OpenAI has developed new tools to detect if an image was created using its DALL-E AI image generator, as well as new watermarking methods to more clearly indicate content generated by AI. In a recent blog post, the company announced the development of provenance methods to track content and verify whether it was AI-generated. This includes an image detection classifier that uses AI to determine if a photo was created by DALL-E, as well as a tamper-resistant watermark that can tag content like audio with invisible signals. The classifier can predict the likelihood that a picture was made with DALL-E 3, even if the image is cropped, compressed, or has changes in saturation. While it can detect images made with DALL-E 3 with around 98 percent accuracy, its performance at identifying content from other AI models, such as Midjourney, is not as strong, only flagging 5 to 10 percent of pictures.
OpenAI has also added content credentials to image metadata, in collaboration with the Coalition of Content Provenance and Authority (C2PA). Content credentials serve as watermarks that contain information about the image’s ownership and creation process. OpenAI, along with companies like Microsoft and Adobe, is part of C2PA, and recently joined the organization’s steering committee. The image classifier and audio watermarking signal are still in the refinement stage, with OpenAI seeking feedback from users to test their effectiveness. Researchers and nonprofit journalism groups can test the image detection classifier by using OpenAI’s research access platform.
OpenAI has been focusing on detecting AI-generated content for several years. In 2023, the company had to discontinue a program that aimed to identify AI-written text due to consistently low accuracy by the AI text classifier. Despite this setback, OpenAI has continued to make progress in developing tools to identify AI-generated content, such as the image detection classifier and audio watermarking signal. The company is committed to refining and improving these tools based on user feedback and testing.
The new image detection classifier and audio watermarking methods are part of OpenAI’s efforts to enhance transparency around AI-generated content. By tracking the provenance of images and content, the company aims to provide users with more information about how images were created and who owns them. This increased transparency can help combat issues like misinformation and deepfakes by enabling users to verify the authenticity of content generated by AI.
In addition to the technical development aspect, OpenAI is also focusing on collaboration with industry partners and organizations like C2PA to establish standards around content provenance and authority. By working together with other companies and stakeholders, OpenAI aims to create a more robust framework for tracking the origin of content and ensuring its credibility. Joining C2PA’s steering committee demonstrates OpenAI’s commitment to contributing to the development of best practices and standards in the field of content authenticity.
Overall, OpenAI’s new tools for detecting AI-generated content, such as the image detection classifier and audio watermarking methods, represent significant advancements in enhancing transparency and accountability in the digital landscape. These tools have the potential to empower users to verify the authenticity of images and content created by AI, ultimately helping to build trust and combat misinformation in an increasingly AI-driven world. Through ongoing refinement and collaboration with industry partners, OpenAI is working to establish standards and practices that promote transparency and authenticity in AI-generated content.