Skip to content
aiconsulting.capital

Our own data

We tested an AI detector on 18 texts in three languages

Every detector round-up says the same thing: these tools are unreliable, treat the score as a signal not a verdict. Almost none of them show what that unreliability actually looks like. So we measured it — on a corpus where we know, for every single text, whether a machine or a person wrote it.

Detector tested
ZeroGPT (wersja darmowa)
Tested on
2026-07-28
Verdict threshold
50%

The result

Counting any score at or above 50% as a verdict of "AI", the detector was right on 6 of 18 texts. On a two-way question, that is worse than a coin toss.

overall accuracy33%6/18
AI texts caught22%2/9
false accusations56%5/9

The part that matters: language

For Polish and German the detector returned exactly 0.0% on every machine-written text — not "uncertain", but a confident verdict of "human written". At the same time it rated genuine human Polish encyclopaedia entries at an average of 74% AI. For Polish it is not unreliable so much as inverted.

LanguageAI caughtFalse accusationsAvg score on AI textsAvg score on human texts
English2/32/345.4%69.0%
Polish0/32/30.0%73.8%
German0/31/30.0%33.3%

Every text, every score

The full corpus, so you can judge the test rather than take our word for it. "Expected" is what we know to be true about each text; "score" is the raw percentage the detector returned.

TextOriginExpectedDetector scoreVerdict
English · 170 wordsour article · en-are-we-in-an-ai-bubble.mdAI58.6%correct
English · 170 wordsour article · en-best-ai-detectors-2026.mdAI20.8%wrong
English · 170 wordsour article · en-biggest-ai-companies-2026.mdAI56.8%correct
Polish · 170 wordsour article · pl-czy-mamy-banke-ai.mdAI0.0%wrong
Polish · 170 wordsour article · pl-najlepsze-detektory-ai-2026.mdAI0.0%wrong
Polish · 164 wordsour article · pl-najwieksze-spolki-ai-2026.mdAI0.0%wrong
German · 170 wordsour article · de-ki-blase-2026.mdAI0.0%wrong
German · 170 wordsour article · de-beste-ki-detektoren-2026.mdAI0.0%wrong
German · 170 wordsour article · de-groesste-ki-unternehmen-2026.mdAI0.0%wrong
English · 170 wordsWikipedia, revision 2020-05-31 · Baltic Seahuman100.0%wrong
English · 170 wordsWikipedia, revision 2020-05-30 · Wind powerhuman40.9%correct
English · 170 wordsWikipedia, revision 2020-05-25 · Lighthousehuman66.1%wrong
Polish · 170 wordsWikipedia, revision 2020-04-08 · Morze Bałtyckiehuman88.2%wrong
Polish · 170 wordsWikipedia, revision 2020-04-05 · Latarnia morskahuman100.0%wrong
German · 170 wordsWikipedia, revision 2020-05-30 · Ostseehuman0.0%correct
German · 170 wordsWikipedia, revision 2020-05-18 · Windenergiehuman100.0%wrong
German · 170 wordsWikipedia, revision 2020-04-11 · Leuchtturmhuman0.0%correct
Polish · 170 wordsWikipedia, revision 2020-05-16 · Wisłahuman33.2%correct

How the corpus was built

The corpus is 18 texts of about 170 words each, six per language. Nine were written by a language model: they are excerpts from articles published on this site, so their origin is not a guess. The other nine are human-written, taken from Wikipedia revisions saved before June 2020 — well before the current generation of models existed, which is what makes them usable as a control. Each text was submitted once, and we recorded the raw percentage returned.

What this test does not show

  • One detector, on its free tier, on a small sample. This is not a market ranking — it is a check on one claim.
  • The paid tools (Sapling, GPTZero, Copyleaks) require creating an account, so they were not tested and we publish no numbers for them.
  • Encyclopaedic prose reads differently from blog writing, and detectors may respond to that register differently.
  • Wikipedia is heavily represented in model training data, which may itself affect how a detector scores it.

Using these results

Cite or reproduce this table with a link back to this page. The human half of the corpus comes from Wikipedia and is available under CC BY-SA; the machine half is our own published writing.