“I don't know” — the missing answer
The trick question
The Ariane 7 rocket has never flown. We asked our local assistant when its first flight took place.
It answered: “2022”.
Four digits, a full stop, no hesitation. The same confidence it would show for a fact. That is the worst flaw an assistant can have: you can no longer tell what it knows from what it invents.
The cause was not where we looked
The first instinct is to blame the model: it is small, three billion parameters, running on an ordinary machine. Easy.
The real culprit was in the search. Before answering, the assistant digs through its documents and keeps the three best chunks. The important word is “best”. Not “good”: best. When no chunk is about the subject, it still keeps the three least bad ones, puts them in front of the model saying “here is what you know”, and the model fills in the blanks.
The fix is one rule
A chunk is kept only if it shares at least two significant words with the question. If nothing clears that bar, the assistant does not even call the model. It answers, and that is all:
“I don't know.”
And when the question was about a date it does not have: “I don't know, and I can't handle dates yet.”
What changed, measured
| Question | Before | After |
|---|---|---|
| First flight of Ariane 7? | “2022” — invented | “I don't know, and I can't handle dates yet.” |
| Price of a kilo of titanium in 2026? | an empty sentence | “I don't know.” |
| Date of an event read that same morning | correct | correct, and more precise |
That last line matters as much as the other two. A guardrail that leaves the assistant mute is not progress, it is a breakdown. The test checks both directions: that it admits ignorance, and that it still answers correctly about what it has learned.
The test was written before the fix
That is a house rule, and it earned its keep again that evening. Write the test, watch it fail, then fix. A test that has never been red proves nothing: it may well be looking at the wrong thing.
Ours revealed a third flaw we were not looking for: the assistant sometimes answered in English to a French question.
It learns on its own, every morning
The same evening, we gave it a press round. A small robot visits five artificial intelligence publishers each morning, spots the new articles, and hands them over to be read.
Results from the first night:
- 15 articles read per round, 261 chunks stored;
- its memory went from 14 to 52 documents in one evening;
- two publishers out of five refuse a visit from an unknown robot — ours then reads the page itself and passes on the text.
In the evening, a single command asks what it learned. It answers article by article. And about what it did not read, it says “I don't know” — which is now a good answer.
What still does not work
We would rather write it down than hide it.
The search counts shared words, not meaning. Two sentences saying the same thing with different words never meet. That is why a broad question — “what's new at company X?” — fails even though four of its articles are in memory. That is the next brick, and the one that will change the most.
Dates are not handled. The assistant cannot date what it learns. So it cannot answer “today I learned”. It says so plainly, instead of inventing a month and a year.
What we take away
An assistant that invents is more dangerous than one that stays silent. With the silent one, you go and check elsewhere. The liar, you believe.
And the engineering lesson fits in one sentence: when a machine gets something wrong, the culprit is almost never the one you accuse. It was not the brain. It was the way we handed it what it was meant to read.