This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on smartphones and digital imagery — real or otherwise — follow Allison Johnson. The Stepback arrives in our subscribers’ inboxes at 8AM ET. Opt in for The Stepback here.

Remember the early days of AI image generation? Oh how we laughed when our prompts resulted in people with too many fingers, rubbery limbs, and other details easily pointing to fakes. But if you haven’t been keeping up, I regret to inform you that the joke is over. AI image generators are getting way better at creating realistic fakes, partly thanks to a surprising new development: making image quality a little bit worse.

If you can believe it, OpenAI debuted its image generation tool DALL-E a little less than five years ago. In its first iteration, it could only generate 256 x 256 pixel images; tiny thumbnails, basically. A year later, DALL-E 2 debuted as a huge leap forward. Images were 1024 x 1024, and surprisingly real-looking. But there were always tells.

In Casey Newton’s hands-on with DALL-E 2 just after it launched in beta, he included an image made from his prompt: “A shiba inu dog dressed as a firefighter.” It’s not bad, and it might fool you if you saw it at a glance. But the contours of the dog’s fur are fuzzy, the patch on its (adorable little) coat is just some nonsense scribbles, and there’s a weird, chunky collar tag hanging to the side of the dog’s neck that doesn’t belong there. The cinnamon rolls with eyes from the same article were easier to believe.

Midjourney and Stable Diffusion also came to prominence around this time, embraced by AI artists and people with, uh, less savory designs. New, better models emerged over the next couple of years, minimizing the flaws and adding the ability to render text somewhat more accurately. But most AI generated images still carried a certain look: a little too smooth and perfect, with a kind of glow you’d associate with a stylized portrait more than a candid photo. Some AI images still look that way, but there’s a new trend toward actual realism that tones down the gloss.

OpenAI is a relative newcomer in the tech world when you compare it to the likes of Google and Meta, but those established companies haven’t been standing still as AI ascends. In the latter half of 2025, Google released a new image model in its Gemini app called Nano Banana. It went viral when people started using it to make realistic figurines of themselves. My colleague Robert Hart tried out the trend and noticed something interesting: the model preserved his actual likeness more faithfully than other AI tools.

That’s the thing about AI images: they often tend toward a neutral, bland middle ground. Your request for an image of a table will look basically right, but it will also feel like the result of a computer averaging out every table it’s ever seen into something lacking any actual character. The things that make an image of a table look like the real thing — or a reproduction of your own facial features — are actually imperfections. I don’t mean the bizarre artifacts of AI trying to understand letters of the alphabet. I mean a little clutter, messiness, and lighting that’s less than ideal. And lately, that also means imitating the imperfections of our most popular cameras.

Google updated its image model less than a month ago, touting Nano Banana Pro as its most advanced and realistic model yet. It’s able to draw from real-world knowledge and render text better, but the thing I find most interesting is that it often mimics the look of a photo taken with a phone camera. Contrast (or lack thereof), perspective, aggressive sharpening, exposure choices — so many of the images this model generated for me bear the hallmarks of phone camera systems.

Whether you’re aware of it or not, you’re probably attuned to this look, too. The small sensors and lenses in our phones use multiframe processing to overcome their limitations compared to a bigger camera, and these photos are optimized for viewing on a smaller screen. Altogether, that means phone photos have a certain “look” compared to a more artistic representation of a scene — boosting shadows to reveal more details and cranking up sharpness to make subjects pop. Apparently, Google’s image generator has absorbed this style, too.

Google isn’t alone in offering a more realistic look to generated images. Adobe’s Firefly image generator has a control labeled “Visual Intensity” that lets you tone down the glowy AI look. The results look less pristine and more like they were captured with a real camera — maybe more of a professional camera than a phone camera, which makes sense given Adobe’s target audience of professionals. But even Meta’s AI generator 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 visuals of security cameras. When the AI only has to be as good as a CCTV, it can be pretty convincing.

There are a lot of good reasons to treat claims of AI’s infinite potential for improvement with skepticism. AI agents still struggle at buying you a pair of shoes. But the imaging models? They have vastly improved, and the evidence is in front of our eyes.

I recently spoke to Ben Sandofsky, one of the cofounders of the popular iPhone camera app Halide, about the AI-imitating-smartphones trend recently. He says that by embracing the strong processing tendencies and familiarity of phone camera photos, which already make our photos look a little untethered from reality, “Google might have sidestepped around the uncanny valley.” AI doesn’t have to make a scene look realistic — in a way, that’s a dead giveaway. It just has to mimic the way we record reality, with all its flaws, and use it as a kind of cheat code to make an image look believable. So how do we believe any photo that we see?

There’s the Sam Altman view, that real imagery and AI imagery will blend together in the future, and we’ll just be fine with that. I think he’s partially right, but I have a hard time believing that we won’t really care what’s real and what’s not. And in order to sort the two out for ourselves, we’re going to need some help. And it appears to be on the way — but it’s not coming as fast as the AI image models are improving.

The C2PA’s Content Credentials standard is gaining some much-needed momentum. On Google’s Pixel 10 series phones, every image taken with the camera gets a cryptographic signature identifying how it was made. This avoids the “implied truth effect,” as Pixel camera head Isaac Reynolds explained to me earlier this year. If you only label AI-generated images as AI, then we assume that everything without a label is real. Actually, though, the lack of a label only means that we don’t know where the image came from. So the Pixel camera labels both AI and non-AI images alike.

Labels are all well and good, but they’re not useful if you can’t see them. That’s starting to change, and earlier this year Google Photos added support for displaying Content Credentials. The company will also make Content Credentials easy to view in search results and ads when they’re present. That last part is the key, though — right now, most images captured with phone cameras today aren’t assigned credentials. For the system to work, hardware makers need to adopt the standard so that images are marked as AI or not at the point when they’re created. The platforms where images are shared need to get on board, too. Until that happens, we’re on our own — and it’s a better time than ever to trust nothing that you see.