Researchers at OpenAI have found that their state-of-the-art computer vision system can be easily fooled
Researchers at OpenAI call this typographic attacks. They use the model's ability to reliably recognize text. It turns out that even a photo with handwritten text can fool AI.