Updated
Updated · Gizmodo · Jul 11
Meta AI Detector Misses Over 50% of Cropped Muse Images
Updated
Updated · Gizmodo · Jul 11

Meta AI Detector Misses Over 50% of Cropped Muse Images

3 articles · Updated · Gizmodo · Jul 11

Summary

  • Reuters found Meta’s preview detection tool identified all 40 Muse Image outputs at full size but flagged only 55% after they were cropped to half or one-third.
  • That undercuts Meta’s claim that its invisible Content Seal watermark would survive cropping, compression, resizing and screenshots and remain detectable.
  • The miss comes as AI-generated deepfakes are surging — cybersecurity firm DeepStrike estimates roughly 900% annual growth from 2023 to 2025 — while detection tools still make frequent errors.
  • For Meta, the setback clouds a key AI launch in a catch-up push backed by multibillion-dollar investment and follows privacy backlash over Muse Image using public Instagram photos without explicit consent, a feature since removed.

Insights

Meta is spending $145 billion on AI, so why can its new tool be defeated by a simple image crop?
As Meta's own AI detector falters, is the concept of online trust becoming a relic of the past?
With deepfake creation costing just a dollar, are multi-billion dollar detection efforts fighting a losing battle?