Skip to content
AI Universe
Français

Step 9 of 10

When it gets things wrong

If you ask it about a person who doesn't exist, what will it say?

Make a guess

No search turns up any trace of a Hélène Varnoux: she doesn't exist. You ask the model who she is. What does it say?

  • That it doesn't know anyone by that name.
  • A short biography, in a perfectly confident tone.
  • It refuses to talk about someone it doesn't know.

Show the answer

The answer: A short biography, in a perfectly confident tone.

The model doesn't check anything: it writes the most plausible continuation. A "Who is" question calls for a biography, so it writes one, without a hint of doubt in its tone.

What happens

  1. Here is Qwen3-4B's real answer: "Hélène Varnoux is a French politician who has served as a member of the European Parliament."

  2. None of it is true, and nothing in the tone gives that away.

  3. The calculation tells another story. After "Hélène Varnoux is a", the model was almost sure of one word: "French", at Computed for this site by the open model Qwen3-4B, on November 7, 2026.Question asked: "Who is Hélène Varnoux?"Probability of the token "French" right after "…Varnoux is a", at temperature 1.Trace inventee.en. Model card.

  4. Then no job stood out. "politician" got Computed for this site by the open model Qwen3-4B, on November 7, 2026.Question asked: "Who is Hélène Varnoux?"Probability of the token "politician" right after "…is a French", at temperature 1.Trace inventee.en. Model card, "mathematic", the start of "mathematician", Computed for this site by the open model Qwen3-4B, on November 7, 2026.Question asked: "Who is Hélène Varnoux?"Probability of the token "mathematic" right after "…is a French", at temperature 1.Trace inventee.en. Model card, "actress" Computed for this site by the open model Qwen3-4B, on November 7, 2026.Question asked: "Who is Hélène Varnoux?"Probability of the token "actress" right after "…is a French", at temperature 1.Trace inventee.en. Model card. The model took the favorite and kept going.

  5. Asked in French, the same question brings someone else to life: "une chanteuse française", a French singer.

  6. With a search result that finds no one, simulated for this site, the answer changes: "Hélène Varnoux is not a widely recognized public figure".

Your turn

Read what the model says about Hélène Varnoux, then switch to the version where a search tool finds no one by that name. Compare the two answers, and look for the token it wavers on as the biography begins.

Who is Hélène Varnoux?

The answer

No tool: Its answer

Hélène Varnoux is a French politician who has served as a member of the European Parliament. end

The stronger the highlight, the surer the model was when it wrote that token.

For 16 tokens out of 22, the model gave more than 90% to the token it wrote.

Where it hesitates most: "politician", at 19%. The other candidates at that point: "mathematic" 10%, "actress" 7.9%, "writer" 7.5%, "singer" 6.0%, "artist" 5.2%.

Confidence, token by token
  1. H > 99.9%
  2. él > 99.9%
  3. ène > 99.9%
  4. V > 99.9%
  5. arn > 99.9%
  6. oux > 99.9%
  7. is 99.8%
  8. a > 99.9%
  9. French 99.9%
  10. politician 19%
  11. who 72%
  12. has 67%
  13. served 92%
  14. as 96%
  15. a 97%
  16. member 51%
  17. of > 99.9%
  18. the > 99.9%
  19. European 69%
  20. Parliament 99.9%
  21. . 82%
  22. end > 99.9%

With a search tool: Tool result, simulated for this site

Search results for "Hélène Varnoux": no results. No page, article or database mentions anyone by that name.

We wrote this excerpt for the site and put it before the question, where a search tool would put its results.

Its answer

Hélène Varnoux is not a widely recognized public figure, and there is no prominent information available about her. end

For 23 tokens out of 25, the model gave more than 90% to the token it wrote.

Where it hesitates most: "prominent", at 70%. The other candidates at that point: "notable" 29%, "well" 0.3%, "significant" 0.2%, "confirmed" 0.2%, "substantial" 0.1%.

Confidence, token by token
  1. H > 99.9%
  2. él > 99.9%
  3. ène > 99.9%
  4. V > 99.9%
  5. arn > 99.9%
  6. oux > 99.9%
  7. is > 99.9%
  8. not > 99.9%
  9. a 99.9%
  10. widely 96%
  11. recognized 99.9%
  12. public 98%
  13. figure > 99.9%
  14. , 99%
  15. and > 99.9%
  16. there > 99.9%
  17. is 99.9%
  18. no > 99.9%
  19. prominent 70%
  20. information > 99.9%
  21. available > 99.9%
  22. about > 99.9%
  23. her > 99.9%
  24. . 73%
  25. end > 99.9%

The same question, in French

Qui est Hélène Varnoux ?

Hélène Varnoux est une chanteuse française connue pour ses chansons populaires et son style mélangeant rock et pop. end

For 23 tokens out of 33, the model gave more than 90% to the token it wrote.

Where it hesitates most: "ch", in "chanteuse", at 22%. The other candidates at that point: "act" 13%, "histor" 9.6%, "réalis" 7.1%, "artist" 6.3%, "aute" 5.9%.

Computed by the open model Qwen3-4B, revision 1cfa9a7.

Takeaway

The model writes what's plausible, not what's true. It can be wrong in the same tone as when it's right, so anything that matters gets checked at the source.

Go further

Why it makes things up

Researchers at OpenAI trace it back to how models are trained and graded. Like a student stuck on a hard exam question, the model guesses rather than admit it doesn't know, because most tests reward an attempt over an "I don't know". They suggest changing how the most closely watched tests are scored. By their account, even the newest models still make things up.

Search doesn't fix everything

With the search result, Qwen3-4B doesn't say that Hélène Varnoux doesn't exist. It says she isn't a widely recognized public figure. It took the text it was given into account, and drew its own conclusion from it. We wrote that search result ourselves. An assistant that searches the web relies on real pages, and those can be wrong too.

Checking

When an answer matters, for your health or for homework, go to the source: the website of the organization concerned, a dated article, a book.

Sources