OpenAI has announced the initiation of new techniques to identify content produced by artificial intelligence (AI). This includes the development of a robust image classifier and tamper-resistant audio watermarking.
Image Classifier for AI-Generated Content Identification
OpenAI’s latest image classifier boasts up to 98% accuracy in discerning whether an image was created by the DALL-E 3 AI image generator. Even in scenarios where images have undergone cropping, compression, or alterations in saturation, the classifier remains effective. However, its capability to recognize content generated by other AI models, such as Midjourney, is notably lower, ranging from 5 to 10%.
Tamper-Resistant Audio Watermarks
In addition to the image classifier, OpenAI has introduced watermarks for audio content produced using its Voice Engine text-to-speech platform, presently in pre-production testing. These watermarks contain vital information regarding the creator and methods utilized in content creation, significantly simplifying authenticity verification.
Participation in C2PA Coalition
OpenAI actively engages in the Content Provenance and Authenticity (C2PA) Coalition, alongside notable companies like Microsoft and Adobe. This month, OpenAI assumed a position on the C2PA steering committee, underscoring its commitment to establishing standards for transparency and authenticity in digital content.
For enhanced authenticity verification, OpenAI has integrated C2PA content credentials into image metadata. These credentials, serving as effective watermarks, furnish information about the image owner and the creation process.
Ongoing Development and User Feedback
The development of the image classifier and audio watermarking technology remains in progress, notes NIXSolutions. OpenAI stresses the importance of receiving user feedback to evaluate the efficacy of these tools. Interested parties, including researchers and non-profit journalists, have the opportunity to test the image classifier through the OpenAI research access platform.
As OpenAI continues to refine its AI content identification tools, we’ll keep you updated on their advancements and effectiveness in ensuring content authenticity.