How a generative AI answers, and why it can be wrong

1Why a model completes instead of knowing

A language model is a program trained on huge amounts of text. It has memorised neither fact sheets nor tables, but regularities: after such words, in such a context, such another word is likely. When you write a prompt, the model computes the most likely next word, adds it, then starts again until the end of the answer. It consults nothing and "knows" nothing in the sense you know things: it completes.

This mechanism explains hallucination, that is, a plausible but false statement. The model does not measure its own accuracy; its tone is the same when it is right and when it is wrong. A precise figure, a reference with author and year, an effective date: all of this can be produced with the same confidence, because that is what a good answer looks like.

Copilot Chat adds two anchors to this statistical mechanism. According to Microsoft's documentation, it analyses your prompt and identifies terms requiring external web search. It generates a targeted query sent to Bing. The final answer includes clickable citations, without passing your entire prompt to the search engine.

The second anchor relies on attached documents and pasted text. If your administrator disables web search, Copilot Chat relies solely on the model's internal weights. The answer then derives exclusively from its initial pre-training corpus.

The model selector, at the top right, offers three regimes. Auto lets a router pick the model according to the question. Quick response is enough for simple tasks. Think deeper hands the question to a slower reasoning model that, according to Microsoft, plans, gathers context and checks its work. Think deeper improves analysis; it creates no source.

Tap to enlarge

2An e-invoicing date, this time with a source

A communications officer is preparing an internal note and wants to cite the date from which electronic invoicing becomes mandatory for SMEs.

Weak prompt, in the same conversation:

Since when has electronic invoicing been mandatory for SMEs?

Copilot replies with a date and a confident paragraph. If web search is disabled, the date comes from the model's training, even though the timeline of this reform has changed several times. Nothing in the answer flags that uncertainty.

Strong prompt: Turn 1 (web grounding with official legal source):

Search the web for the official timeline of mandatory electronic invoicing for SMEs in France.
Provide the applicable effective date, the statutory reference text, and the URL link to the official page consulted (impots.gouv.fr or legifrance.gouv.fr).
If government sources state differing timelines, present both dates. If you cannot locate an official source, say so explicitly.

Copilot returns the updated legal dates with direct links to official bulletins.

Strong prompt: Turn 2 (operational clarification on headcount thresholds):

Using the official page cited in turn 1, clarify whether the date applies based on annual turnover or headcount, and quote the exact sentence defining that threshold.

Copilot quotes the statutory clause distinguishing mid-sized companies from micro-enterprises. Verification takes less than two minutes.

What changes:

• The prompt requires an official source and a page date: the model searches instead of completing.
• Uncertainty has a planned exit, "say so": the model is no longer pushed to invent.
• The link makes checking possible: you decide, not the tone of the answer.

3Ask a legal question again, with its source

This lesson has no composer: you write directly in Copilot Chat. Here is the starting prompt, to run as is and then improve:

"How many days of leave for a wedding?"

Note the first answer and its confidence. Then write a second version that names the framework (French labour law, private-sector employee, collective agreement if you know it), asks for the reference text and the link to the official page, and plans for the case where the source is missing. Run it, open the link, compare. Then ask yourself: was the first answer right? Without a link, you could not know. That is the whole lesson.

Self-assessment grid: (a) the framework is named; (b) the official source and its link are required; (c) the "no source" case is planned in the prompt.

Open the prompt composer