AI image generators keep getting worse

it is stepbackA weekly newsletter that presents an essential story from the tech world. For more on smartphones and digital imagery – real or otherwise – follow Alison Johnson. Stepback arrives in our subscribers’ inboxes at 8am ET. opt in for stepback Here.

Remember the early days of AI image generation? Oh, how we laughed when our signs resulted in people with too many fingers, rubbery limbs and other details easily pointing out the fakes. But if you haven’t been paying attention, I’m sorry to tell you that the fun is over. AI image generators are getting much better at creating realistic fakes, partly thanks to a surprising new development: tweaking image quality a bit. much worse,

If you can believe it, OpenAI launched its image generation tool DALL-E less than five years ago. In its first iteration, it could only generate 256 x 256 pixel images; Basically little thumbnails. A year later, DALL-E 2 marked a huge leap forward. The images were 1024 x 1024, and looked surprisingly realistic. But there were always arguments.

Casey Newton with DALL-E 2 in his hands shortly after it launched in beta, he included an image created from his prompt: “A Shiba Inu dog dressed as a firefighter.” It’s not bad, and if you saw it at a glance it might fool you. But the shape of the dog’s fur is blurry, the patch on his (adorably small) coat is just some nonsense handwriting, and there’s a weird, thick collar tag hanging on the side of the dog’s neck that doesn’t belong there. It was easier to believe than the cinnamon rolls with eyes from the same article.

Midjourney and Stable Diffusion also came to prominence around this time, adopted by AI artists and those with less tasteful designs. Newer, improved models emerged over the next few years, reducing flaws and adding the ability to render text somewhat more accurately. But most AI-generated images still have a certain look: a little too smooth and perfect, with a kind of glow you’d associate more with a stylized portrait than a candid photo. Some AI images still look the same, but there’s a new trend toward Real Realism that minimizes glare.

OpenAI is a relative newcomer to the tech world when you compare it to companies like Google and Meta, but even as AI continues to grow, these established companies aren’t standing still. In late 2025, Google released a new image model called Nano Banana in its Gemini app. It went viral when people started using it to create their own realistic sculptures. My colleague Robert Hart tried this trend and noticed something interesting: The models preserved their true likeness more faithfully than other AI tools.

That’s the thing about AI images: they often gravitate toward a neutral, soft middle ground. Your request for an image of a table will seem basically correct, but it will also seem like the result of averaging every table ever seen by a computer into something devoid of any real character. The things that make a table image look like the real thing – or reproduce your facial features – actually have imperfections. I don’t mean bizarre artifacts of AI trying to decipher the letters of the alphabet. I mean a little bit of clutter, mess, and lighting that is less than ideal. And lately, that’s also meant copying the flaws of our most popular cameras.

Google updated its image model less than a month ago, calling the Nano Banana Pro its most advanced and realistic model yet. It is able to draw from real-world knowledge and render text better, but what I find most interesting is that it often mimics a photo taken with a phone camera. Contrast (or lack thereof), perspective, aggressive sharpness, exposure options – many of the images produced for me by this model bear the hallmarks of phone camera systems.

Whether you’re aware of it or not, you’re probably used to this look, too. The smaller sensors and lenses in our phones use multiframe processing to overcome their limitations compared to larger cameras, and these photos are optimized for viewing on smaller screens. Overall, this means that phone photos have a certain “look” compared to a more artistic representation of a scene – boosting shadows to reveal more detail and increasing sharpness to make subjects pop. Apparently, Google’s image generator has also absorbed this style.

Google is not alone in providing a more realistic look to the generated images. Adobe’s Firefly image generator has a control labeled “Visual Intensity” that lets you reduce the bright AI look. The results look less pristine and more like they were captured with a real camera – perhaps more of a professional camera than a phone camera, which makes sense given Adobe’s target audience of professionals. But Meta’s AI generator also has a slider for “Stylization”, which dials the realism up or down accordingly. Elsewhere, video generation tools like OpenAI’s Sora 2 and Google’s Veo 3 have been used to create viral clips mimicking the low-resolution, grainy views of security cameras. When AI only needs to be as good as CCTV, this can be quite reassuring.

There are good reasons to be skeptical of claims about the infinite potential for improvement in AI. AI agents are still struggling to buy you a pair of shoes. But the imaging model? they have extremely Improvements have taken place and the proof is before our eyes.

I recently talked to Ben Sandofsky, one of the co-founders of the popular iPhone camera app Halide, about the AI-imitating-smartphone trend. He says that by adopting the strong processing tendencies and familiarity of phone camera photos, which already makes our photos slightly irrelevant to reality, “Google may have taken the edge off the uncanny valley.” AI doesn’t have to make a scene look realistic – in a way, it’s a useless gift. It simply has to mimic the way reality is recorded with all its imperfections, and use it as a kind of cheat code to make the image believable. So how are we to believe every photo we see?

Sam Altman’s view is that real imagery and AI imagery will merge together in the future, and we will be fine with that. I guess he’s partially right, but I have a hard time believing that we wouldn’t really care about what’s real and what’s not. And to solve both of these on our own, we’re going to need some help. And it appears it’s on the way – but it’s not coming as fast as AI image models are improving.

C2PA’s Content Credential standard is gaining some much-needed momentum. On Google’s Pixel 10 series phones, Everyone The image captured by the camera receives a cryptographic signature that reveals how it was created. This avoids the “implicit truth effect,” as Pixel camera chief Isaac Reynolds explained to me earlier this year. If you only label AI-generated images as AI, we assume everything unlabeled is real. In reality, however, the lack of a label only means that we don’t know where the image came from. So Pixel Camera labels both AI and non-AI images equally.

Labels are all well and good, but they’re not useful if you can’t see them. This has begun to change, and earlier this year Google Photos added support for displaying content credentials. The company will also make it easier to see content credentials when they are present in search results and ads. However, that last part is key – right now, most images captured with phone cameras are not credentialed. For the system to work, hardware manufacturers need to adopt standards so that images can be marked as AI or not at the point they are created. The platforms on which images are shared also need to be involved. Until that happens, we’re on our own – and it’s a better time than ever to trust what you see.