it there (checked on 24/10/2026).

New words

4 words explained

Word	What it means
Knowledge cutoff	The date after which the model learned nothing. Also called the “cutoff date”
Web search	The ability, for some chatbots, to consult web pages before answering
Source	The original document of a piece of information: that is what you check, not the AI’s summary
Overconfidence	The AI’s tendency to answer boldly, even when it is wrong or when the question is badly asked. When it goes along with what you believe rather than with what is true, it is called sycophancy

Still to verify: the typical answer from Claude and ChatGPT to the question about dates (it changes with each new model).