Tuan Hoang

Tuan Hoang leads detection at amige. He sets how the detector panel is assembled and scored, and reviews every model guide, glossary entry, and comparison on this site for accuracy.

The panel is heterogeneous on purpose. When several detectors are trained on different data, their errors are partly uncorrelated, so combining them is more robust than trusting any one. He keeps the roster mixed across vendors for that reason, scores a scan by reading where the detectors agree and where they split rather than averaging them into one number, and weights the loudest deepfake signal instead of letting a confident classifier get diluted. The wide “uncertain” band in the middle is deliberate: amige. reports that it isn’t sure when it isn’t.

He also sets the claims line. amige. doesn’t headline a single accuracy number, because every detector in this category advertises around 99% and measures lower in independent tests, and one number on a homepage invites the overconfident misuse the product is built to avoid. Every verdict reads as a probabilistic estimate, attribution stays a best guess (“looks like Midjourney v6”, with the question mark), and the text surfaces name the higher false-positive risk on non-native-English writing.

His focus is the gap between lab benchmarks and real-world detection: where detectors break on compressed, cropped, and re-shared media, where a paraphraser or humanizer defeats text detection, and why a panel of independent classifiers fails less often than any single one. The articles here cite primary sources, from the RAID benchmark to the C2PA specification, and name their limits on the page.

Every article in the amige. corpus carries a “last reviewed” date. When a model ships a new version or a benchmark updates, the affected guides get re-checked against primary sources, and the date moves with the review.