OpenAI working on new AI image detection tools

OpenAI has added a caller instrumentality to observe if an image was made pinch its DALL-E AI image generator, arsenic good as new watermarking methods to much intelligibly emblem contented it generates.

In a blog post, OpenAI announced that it has begun processing caller provenance methods to way contented and beryllium whether it was AI-generated. These see a caller image discovery classifier that uses AI to find whether nan photograph was AI-generated, arsenic good arsenic a tamper-resistant watermark that tin tag contented for illustration audio pinch invisible signals. 

The classifier predicts nan likelihood that a image was created by DALL-E 3. OpenAI claims nan classifier useful moreover if nan image is cropped aliases compressed aliases nan saturation is changed. While nan instrumentality tin observe if images were made pinch DALL-E 3 pinch astir 98 percent accuracy, its capacity astatine figuring retired if nan contented was from different AI models is not arsenic good, flagging only 5 to 10 percent of pictures from different image generators, for illustration Midjourney.  

OpenAI antecedently added content credentials to image metadata from nan Coalition of Content Provenance and Authority (C2PA). Content credentials are fundamentally watermarks that see accusation astir who owns nan image and really it was created. OpenAI, on pinch companies for illustration Microsoft and Adobe, is simply a personnel of C2PA. This month, OpenAI besides joined C2PA’s steering committee. 

The AI institution besides began adding watermarks to clips from Voice Engine, its text-to-speech platform presently successful constricted preview.

Both nan image classifier and nan audio watermarking awesome are still being refined. OpenAI says it needs to get feedback from users to trial its effectiveness. Researchers and nonprofit publicity groups tin trial nan image discovery classifier by applying it to OpenAI’s investigation entree platform. 

OpenAI has been moving connected detecting AI-generated contented for years. However, successful 2023, it had to extremity a programme that attempted to identify AI-written text because nan AI matter classifier consistently had debased accuracy.