Primary Source
FREN

Applied AI at work in Luxembourg
Black Friday price 47.20 € until 30 November, 23:59, VAT included, 12 months of access.

Enrol for 47.20 €

Applied AI at work / Module 1 / Lesson 1.1

1.1 What happens when you press Enter

Video lesson 1 of 5 in module 1 · 3:48 · English captions

Speed
  1. 1.1
  2. 1.2
  3. 1.3
  4. 1.4
  5. 1.5
  6. Exercise
  7. Test

Check yourself: 3 questions

Instant feedback with the reason. Not graded: retake it as often as you like.

1. A plain chat answer, with no file attached and no grounding, comes from:

2. Your tool states a figure in a confident, well written sentence. What does that tell you about the figure?

3. You ask the same question twice and get two different launch dates. Why can that happen?

Key points

  1. A model writes the likely next piece of text. It looks nothing up unless the tool grounds it.
  2. Fluent is not correct, and the same prompt can give a different answer.
  3. The model drafts. You verify every fact against a source you can open.

Your exercise

Try it now: the same question, with and without a source

  1. With no text attached, ask your tool: When was Pinecrest Meridian SICAV launched?
  2. Ask again in a new chat, and note whether the answer changes.
  3. Paste the prompt below this video: three facts, the NOT FOUND rule and a question about the auditor.
  4. Check the result: the auditor must come back as NOT FOUND.
The NOT FOUND prompt
Answer only from the text below. If the answer is not in it, write NOT FOUND. Quote the line you used for each fact.

Text:
Launch date: 14 March 2017
Management company: Alzette Valley Asset Management S.A.
The SICAV has exactly three sub-funds.

Question: When was Pinecrest Meridian SICAV launched, how many sub-funds does it have, and who is its auditor?
You are watching a free lesson.The full course: ten modules of short video lessons, a practical exercise and a scored test per module, a final test and a certificate. Black Friday price 47.20 € until 30 November, 23:59, VAT included, 12 months of access.Enrol for 47.20 €
Transcript of the video

Module one, lesson one. What actually happens when you press Enter. A colleague sends you a summary that an AI assistant drafted. It reads perfectly: confident, fluent, well organised. And one of its dates appears in no document anyone gave the tool. To see how that happens, and how to catch it, you need one idea about how these tools work. A large language model, an LLM, produces text one small piece at a time. At each step it looks at everything it has seen so far: your prompt, any files the tool attached, and the text it has already written. Then it picks a likely continuation. It learned those likelihoods from a very large body of training text. It adds that piece, and repeats. Notice what is missing. Nothing in this loop checks a fact. Three consequences follow. First, the model does not look facts up unless the tool makes it. A plain chat answer comes from patterns learned in training. Some tools add a retrieval step first, called grounding. They search your files, your mailbox or the web, paste what they find into the model's input, and then generate. Microsoft 365 Copilot works this way, with the emails, chats and documents you have permission to access. Second, fluent is not the same as correct. OpenAI researchers wrote in 2025 that models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty. Third, the same prompt can give a different answer. Microsoft's own application card for Copilot calls AI services fallible and probabilistic. Watch it happen. We ask a plain chat tool, with no document attached, about Pinecrest Meridian, the fictional fund used in this course. No training text describes it. The answer comes back in seconds, fluent and specific. A launch year: wrong, the fact sheet says fourteen March 2017. And five sub-funds: wrong again, the fund has exactly three. The model did what it always does. It produced likely-sounding text. Now the fix. Attach the fact sheet, and give the tool one rule: answer only from this document, write NOT FOUND if the answer is not there, and quote the line you used. This time each fact arrives with its evidence. You can search for each quote in the source in seconds. And anything the document does not hold would show up as a visible gap, not as a guess. Three things to keep. A model writes the likely next piece of text, and looks nothing up unless the tool grounds it. Fluent is not correct, and the same prompt can give a different answer. So the model drafts, and you verify. The rest of this module is about doing the verifying quickly. Your turn. It takes three minutes. Ask your tool, with nothing attached, when Pinecrest Meridian was launched. Ask again in a new chat, and see whether the answer changes. Then paste the prompt below this video. It holds three facts, the NOT FOUND rule, and one question the facts cannot answer. A good answer gives you the auditor as NOT FOUND. When you are done, take the three-question check under the video.

Sources cited in this lesson