- OpenAI is developing a classification tool that can detect whether DALL-E 3 created an image.
- This tool is 98% accurate, but only works on images created with DALL-E 3.
- OpenAI invited researchers to help further develop the tool.
Advances in generative artificial intelligence over the past few years have made it even more difficult to distinguish between images created by bots and real art. OpenAI is acutely aware that as technology improves, it will become more difficult to tell the fake from the real over time.
The company announced that it is working on a tool that can detect whether a generative AI created an image. Please note here that OpenAI is developing this tool only to detect whether its DALL-E 3 model was used to create an image, and currently The tool is that he's not 100% accurate.
“This classifier accurately identifies images generated by DALL・E 3 and does not trigger for non-AI generated images. Approximately 98% of DALL・E 3 images are correctly identified; Approximately less than 0.5% of non-AI-generated images were incorrectly tagged as DALL・E 3, while the classifier made minimal common changes such as compression, cropping, and saturation changes. Processing. Performance Impact. However, other changes may reduce performance,” OpenAI wrote.
Changes that affect this performance include adjusting the image hue and introducing a “moderate amount of Gaussian noise.” However, in most cases, the classifier is very accurate, even when the actual images are labeled as being created by DALL-E 3. Of course, we should point out that OpenAI has been burning detection information into his DALL-E3 since the beginning of this year.
OpenAI says it is making applications available to research institutions, journalists, and other testers to access this image detection classifier. The goal is to allow independent research to evaluate the effectiveness of the tool and help make improvements. If you are interested in testing classifiers, please visit here.
As mentioned above, this tool is not intended for use with other image generation tools, but OpenAI says that this classifier flagged 5 to 10 percent of images generated by other models in the dataset. is.
“In addition to investing in C2PA, [Coalition for Content Provenance and Authenticity], OpenAI is also developing new provenance methods to enhance the integrity of digital content. This includes implementing tamper-evident watermarking (marking digital content such as audio with an invisible signal that is intended to be difficult to remove) and using artificial intelligence to ensure that content is generatively modeled. Contains an implementation of a detection classifier, a tool that evaluates the likelihood of origin. “These tools are intended to increase resistance to attempts to remove signals about the origin of content,” he wrote the AI company.
While this is good news, the obvious question is whether this detection tool has arrived too late. Artificial intelligence is being exploited by bad actors in a variety of ways, and tools like this can sometimes be useful, but it depends on how artificial intelligence is developed and how quickly tools like this are developed and deployed. Determined by
OpenAI is developing a similar tool for Sora, a video generator.
[Image – Jonathan Kemper on Unsplash]
