Meta’s AI detector can’t catch its own cropped fakes

Meta's AI detector, designed to identify images generated by its Muse Image model, fails to detect over half of the images once they are cropped, revealing limitations in its watermarking approach. The invisible watermark, Content Seal, is intended to survive common edits but can be lost with heavy cropping, a common image modification. This highlights challenges in relying solely on watermarking for detecting AI-generated images, a problem shared by other companies like Google and OpenAI.