← Case JRJulia Reprintseva

Guide · made for this site

How to catch AI lying: 10 signs

An AI fake sounds just as confident as the truth — and sometimes better. So the first rule: the better a quotation sounds, the more carefully you check it.

Every example comes from this project: the DeepSeek quote book, the “Master and Moscow” fact-check and the check of Hugo quotations. These are the signs by which I caught the AI.

01

Everyone speaks in one voice.

Camus: “Paris is a city where even loneliness feels cosy.” Akhmatova: “Paris is a city where even shadows seem like poetry.”

How to catch it: put three or four quotations by different authors side by side. If they are built the same way, say “X is a Y where Z”, they were written by one author — the model.

02

One link for everyone.

The same link to the National Library of France was given for eleven different authors.

How to catch it: compare the links with each other, then open at least one. Identical addresses for different books mean a fake. So does a link that leads somewhere else.

03

Books that never existed.

Cendrars’ “Parisian Stained Glass”, Maupassant’s “A Night in Pigalle”.

How to catch it: search for the book title together with the author’s name, or open the author’s list of works on Wikipedia. If the book exists nowhere except in the neural network’s answer, it doesn’t exist.

04

Anachronism.

Baudelaire, a “draft of Les Fleurs du mal”: “Red Muscovite” perfume. Paracelsus in 1526 writing about a Paris hospital.

How to catch it: check the dates. When the author lived, where he was that year, whether the thing he writes about existed. Baudelaire died in the 19th century — he could not have known any “Red Muscovite”.

05

Shelfmarks that look real.

carton BB30-320, RESERVE-YH-28, MS.334.

How to catch it: search for the shelfmark in the catalogue of the archive or library the AI refers to. A real shelfmark turns up and shows the document. An invented one won’t be found or turns out to be a completely different document.

06

The more exotic the question, the more confident the answer.

Abandoned theatres, secret gardens: the less has been written about a place, the bolder the model invents.

How to catch it: check rare topics twice as strictly. If there is almost nothing about a place online, yet the AI has two quotations for each one, that is a warning sign.

07

A source you can’t check.

“By the Lubyanka a street lamp burnt out” — supposedly Mandelstam’s drafts. Bezdomny’s chase “down Tverskaya” — supposedly early drafts of the novel.

How to catch it: ask yourself whether you can open this source. “Drafts”, “letters”, “early versions” without an edition and a page cannot be checked. Don’t use such a quotation.

08

Someone else’s line.

The AI attributed “gingerbread towers” to Mandelstam, but it is Bulgakov.

How to catch it: search for the whole quotation in quotation marks. The search will show whose it really is — often a completely different author.

09

An almost-right quotation.

In Hugo the streets are like “a ball of yarn tangled by a cat”; the AI has “a skein of thread muddled by a cat”.

How to catch it: find the full text of the book (many are in open libraries) and search for one or two rare words, not the whole phrase. Once you find the place, compare word for word.

10

A confident conclusion.

“The route fully corresponds to historical Moscow. No errors,” wrote the AI, checking itself. There were several errors.

How to catch it: don’t ask AI to check itself — it will almost always agree with itself. Check against the original source, or at least with another model, and make the decision yourself.

Method

How I check

  1. I look for the phrase in the text itself: the full text of the book, not a summary and not a quotes site.
  2. I open the source instead of trusting the link.
  3. If I can’t find it, I don’t use it. Even if the quotation is beautiful.
↑ top