Meta AI Detector Fails on 55% Cropped Images in Reuters Test

On July 10 Reuters found Meta’s new AI-image detector failed to identify 55% of cropped versions from its Muse model, exposing limits in watermark reliability ahead of U.S. midterms. The setback follows Meta’s July 9 admission that employee keystroke data from its Model Capability Initiative was moved to an unauthorized internal location, prompting an immediate pause and investigation by CTO Andrew Bosworth. Earlier the same day, a July 2 patent filing surfaced for an AI wearable designed to track sighs, laughter, voice tone and medication intake, amplifying scrutiny over Meta’s expanding data practices.