ChatGPT's Latest Update Makes It Harder Than Ever to Spot AI-Generated Images
AI-generated images are getting harder and harder to spot. There are the usual tells, of course, but those tells are minimizing with each new AI image model. You could count on AI-generated hands including too many or too few fingers on people, for example; nowadays, that's not necessarily the case.
One particular area of weakness for AI image models has been text generation. The image itself might look convincing, but take a close look at the words, and you'll often notice they're not really right. Maybe some are accurate, while others are pretty close, but, in many cases, you'll notice a lot of inconsistencies: Too many repeating letters, letters that aren't really letters, characters that blend and squiggle in and out of one another. A lot of these quirks remind me of how Star Wars' language looks, at least when AI is trying to replicate American English.
OpenAI's Images 2.0 is capable of generating highly realistic AI images
But the latest AI models are getting much better with text generation. In fact, OpenAI's latest model for ChatGPT, Images 2.0, can render highly realistic text, and a lot of it—to the point where I'm not sure many of us (or any of us) will be able to spot it. According to OpenAI, Images 2.0 is the company's first image model with thinking capabilities: That means the model can take its time breaking down each step of a request, which may generate more detailed or accurate images, as well as the ability to generate up to eight images from one prompt (though this is only available for paid subscribers). Free users can still take advantage of Images 2.0 perks, like how it searches the web for information and double-checks its work. The company says that "results feel less AI-generated and more intentionally designed," which essentially means images generated using ChatGPT are going to be much more difficult to spot going forward.
The company seems very confident in this latest model. It touts the number of different types of images it can create—not just photorealistic pictures, but screenshots of a computer's UI, a magazine collage, a mound of rice (that's a lot to generate), a magazine page, and a handwritten essay. They mean handwritten, too, down to a coffee stain on the paper. You can scroll through these examples on OpenAI's official announcement post to see how shockingly realistic they are. The post includes other examples, like highly realistic photographs, graphic novel pages, movie posters, and images with different aspect ratios—down to the iPhone's panorama view.
All of these developments are as impressive as they are distressing, but for me, it's the improvements to text that really take this to another level. Many models are getting quite good at generating images that trick users into thinking they're real, but the level of detail in text and writing on these examples is something I haven't yet seen. I asked ChatGPT to generate me a menu for an Italian restaurant, highlighting five dinner courses and two desserts—I left the specific dishes up to the AI's discretion. It managed to generate something realistic, with dinner and dessert entries without mistakes—as far as I could tell. I then asked it to generate a newspaper entry announcing that the Red Sox and Yankees would be switching cities. It did that too, again without any obvious mistakes.
I'm not saying these images are perfect: They still have an AI "sheen" to them, that a trained eye or close observer would be able to notice. OpenAI says that Images 2.0 struggles with certain complex task, like puzzles, as well as details found on hidden or oddly placed areas, like reversed surfaces. But none of that really matters when the images this model produces are impressive enough to fool most people who scroll past them. Infographics, photographs, maps, comics, movie posters, you name it: People are going to be using this tool, and you're going to start seeing a lot more AI images in your life—often without ever knowing it.