OpenAI, the company behind ChatGPT, is working on a new tool that will help you detect if an image has been created by DALL-E, their artificial intelligence model that generates images from descriptions.
The announcement, published on the official OpenAI blog, states that they have started developing new methods to track AI-generated content.
To begin, the company has announced its participation in the C2PA (Coalition for Content Provenance and Authenticity), an initiative that seeks to combat misinformation and enhance online transparency. The C2PA protocol enables creators to tag and certify their content to verify its true origin.
OpenAI will use this protocol in the metadata of all images created and edited by DALL-E 3. It will also be incorporated into videos created with Sora when the model is publicly released.
On the other hand, the company has also announced the arrival of a new detector that will be used to determine if a photo has been created by AI or not. Specifically, it predicts the probability that an image has been created by DALL-E 3.
OpenAI claims that the tool works even if the image is cropped, compressed, or even if the saturation has been altered. Although it has an accuracy of 98% for images created with DALL-E, the story changes with other models, such as Midjourney. There, its reliability can drop to 5%. On the other hand, they have also started adding watermarks to clips created with Voice Engine, their text-to-speech conversion platform.
However, it is important to keep in mind that the tools proposed by OpenAI are far from being in their final version. In fact, the company has opened up applications to test the image detector and verify its effectiveness firsthand.
Although OpenAI has been working on the creation of detection tools for years, not all of their projects have been successful. In 2023, they had to abandon a program that would identify text written by AI. Its low accuracy prevented it from moving forward.
Source: The Verge
