Notebook · Note 003 · Content

AI leaves a different fingerprint in every language

If you translate a list of "AI-smelling words" written for English, you end up with a useless list. What bugs a German editor is not the same as what bugs a Polish editor.

Serhat Belen

Founder, BLN Global · · 6 min read

Short answer

Translate one of the supposed AI word lists made for English into another language and you get a useless list. Each language has different tells. In German, the problem is sentence structure rather than word choice. In Arabic, it is often a preposition error carried over from translation. We created separate lists for nine languages and checked 181 patterns across 113 thousand words. The result was zero violations. Passing an audit still does not make the writing good.

This site's copy was written in Turkish, then published in nine languages. It was not translated. It was rewritten for each language. That difference matters, because translation smell and AI smell often show up in the same sentence.

Online lists of "AI words" are for English: delve, tapestry, in today's fast-paced world. Translating this list into German does nothing. What makes a German reader raise an eyebrow is not the word, structureis.

Every language has its own tell

Everyone already has the English list: delve, tapestry, smooth, in today's fast-paced world. Everyone knows these words by now. The problem is that the list only covers English.

In German, it is not the word that jars the reader structureis this: piling up nouns where verbs should work, and hiding in the passive voice. Polish catches something else. Arabic catches something completely different. In Arabic, the clearest sign is a preposition error leaking in from translation, because it gives away that the text was translated, not written.

We built the lists for all nine languages separately, using their own style guides and editorial discussions. We don't publish the lists. They're part of the work our clients pay for. The list isn't the main point anyway: Translating the English list does not work.

Here is the common thread: in each language, the tell is its easiest pattern to memorize. The model memorized it because it appears most often in texts.

Measuring is not the same as reading

Once the list exists, reading pages by eye is not enough. There are 150 pages. We wrote a check that counts patterns. Output:

EN · 15 pages · 13028 words · clean DE · 15 pages · 11968 words · clean FR · 15 pages · 13947 words · clean ES · 15 pages · 13298 words · clean IT · 15 pages · 13276 words · clean NL · 15 pages · 12730 words · clean PL · 15 pages · 11937 words · clean RU · 15 pages · 11544 words · clean AR · 15 pages · 11254 words · clean total dangling strings: 0
9 languages, 181 patterns, 112.982 words. "Clean" means no pattern crossed the threshold.

The tool's own mistake: flattening text into one line

On its first run, the audit flagged dashes in the German pages. The text was correct. The tool was wrong. While extracting text from HTML, it replaced tags with spaces. The page collapsed into one line, making the mark between a label and its description look like part of the prose.

s = re.sub(r'<[^>]+>', '\n', s) # label = block boundary
The false alarm stopped when labels became line breaks, not spaces.

Second false alarm: Polish quotation marks

Polish review ” treated the character as an AI sign. But in Polish, that is exactly the closing quote: „…”. So the tool, the language's right flagged the spelling as an error.

The rule was narrowed: an English quotation mark leaking into Polish text is a real signal, not that language’s own closing mark. This was the second time the site taught the same lesson: a quarter of the accessibility warnings were false too. Fixing the text before the tool means breaking what is right.

Passing an audit does not mean it is well written

Zero hits do not prove the text is good. They only show known patterns were not found. Text can pass every list and still feel lifeless.

The audit catches what you miss and stops a pattern from silently spreading across a hundred and fifty pages. A speaker of the language still decides if the text truly reads well. The tool does not replace that judgment. It saves them time.

For the product-side version of the same discipline AdForge, which generates ad copy you can check the page: there too, every generated text passes a threshold before publishing.

FAQ

Can you just translate the English list of AI words?

It doesn't work. That list contains words specific to English. In German, what bothers the reader often isn't a word at all. It's how the sentence is built. A translated list flags the wrong things and misses the real signs.

Can you publish text once it passes review?

An audit only says the known patterns are not present. A text can pass every list and still feel dead. The audit catches what people miss and stops one pattern from spreading quietly across hundreds of pages. A person who speaks the language still decides whether it reads well.

Does Google penalize AI-generated content?

The production method is not the point. Purpose and quality are. Google treats scaled content production for ranking manipulation as spam under its policies. It does not ban automation itself, and it recommends disclosing its use where that makes sense.

What to do when an audit tool flags an issue

First, verify the warning. In our case, the Polish audit mistook the language’s own closing quote for an error. In the accessibility audit, 6 of 24 contrast warnings were not real. A warning is a claim, not proof. Fixing without checking can break what is already right.

Sources

Related work