OpenAI creates a tool to detect images generated by its AI; But it is not enough to eliminate deepfakes

The emergence of generative AI has given millions of Internet users access to more or less realistic rendering tools, and in less than a year, generative AI programs such as DALL-E and even Midjourney have made it possible to flood the web. millions of synthetic creations sometimes very, very realistic. The problem is so serious that a scientific study claims that artificial intelligence and deepfakes can even corrupt our memory.

According to the co-creator of ChatGPT, the technology Detection of AI generated images would not be far from infallible, “99% reliable”. This is more than an honorable score, but the content detector created by the company’s AI has one weakness, and most importantly: it can only detect “fake” content if they are designed with DALL-E. Obviously, the GPT-3 based model is very popular, but in reality the value of images created with this multi-modal AI is 6% of 15 billion images created With this technology since its inception

OpenAI is working on an AI image detector that has a major weakness

According to the EveryPixel website, “more than 15 billion images were created using text-to-image conversion algorithms” in one year. In comparison, it would be necessary 150 years for photographers to take so many photosfrom 1826 to 1975 to be exact. Deepfakes and synthetic images still have a bright future ahead, especially as the public has more and more weapon options. A wide variety of services, such as Midjourney, DALL-E, Stable Diffusion, and even Microsoft Paint, offer creating images using artificial intelligence.

Other big names in tech are focused on delivering Labels and markings intended to identify origin contents. Adobe is particularly threatened by the popularity of these productive AIs. The Photoshop editor has therefore launched a universal symbol to identify AI-generated images and wants all market players to adopt it.

Source : Bloomberg