Methodology — How Our AI Detector Works

This page explains how the AI Detector produces its scores and, just as importantly, where those scores fall short.

The model

Detection is powered by a RoBERTa-based classifier. RoBERTa is a transformer language model; here it is applied to the task of separating AI-generated text from human-written text. When you submit text, it is split into sentences and each is evaluated, producing two outputs:

• an overall AI-probability score from 0–100%;
• a sentence-level breakdown that labels each sentence as AI, mixed, or human, with its own probability.

How to read the score

As a guide: above ~70% the text shows strong AI-like patterns; 40–70% suggests a mix of AI and human writing; below 40% reads as primarily human. These thresholds are interpretive guides, not hard rules.

Accuracy and limitations

AI detection is inherently probabilistic, and no detector is fully reliable. The known limitations of this tool include:

• False positives: plain, formulaic, or simple human writing can be flagged as AI.
• False negatives: AI text that has been paraphrased, edited, or “humanized” can pass as human.
• Short text: accuracy drops sharply on very short passages — we require at least 50 characters, and results are far more stable on longer text.
• Domain & format: performance is strongest on general English prose and weaker on code, lists, or specialised formats.

Because of these limits, the result should be treated as one signal among many. We do not recommend relying on it alone for consequential decisions such as academic discipline or employment.