An AI detector is software that estimates how likely it is that a text was written by a language model such as ChatGPT, Gemini or Claude: it returns a probability, not proof. In our test on 28 October 2026, run on 20 Italian texts (half written by people before 2020, half generated by three AI models), three open-source detectors caught 0, 2 and 9 of the 10 AI texts, and the one that caught the most also flagged 3 of the 10 human texts as AI.
In short: the same text can come out "human" or "AI" depending on the tool, and outside English the margin of error is wider than vendor home pages suggest.
Paste one paragraph into three detectors and you will often get three different answers.
So which one do you believe?
None of them, if the question is "who wrote this". The two mistakes we see most often:
- taking the "99% accuracy" on the home page at face value, a figure nobody can reproduce;
- using the score to decide what to publish, as if Google rewarded texts that "pass" a detector.
The previous version of this page was a list of 15 tools with accuracy figures supplied by their makers. We rewrote it around a real test, because those figures told readers nothing about how the tools behave on their own content.
Macropix, which makes custom LED screens, reached first position in Italy for "ledwall" (4,400 Italian searches a month) with one page per product line (the full story is in our piece on SEO for manufacturers). Sunscape, which develops solar and agrivoltaic plants, went from 329 to 5,809 organic impressions a month in six months from March 2026.
What has that got to do with detectors?
Neither Google nor the AI answer engines asked those pages to "sound human". They ask for pages that are useful and verifiable. In this article:
- our test on 20 texts, with method and results;
- what independent studies found, from OpenAI to Stanford;
- what Google and AI search look at instead of a detector score.
Let's get into it.
What AI detectors are (and what they actually measure)
An AI detector is a statistical classifier: it compares a text with what a language model would have written and estimates the chance that it came from a machine. It cannot see the document's history and does not know who wrote it.
- Trained classifiers: a model (often RoBERTa) is trained on millions of texts labelled "human" or "AI". They do well on text that resembles their training data and much worse on anything else.
- Zero-shot methods: they measure how predictable a text is for a language model, its perplexity. AI text tends to be more predictable. Binoculars, from the University of Maryland, compares two models to cut false alarms.
One point almost everyone gets wrong. The score is a probability, not a proportion. The model card for Fakespot, the Mozilla lab, says so plainly: "80% AI" means an 80% chance the text is AI, not that 80% of it was generated.
What about watermarks?
Different thing. Article 50 of the EU AI Act requires providers of systems that generate text, images, audio or video to mark outputs "in a machine-readable format". That is tracking provenance at the source, not guessing it afterwards. We cover publishers' obligations in our piece on AI-generated content.
The test: 20 Italian texts, 3 detectors (nothing edited)
We ran three open-source detectors on 20 Italian texts whose origin we know for certain. It is a small test built to be repeatable: anyone can rerun it with the same texts and tools. We used Italian on purpose: most detectors are trained on English, and this shows what happens when they leave it.
The texts
- 10 human texts: the opening of 9 Italian Wikipedia articles as saved in November and December 2019, before ChatGPT existed (from the Colosseum to railways), plus the first lines of chapter I of Manzoni's The Betrothed, 1840 edition, from Wikisource.
- 10 texts generated on 28 October 2026 on the same topics with the same prompt ("about 150 words, encyclopaedic style, prose only"): 3 by GPT-5.5 and 1 by GPT-4o-mini from OpenAI, 3 by Google's Gemini 3.6 Flash, 3 by Anthropic's Claude Sonnet 5. For the novel, we asked Claude for the opening of a historical novel set on Lake Como in the 1600s.
Human texts run from 75 to 200 words, AI texts from 133 to 160. Nothing was edited apart from standardising apostrophes.
The detectors
- RADAR, presented at NeurIPS 2023 by researchers from the Chinese University of Hong Kong and IBM Research (paper).
- Binoculars, a zero-shot method from the University of Maryland (ICML 2024), using its low false-positive threshold.
- Fakespot AI Text Detector, a RoBERTa-base model trained by Mozilla's Fakespot lab.
We queried their public demos on Hugging Face, with a 50% threshold. Fakespot's model card states it was trained on English text, and the RADAR and Binoculars papers report results on English datasets.
And GPTZero, ZeroGPT, Originality.ai?
Not in this test: querying them repeatably requires a paid API subscription. We cover what can be verified about them further down.
The results: 0 to 9 AI texts caught out of 10
| Detector | AI texts caught (of 10) | Human texts flagged as AI (of 10) | Correctly classified (of 20) |
|---|---|---|---|
| RADAR | 0 | 0 | 10 |
| Binoculars | 2 | 1 | 11 |
| Fakespot | 9 | 3 | 16 |
Three tools, three opposite behaviours.
RADAR called all 20 texts human, never giving more than a 4.2% chance of AI. No false positives, but only because it never said "AI".
Binoculars was cautious: it caught GPT-4o-mini's marketing text and Claude's Lake Como text, and got one 2019 Wikipedia article wrong, the one on railways.
Fakespot was the only one to catch almost every AI text: 7 of 7 for GPT and Gemini, 2 of 3 for Claude. But it gave a probability above 90% to two Wikipedia articles written by people in 2019.
| Text | Written by | AI probability (Fakespot) | Verdict |
|---|---|---|---|
| Olive oil (Wikipedia, 14/11/2019) | People | 92% | Wrong |
| Lake Como (Wikipedia, 30/12/2019) | People | 90% | Wrong |
| Leonardo da Vinci | Claude Sonnet 5 | 12% | Wrong |
| The Betrothed, ch. I (1840) | Alessandro Manzoni | 21% | Right |
| Opening of a novel set on Lake Como | Claude Sonnet 5 | 93% | Right |
Why those two Wikipedia articles?
They are the most regular texts in the sample: definitional sentences, neutral vocabulary, no digressions. That is exactly the style language models imitate, and the style the detector learned to call "AI". The tidier and more impersonal a human text, the higher the risk of a false flag.
The limits, stated up front: 20 texts is a small sample, several human texts are under 1,000 characters (the length below which OpenAI called its own detector "very unreliable"), and we tested open-source tools rather than commercial ones. The direction, though, matches the larger studies.
What independent studies found (and why "99%" does not hold up)
The published studies are almost all on English texts. They say the same thing as our test, on much larger samples.
- OpenAI withdrew its own detector on 20 July 2023. It identified 26% of AI texts as "likely AI-written" and wrongly flagged 9% of human texts, and OpenAI said it performed "significantly worse" in languages other than English (OpenAI, 2023).
- On 91 TOEFL essays by non-native English speakers, 7 detectors produced an average false-positive rate of 61.22%, and 18 essays were flagged by all seven. Essays by native-speaking US students were classified almost perfectly (Liang et al., Stanford, published in Patterns, 2023).
- 14 tools tested, none "accurate or reliable", with a bias towards classifying text as human and a sharp drop on reworked text (Weber-Wulff et al., International Journal for Educational Integrity, 2023).
- More than 6 million texts, 11 generative models, 12 detectors: the RAID benchmark found detectors "easily fooled" and questions 99% accuracy claims made without shared data (Dugan et al., ACL 2024).
So where does "99%" come from?
From internal tests on data chosen by the vendor. A number you cannot reproduce is not data, it is a slogan. Binoculars is the clearest example of the gap: its paper reports catching more than 90% of ChatGPT texts at a 0.01% false-positive rate on its own benchmarks. On our Italian texts it caught 2 out of 10.
A wrong score costs more than a text written with AI.
In our test, two Wikipedia articles from 2019 scored 90% "AI". If you reject content or suppliers on that number, you are measuring the wrong thing. We measure what matters: whether your site shows up on Google and in ChatGPT answers, and who gets cited instead of you.
Book a strategy callDo they work worse outside English? (yes, and almost nobody measures it)
Almost certainly, for a simple reason: most detectors are trained mainly on English. Fakespot's model card says so, and OpenAI said the same of its own classifier.
Commercial services advertise multilingual support (Compilatio claims more than 35 languages). But we could not find a single published independent study on detector accuracy for Italian texts, and the language figures in circulation come from the vendors themselves. The Stanford study adds a warning for anyone assessing English written by non-native speakers: that is where false positives cluster.
Which AI detector to use (and for what)
It depends on the job.
- If you run an editorial team or buy content: a detector can point to a piece worth a closer look, never replace reading it. Proof of who wrote what is the version history, not a score.
- If you mark essays or theses: with OpenAI's 9% false-positive rate, a cohort of 100 students means 9 wrong accusations.
- If you do SEO: you do not need one. Google does not use detector scores to rank pages, as the next section shows.
And if I have to pick one anyway?
Use at least two, on texts over 1,000 characters, pasted without formatting. If they disagree, and in our test at least two of the three tools contradicted each other on 12 of the 20 texts, the honest answer is "we don't know".
Do not confuse these tools with the ones that measure how often a brand appears in ChatGPT and Gemini answers: for those, see our guide on how to choose an AI visibility tool.
AI detectors and SEO: what Google looks at instead
Google does not penalise text for being written with AI. In guidance published on 8 February 2023 it says it rewards "high-quality content, however it is produced", and that using automation "with the primary purpose of manipulating ranking" is spam.
The test is purpose, not the tool. We collected the data on how AI content gets indexed and ranked in AI content and SEO, and how to show quality in our piece on E-E-A-T.
The same goes for AI answers. There is no evidence that AI Overviews, AI Mode or ChatGPT run pages through a detector before citing them. What makes a page citable is a clear answer up front, sourced data and a fact nobody else has. In our study of the sources ChatGPT and AI Overviews cite, the pages cited were often from companies in the sector with specific information, not the most "human-sounding" ones.
So does a test like this help there too?
Yes, and it is why we ran it: original data is the kind of content an AI can cite and cannot make up. The full picture on measuring presence in AI answers is in our guide to AI visibility.
How to use them without doing damage (five rules)
- Never as proof. A score is not enough to reject a text, a supplier or a piece of coursework.
- Long texts. Under 1,000 characters even OpenAI considered its tool very unreliable.
- Clean text. Paste without HTML or formatting: in a 2024 test by Textbroker on German texts, formatting alone moved the same text from 0% to 100% AI.
- Two tools, not one. If they disagree, the score is worthless.
- Ask for drafts. Version history, notes and sources say who wrote what better than any model.
And for publishers: do not write to "pass" a detector. A text rewritten to lower its score does not become more useful to the person reading it.
FAQs
No, not enough to base a decision on. In our test on 20 Italian texts, three open-source detectors caught 0, 2 and 9 of 10 AI texts, and the most sensitive flagged 3 of 10 human texts as AI. Independent studies in English, from OpenAI to Stanford, reach the same conclusion.
Yes. In our test two Wikipedia articles saved in 2019, before ChatGPT existed, scored 90% and 92% AI probability with Fakespot. Tidy, definitional and impersonal writing is most at risk.
Not reliably, as far as anyone has shown. Commercial services claim multilingual support, but we found no published independent study for Italian, and OpenAI said its own classifier performed significantly worse outside English.
No. Google says it rewards quality content however it is produced, and treats as spam only automation used to manipulate rankings. It does not use detector scores to rank pages.
It depends: for English, independent tests favour different tools for different kinds of text, and for other languages nobody has verified data. If you use one, compare it with a second and treat the result as a hint.
So?
A detector estimates, it does not prove.
On 20 Italian texts: 0, 2 or 9 AI texts caught out of 10, depending on the tool.
Two Wikipedia articles from 2019 scored 90% "AI".
For Italian, we found no published independent studies.
Google judges a page by its purpose, not the tool that wrote it.
The question "who wrote this" is moving from the text to its source. OpenAI withdrew its detector saying it would work on provenance, and the AI Act asks providers to mark outputs at the source. In a few years we will ask a file where it came from, not ask a model what it thinks. In the meantime, if you want to know whether your site appears where customers search, we measure it through our AI monitoring and GEO services.
Speak soon!
