Methodology & limits
How TrueState reaches a verdict
Each check combines measurements we compute on our server with an analysis by Google's Gemini 3 model. We then calibrate the result: probabilities are clamped between 3% and 97%, every range is at least 10 points wide, and anything straddling the middle is labeled Uncertain.
Likely human
Whole range at or below 40%
Uncertain
Range touches the 40–60% zone
Likely AI
Whole range at or above 60%
Text
What it can do
Spot the habits of language models: stock transitions, uniform sentence rhythm, generic hedging, missing concrete detail. We also measure sentence-length variation and vocabulary diversity.
What it can't
Reliably catch AI text that a person has rewritten, and it can wrongly flag non-native writers or formulaic genres like lab reports.
Images
What it can do
Read EXIF camera data, C2PA content credentials and generator tags left by tools like Stable Diffusion, then have Gemini 3 inspect hands, lettering, lighting, reflections and textures.
What it can't
See through heavy compression, screenshots or small images, which wipe out artifacts. Missing metadata is normal after social media upload.
Audio
What it can do
Listen for synthetic-speech tells: flat prosody, missing breaths, robotic artifacts, room acoustics that don't match, and odd phrase transitions.
What it can't
Unmask a high-quality voice clone with confidence, especially over phone-quality or noisy recordings.
For schools and newsrooms
- ★ Use a result to start a conversation or an investigation — never to end one.
- ★ Ask for drafts, version history, original files or the source recording.
- ★ Reverse-image search and contact the original publisher before you publish.