Meta's Image Detector Defeated by Cropping
filiksyos · x · 2026-07-13
Reuters tested Meta's new AI image detector using samples from Meta's own Muse Image model.
Results showed that while original images could be verified, cropping them to about a third or half of their original size caused the detector to fail 55% of the time. Meta explained this is a preview version and heavy cropping might degrade the signal. The author argues this isn't just a content moderation issue, but an "architectural problem":
- Watermarking/provenance relying on a single hidden marker is fragile.
- Attackers only need to weaken the marker to create uncertainty.
- Platforms need confidence levels, not a binary "AI/non-AI" judgment.
The article advises companies to treat detection, monitoring, watermarking, evals, and audit logs as a holistic "measurement system" and subject them to adversarial testing. Simply "hoping the watermark survives" does not constitute a real provenance strategy.
More from Models
- Claim says Kimi was distilled from Fable, sparking a model-attribution jab — cephaloform · 2026-07-22
- Gary Marcus says LLM math skills are like knowing only a car’s engine size — GaryMarcus · 2026-07-22
- OpenAI’s Codex + GPT-5.6 Sol hits 99% recall in Project APE verification tests — soumitrashukla9 · 2026-07-22
- OpenAI-linked paper says capability RL can make models more reward-seeking — MariusHobbhahn · 2026-07-22
- Macaron V1 adds LoRA RL on GLM 5.2 and claims SOTA benchmark gains — Xianbao_QIAN · 2026-07-22
- OpenAI rolls out voice in GPT-Live, but the UI obscures search and reasoning — Graham_dePenros · 2026-07-22