Analyzed examples

What our analyses reveal, concretely

Three representative cases run through AITrustLens: a generated portrait, a chatbot-written essay, and a locally retouched photo. For each: what we submitted, what the analysis returned, and how to interpret it.

Image

"Photo" portrait generated with Midjourney

A photorealistic portrait posted on social media, presented as a real person's photo.

Verdict : 96% AI

AI probability: 96% — confidence 91% (both image models agree)
Likely source: Midjourney (generator attribution)
EXIF: no camera metadata — consistent with generation

How to read it

Three independent signals converge (detection, attribution, missing EXIF): we can state with high assurance this "portrait" is not a photo. The convergence is what makes the verdict solid, not any single score.

Test the same type of content

Article

Essay written with a chatbot, hand-edited

An 800-word argumentative text whose introduction and conclusion were rewritten by the author, the rest generated.

Verdict : 68% of document flagged AI

Body: segments highlighted red (very uniform writing, smooth vocabulary)
Introduction and conclusion: not highlighted (human rhythm variations)
Overall confidence: 74%

How to read it

Per-segment percentages tell the story better than a global score: the human/AI mix is visible passage by passage. It's a starting point for a conversation, not proof of fraud — text detection remains the most fallible modality.

Test the same type of content

Forgery

Authentic photo… locally retouched

A press photo where a background element was digitally erased before publication.

Verdict : Localized manipulation detected

AI (generation) probability: 12% — the image is NOT generated
Manipulation heatmap: red zone concentrated on the erased region
Manipulation probability: 81%

How to read it

The subtlest case: a real photo can lie. Generation detection says "authentic", but forensics precisely locates the retouch. That's why we combine both verdicts on every image.

Test the same type of content

These examples are representative of product output on typical cases. Exact scores vary with quality, compression and generator versions — which our examples honestly show too.

Your turn to test

Open the analyzer Read the methodology