Methodology

Understanding our metrics

Every AITrustLens analysis returns several indicators. Here is exactly what they measure, how to read them — and above all, their limits. Being transparent about what detection can and cannot do is core to our approach.

AI probability

What it measures: The probability (0–100%) that the content was AI-generated, estimated by neural networks trained on millions of real and generated examples.
How to read it: Above 80%: very likely generated. Between 40–60%: uncertainty zone — don't conclude. Below 20%: very likely authentic.
Limits: The newest generators (Midjourney v7, Flux…) regularly defeat the best detectors. Highly stylized photos (bokeh, filters) can trigger false positives.

Confidence

What it measures: How sure the model is of its own verdict. High when our different models agree with each other, low when they diverge.
How to read it: An 85% AI probability with 60% confidence deserves more caution than 75% probability with 95% confidence.
Limits: Confidence measures internal model agreement, not absolute truth: models can be unanimous… and wrong together.

% of document (text)

What it measures: For text, we split the document into segments and score each one separately — like academic tools do. The percentage shows how much of the document reads as AI-written, with suspect passages highlighted.
How to read it: An essay at 90% AI with every segment highlighted tells a different story than a 30% text where only the introduction is suspect.
Limits: Text detection remains the least reliable modality: frequent false positives on non-native speakers, technical or heavily edited writing. NEVER use it as sole proof against a student.

Source authority score

What it measures: For articles, a 0–100 score reflecting the editorial reliability of the publishing domain (news agencies, scientific journals, established media, blogs, social networks…).
How to read it: Reuters ≈ 95, a scientific journal ≈ 98, an anonymous blog ≈ 30. Always cross-check a low score against the alternative sources we suggest.
Limits: Domain authority doesn't guarantee the truth of ONE specific article — big outlets get things wrong, small blogs can be right.

Manipulation heatmap (images)

What it measures: Our forensics model (MMFusion-IML, 2024 academic research) locates retouched areas in an image: added, removed or moved elements. The result is a heatmap overlaid on the image.
How to read it: Red zones concentrated on a face or a specific object indicate targeted retouching. Diffuse noise across the whole image is less meaningful.
Limits: Compression (social networks, screenshots) degrades manipulation traces and can create artifacts. Heavily recompressed images give less reliable results.

Golden rule: none of these indicators is proof. They are signals meant to guide human verification — cross-checking sources, context, common sense. That's why every analysis also includes alternative source suggestions and fact-checking leads.

See these metrics in action

Analyze content See analyzed examples