Sliced, rotated, emoji-covered: AI image detector Pangram easily fooled
goodside · x · 2026-11-14
goodside tested the Pangram AI-image detector by slicing a ChatGPT-generated image into four pieces, randomly transforming and rotating each with a different color filter, then layering them over a real nature photo partially covered by garish text and emojis. The detector failed — in the screenshot it only marks three dashed regions and the overall score leans toward "No AI signal".
A concrete demonstration that mainstream AI-image detection has poor robustness to trivial adversarial edits, undermining confidence in detector-based AI content moderation.
More from Fun
- New Yorker interview: AI is a brute-force problem that will 'cook the planet' — kevinnbass · 2026-11-14
- LLMs playing AoE2 keep mistaking gold piles for berries, showing vision is unsolved — max_paperclips · 2026-11-14
- AI-Generated Retro Mall Video Looks Eerily Like Real 1980s Footage — heresalexandria · 2026-11-14
- Yacine jokes about being 'washed': 'my contributions are in my experience' — yacineMTB · 2026-11-14
- NetHack transpiler contest: serteal first to hit 100% on all public test sessions — davidbau · 2026-11-14
- "By Opus 5.5": viral painting of a knight robot in an autumn forest fuels model rumors — repligate · 2026-11-14