OpenAI has created a software tool to help detect images created by its text-to-image AI DALL-E. It’s designed to tackle the growing issue of deepfakes. OpenAI is also the company behind the generative AI ChatGPT chatbot.
DALL-E Detection Classifier
The new tool is called the DALL-E Detection Classifier. OpenAI says that the software is a binary classifier designed to predict the likelihood that an image was created by the DALL-E 3 model. It reports a binary true/false outcome. You also get a numerical score to indicate how likely it is that an image was AI-generated. OpenAI says that the deepfake detector will correctly identify 98.8 percent of images created by DALL-E 3. However, the company says it only “flags approximately 5 to 10 percent of images generated by other AI models.”
How it works
The Detection Classifier works by reading metadata within the images. OpenAI says it has added tamper-resistant metadata to the images created and edited by DALL-E 3. The detector therefore reads the metadata to prove the content comes from a particular source. As such, it knows when an image was generated by DALL-E 3, leading it to deem the image to be artificially generated.
C2PA standard
Along with Google and Meta, OpenAI is joining the steering committee for the Coalition for Content Provenance and Authenticity (C2PA). OpenAI said that the metadata it has used with DALL-E 3 follows the standard for digital content certification set by the C2PA. This standard has also been likened to the nutrition labels found on food. In addition, OpenAI has confirmed that its AI video generator Sora will also add the C2PA metadata to its files.
When will the deepfake detector be released?
OpenAI is currently recruiting researchers to help the company “stress test the classifier across diverse image types, such as medical images and photorealistic simulations.” It also wants testers to investigate potential biases, and how well the classifier performs compared to humans at identifying AI-generated content. However, applications don’t close until the end of July, and researchers won’t find out if they have been accepted until the end of August. As such, the DALL-E Detection Classifier won’t be rolled out any time soon.
What we think
Although generative AI has a lot to offer, there is growing concern over the issue of deepfakes. With an election in the US later this year, AI companies are therefore keen to be seen to be trying to tackle the problem. OpenAI acknowledges that the metadata isn’t a complete solution as it could be removed from an image. However, the company says that it isn’t easy to fake or alter the metadata, so it can help to identify AI-generated files. It’s also important to give OpenAI credit for the steps they are taking. Hopefully, the research will lead to more robust solutions to tackle deepfakes in the future.
