AI detectorItalian AI text detector
Italian texts only. The model is trained on Italian: a text in English or in any other language gives no reliable result. The page is in English, the text you paste has to be in Italian.
0 words. At least 60 are needed.
How it works
The detector looks at the text in three ways and combines the measures with a classifier trained on Italian texts written by people and by language models.
How predictable the text is. A language model almost always picks the most likely word, a person much less so. Two small models read the text word by word and measure how much they expected each one, with the Binoculars and Fast-DetectGPT methods, published in 2024.
Epanorthosis. It is the upward self-correction, as in «non è un corso, è un percorso» ("it is not a course, it is a journey"). Federico Boggia's paper measures its density genre by genre against the human one, and the patterns it uses are the same ones the detector uses.
Calques from English. Models learn mostly from English, and when they write in Italian they carry it along: «giocare un ruolo» (to play a role), «indirizzare un problema» (to address a problem), «è importante notare che» (it is important to note that), the comma before «e» in the last item of a list, headings with every initial in capitals.
No detector is proof. A text written by a person can look generated, and a generated text rewritten by hand can pass as human. The result is there to raise a question; it is not enough to accuse anyone of copying.
Add the Binatomy AI detector to your service via API
You send the text as JSON and get back the same analysis as this page, with the positions of the sentences and the signs. Choose a plan, pay online and the key arrives straight away, on screen and by email. You can change or cancel the plan yourself, whenever you like. One analysis covers up to 3,000 words. The API, too, reads only texts in Italian.
Base
19 € a month
+ VAT
2.000 analyses a month
For a website, an app or a school that checks a few dozen texts a day.
Pro
49 € a month
+ VAT
10.000 analyses a month
For platforms and newsrooms that check hundreds of texts a day.
Custom
For large volumes or a dedicated installation
The detector on your servers, a calibration on your texts or a volume the plans do not cover. We can talk it through on a call.
Book a callAlready subscribed? Manage it here (page in Italian): analyses used, new key, payment method, receipts and cancellation.
How to call it
curl -X POST https://binatomy.com/api/public/rilevatore-ai \
-H "Authorization: Bearer rai_your_key" \
-H "Content-Type: application/json" \
-d '{"testo": "Your text in Italian, at least 60 words...", "genere": "giornalistico"}'The response:
{
"status": "success",
"probabilita": 0.91,
"fascia": "generato",
"parole": 412,
"famiglie": [
{ "chiave": "prevedibilita", "nome": "Prevedibilità per un modello linguistico", "contributo": 2.4 },
{ "chiave": "calchi", "nome": "Calchi dall'inglese", "contributo": 0.6 }
],
"epanortosi": { "occorrenze": 3, "densita": 72.8, "genere": "giornalistico", "base_umana": 2.4, "indice": 30.33 },
"frasi": [ { "inizio": 0, "fine": 118, "prevedibilita": 0.82 } ],
"evidenziati": [
{ "inizio": 140, "fine": 154, "famiglia": "calco", "chiave": "giocare_ruolo",
"perche": "«giocare un ruolo» ricalca «to play a role»: in italiano si svolge" }
],
"conto": { "piano": "Base", "analisi_usate": 37, "analisi_incluse": 2000, "questa_richiesta": 1 }
}- The key goes in
Authorization: Beareror inX-Api-Key. probabilitagoes from 0 to 1: the higher it is, the more the text looks generated.fasciaisumano(human),incerto(uncertain) orgenerato(generated).genereis optional and feeds the epanorthosis index:argomentativo,giornalistico,enciclopedico,accademico,didattico,narrativo,conversazionale,oratorio,promozionale.inizioandfineare character positions in the text you sent, without leading and trailing spaces.- Errors have
status: "error", amotivo(chiave,quota,corto,lungo,lingua,rete) and amessageto show to the user, in Italian. TheX-Analisi-Rimasteheader says how many analyses are left in the month. - The text is not stored: it is analysed and thrown away.
Use it inside Claude
The detector is also a connector for Claude, and there it helps you write. After a draft in Italian, Claude checks the style signs, rewrites only the flagged passages and checks again, until the epanorthosis index sits at the human rate for the genre. The goal is the paper's: bring the figure down to the human rate, without removing it altogether.
The connector address:
https://binatomy.com/api/mcp/rilevatore
- How to add it
- In Claude: Settings, then Connectors, then "Add custom connector", and paste the address.
- The style check
- It is free and asks for nothing: it finds epanorthosis, stock phrases and calques, with the piece of text and the reason.
- The full analysis
- It uses the analyses of your API subscription. In the connector, add the
X-Rilevatore-Chiaveheader with your key. - What to ask
- "Write a post in Italian about our new service and run it through the detector before giving it to me", or "Check the style of this Italian newspaper article and tell me what you would change".
Domande frequenti
Does it work on texts in English?
No. The detector is trained on texts in Italian written by people and by language models, and its signals, such as calques from English, only make sense in Italian. On a text in another language the result means nothing.
How does it tell whether an AI wrote the text?
It measures how predictable the text is to two small language models, counts self corrections of the «not a course, a journey» kind and calques from English, then combines the measures with a classifier trained on Italian texts written by people and by models.
Is the result proof that a text was generated?
No, and no detector is. A person can write like a model, and a generated text edited by hand can pass as human. The result is a reason to reread the flagged sentences, not a verdict on its own.
Is it free?
Up to 500 words per text, yes, with ads; the text needs at least 60 words. Above 500 the extended analysis reads the whole text and is paid online per use, with no account. For software there is an API with a key on a monthly plan.
Where does my upload go?
To a server of ours, and we say so plainly because this is one of the few tools that cannot run inside the browser: the processing needs a real machine. The file is uploaded, processed, and deleted automatically within an hour. Nobody opens it, it does not end up in an archive, and it is never used to train anything.
Do you use it for anything else?
No. The file does the job you asked for and then it is gone: we do not link it to you, we do not resell it, we do not pass it to third parties and we do not train any model on it. Most other Binatomy tools upload nothing at all and work only in your browser; this is one of the few that cannot.
Do you profile me? Are you GDPR compliant?
We do not profile you: there is no account to create, we never ask for your email to use a tool, and we keep no profile of what you do. We are an Italian company and we work under the GDPR and the other European rules, with the privacy notice and the consent request in plain sight. One thing we say in full rather than hiding it behind a slogan: the ads are Google's, and Google may use cookies to personalise them. You decide from the consent banner: if you say no the ads stay, but they are not personalised, and the tool works exactly the same.
If it is free, how do you make money?
From advertising, and that is all. We do not sell your data, there is no hidden paid plan, there is no limit that eventually asks for your card, and you do not have to sign up. The ads pay for the servers and for the time we spend writing the tools: that is how we can keep them open to everyone.