OpenAI engaged on new AI picture detection instruments

OpenAI has added a brand new device to detect if a picture was made with its DALL-E AI picture generator, in addition to new watermarking strategies to extra clearly flag content material it generates.

In a blog post, OpenAI introduced that it has begun creating new provenance strategies to trace content material and show whether or not it was AI-generated. These embody a brand new picture detection classifier that makes use of AI to find out whether or not the photograph was AI-generated, in addition to a tamper-resistant watermark that may tag content material like audio with invisible indicators. 

The classifier predicts the chance {that a} image was created by DALL-E 3. OpenAI claims the classifier works even when the picture is cropped or compressed or the saturation is modified. Whereas the device can detect if photographs had been made with DALL-E 3 with round 98 p.c accuracy, its efficiency at determining if the content material was from different AI fashions just isn’t nearly as good, flagging solely 5 to 10 p.c of images from different picture mills, like Midjourney.  

OpenAI beforehand added content credentials to image metadata from the Coalition of Content material Provenance and Authority (C2PA). Content material credentials are basically watermarks that embody details about who owns the picture and the way it was created. OpenAI, together with firms like Microsoft and Adobe, is a member of C2PA. This month, OpenAI additionally joined C2PA’s steering committee. 

Each the picture classifier and the audio watermarking sign are nonetheless being refined. OpenAI says it must get suggestions from customers to check its effectiveness. Researchers and nonprofit journalism teams can take a look at the picture detection classifier by making use of it to OpenAI’s analysis entry platform. 

OpenAI has been engaged on detecting AI-generated content material for years. Nevertheless, in 2023, it needed to finish a program that tried to identify AI-written text as a result of the AI textual content classifier persistently had low accuracy.