For model version 2024-11-04-base: These subclass categories are currently in beta.
Starting with our November 2024 AI detector model (version: 2024-11-04-base), we are capable of not only distinguishing AI generated text from human, but also goes one step further: we can classify AI generated texts as “pure_ai” or “ai_paraphrased” text.
Our basic scan JSON responses now also contain a separate “subclass” object that provides information about detected subclasses and their confidence categories and thresholds.
As of this version (2024-11-04-base) we support two more new subclasses under the “ai” classification category: “pure_ai” and “ai_paraphrased”.
Flagging text as AI paraphrased is triggered when confidence scores exceed the specified subclass confidence thresholds (this is contained within JSON response under a new "subclass" key as shown below):
{
# Rest of the API response as usual...,
"subclass": {
"ai": {
"predicted_class": "ai_paraphrased",
"result_message": "We are highly confident that this text has been rewritten by AI, an AI paraphraser or AI bypasser",
"confidence_score": 0.9883339110025372,
"confidence_category": "high",
"class_probabilities": {
"pure_ai": 0.01166608899746283,
"ai_paraphrased": 0.9883339110025372
},
"confidence_scores_raw": {
"identity": {
"pure_ai": 0.01166608889746283,
"ai_paraphrased": 0.9883339110025372
}
},
"confidence_thresholds_raw": {
"identity": {
"pure_ai": {
"reject": 0.7,
"low": 0.8,
"medium": 0.92
},
"ai_paraphrased": {
"reject": 0.85,
"low": 0.9,
"medium": 0.95
}
}
}
},
"human": {},
"mixed": {}
}
}
To see the rest of our API documentation, you can refer to our docs: https://gptzero.stoplight.io/
In beta, our solution can successfully detect AI paraphrased text generated from certain AI tools with a false positive rate of <0.1%. We aim to continuously improve our model to enhance its coverage to all the prevalent AI bypass tools in the market. While our detection system is robust, it is not infallible. We strongly encourage users to adopt our AI authorship tools such as our Docs editor and Origin Chrome extension to transparently verify work.