OpenAI has added a new tool to detect whether an image was generated by the DALL-E image generation model, as well as a new way to place a clear watermark in content generated by the model.
In a blog post, the AI startup announced that it has begun developing new methods to track content and prove whether it was created by artificial intelligence.
This includes a new image recognition classifier that uses AI to determine if an image was AI-generated, as well as a tamper-evident watermark that can be placed on content such as audio with an invisible cue.
The classifier predicts the probability of generating an image via DALL-E 3. OpenAI claims that the classifier works even if the image is cropped, compressed, or its saturation changes.
The tool can detect whether an image was generated by DALL-E 3 with 98% accuracy, but it does not perform well at detecting whether content was generated by other image generation models.
OpenAI has already added content identification information to the metadata of C2PA Alliance images. Content IDs are watermarks that contain information about who owns the image and how it was created.
OpenAI is a member of the C2PA alliance alongside Adobe and Microsoft, and this month the AI startup joined the C2PA steering committee.
OpenAI has also begun watermarking clips on its audio-enabled text-to-speech platform.
The AI startup continues to improve its image classifiers and audio watermarking signals.
OpenAI says it needs user feedback to test its effectiveness. Researchers and nonprofit news groups can test image recognition classifiers by submitting them through the OpenAI Research Access platform.
OpenAI has been working on recognizing AI-generated content for years, but in 2023 it had to remove a program that attempted to recognize AI-generated text because its AI text classifiers were consistently less accurate.