The Latest Update to ChatGPT Makes Detecting AI-Generated Images More Difficult Than Ever.
AI-generated images are becoming increasingly difficult to recognize. Of course, there are common cues , but with each new AI-generated image model, they are reduced to a minimum. For example, one might expect AI-generated images of hands to show too many or too few fingers on humans; today, this is not always the case.
One of the weaknesses of AI image processing models is text generation. The image itself may look convincing, but if you look closely at the words, you’ll often notice that they’re not quite right. Some words may be accurate, others fairly close to the original, but in many cases you’ll notice a multitude of inconsistencies: too many repeated letters, letters that aren’t actually letters, symbols that merge and intertwine with each other. Many of these oddities remind me of the language of Star Wars , at least when the AI tries to reproduce American English.
OpenAI’s Images 2.0 system is capable of generating highly realistic AI-generated images.
The latest AI models have significantly improved their ability to generate text. In fact, OpenAI’s latest model for ChatGPT, Images 2.0, is capable of producing highly realistic text , and at such a high volume that I’m not sure many of us (or anyone ) will be able to notice it. According to OpenAI, Images 2.0 is the first image processing model with intelligent capabilities: this means the model can slowly break down each step of a query, which can generate more detailed or accurate images, as well as the ability to generate up to eight images from a single query (though this is only available to paid subscribers). Free users can still take advantage of Images 2.0’s features, such as web searches and cross-checking results. The company claims that “results appear less artificially generated and more thoughtful,” which essentially means that images generated by ChatGPT will be much harder to spot in the future.
The company seems very confident in its latest model. It boasts about the number of different types of images it can create—not just photorealistic pictures, but also screenshots of a computer user interface, a magazine collage, a pile of rice ( that’s a lot to generate), a magazine page, and a handwritten essay. They’re also talking about handwritten images, right down to a coffee stain on paper. You can view these examples in OpenAI’s official blog post to see how strikingly realistic they are . The post also includes other examples, such as highly realistic photographs, graphic novel pages, movie posters, and images with various aspect ratios—even a panoramic view from an iPhone.
All of these developments are impressive and alarming, but for me, it’s the improvements in text processing that take this to a whole new level. Many models are quite good at creating images that trick users into thinking they’re real, but the level of detail in the text and formatting in these examples is unlike anything I’ve seen before. I asked ChatGPT to generate a menu for an Italian restaurant, identifying five entrees and two desserts—I left the specific dishes up to the AI. It managed to generate something realistic, with error-free entries for both entrees and desserts, as far as I could tell. Then I asked it to generate a newspaper article about the Boston Red Sox and New York Yankees moving to another city. It did that too, again without any obvious errors.
I’m not saying these images are perfect : they still have a hint of AI brilliance that a trained eye or careful observer will be able to detect. OpenAI claims that Images 2.0 struggles with some complex tasks, like puzzles, and with details located in hidden or unusually placed areas, such as on upside-down surfaces. But none of that matters when the images this model creates are impressive enough to fool most people who simply skim them. Infographics, photographs, maps, comics, movie posters—you name it: people will use this tool, and you’ll start seeing far more AI-generated images in your life—often without even realizing it.