Meta AI detector struggles to identify cropped images

Meta unveiled a watermark called Content Seal that it embeds invisibly in images generated by its new Muse Image model. The company also released a preview detection tool that claims to recognise the watermark even after editing such as cropping, resizing, compression or screenshots. Reuters tested the system with 40 Muse‑generated pictures and found it flagged all unaltered images correctly, but when the images were cropped to one‑half or one‑third of their original size the tool identified only 55 % of them as AI‑generated. Meta said the detector is still in a preview stage and will be refined. The findings highlight challenges in reliably spotting AI‑generated visual content, especially as the volume of deepfake images continues to rise sharply worldwide.