I’m starting to understand where Google’s visual AI model comes from, because after a few days of playing with it, I’d sum it up like this: bananas. The images it generates are so realistic they’re bananas. I feel like I’m going crazy after looking at them for too long. And if I had to point out one reason why Nano Banana Pro’s images look so much more realistic than the earlier AI bug, it’s this: they look like photos taken with your phone’s camera.
Sure, the clues are there if you look for them. Take the photo of the (not real!) couple on a city sidewalk at the top of this article. The streetlight in the background doesn’t look quite right to me, and some of the building facades – especially further in the background – look a bit strange and blocky. But if I was just scrolling through this photo on social media? There’s no way I could register this as AI. The subjects look realistic, but I think the fact that the photo doesn’t look too perfect is what sells it.
Dazzling, flat exposure, immense depth of field, slightly crunchy details: all this reminds me of a phone camera. Ben Sandofsky, co-founder of the popular Halide iPhone camera app, agrees. In the AI-generated shuttle photo above, he noted “the aggressive image sharpening seen in smartphone photos. It’s a visual trick that helps create the pop effect.” Another characteristic of photos taken with a phone? Noise. “Most AI-generated photos appear too clean. The texture of these photos appears to come from a small smartphone sensor.”
So where does Google’s AI get ideas about phone photos? Google Photos seems like an obvious – and very problematic – place to go, but Elijah Lawal, global communications manager for Gemini, says that “with Nano Banana, we don’t use Google Photos.” It also tells me that the Nano Banana Pro wasn’t specifically aimed at creating the look of a phone camera. “One of the huge improvements is the ability to connect to Google Search,” he says. If you ask it to create an infographic about today’s weather, it can check the temperature – previously you had to provide more of this information in the prompt.
According to Lawal, this is constrained to text search, not image search. But I can go get it Real-world information itself can be a key ingredient here. Nano Banana Pro is particularly good at adding elements to images that make sense in a given context – even if you never specifically asked for them. Can add historical items, such as period-appropriate clothing and cars, without explicit instruction. Even added a watermark for Northwest Multiple Listing Service when I asked her to create a counterfeit Zillow listing for a counterfeit house in Seattle. Becomes better at understanding a task and adding tiny details without prompting.
Photo: Nano Banana Pro
I asked Gemini for a Zillow ad for a Craftsman-style home with white paint and black trim in West Seattle. A long-winded text list appeared describing the place, but after another prompt, I used Nano Banana Pro to create an image that matched the description. I didn’t specifically ask for this, but the photo has a copyright date of 2023, which is extremely witty, and a watermark similar to the one found on basically every real estate photo you can find in the greater Seattle area. Interestingly, this is not the current logo, but a previous version, which is the same in every photo of the house I bought in 2018.
I asked Google where Nano Banana could come up with such a thing, and DeepMind product manager Naina Raisinghani suggested it was a hallucination, making the following statement: “Nano Banana Pro provides significant improvements in character consistency, image generation, and search-based accuracy. While this is our most accurate image model to date, AI hallucinations may occur. If the image is not quite correct, we encourage you to try again, as another attempt will often produce a result more consistent with yours intentions. “The thing is, adding a watermark to the property listing service seems to work exactly as intended.
Watermark or not, I think the little print on the “for sale” sign might indicate artificial intelligence, or maybe the potted plants on the porch look a little too perfect, but honestly? It’s challenging for me to believe that this house isn’t real, even though I know deep in my heart that it isn’t. I wouldn’t give it a second thought if I came across it on a real estate website, and the watermark would certainly assist sell it as true. If AI is getting this good at imitating things that signal a photo is real, then guys: we’re cooked.
This is what worries me most: AI signals are getting harder to detect, and Nano Banana is getting better at mimicking the small details that make an image seem real. We gave him some vague instructions on how to present a Edge a reporter covering a live event; added details such as microphone with Edge logo in the reporter’s hand and a chyron at the bottom of the screen. No spelling mistakes or letters that look foreign. No hands with six fingers. Nothing to clearly indicate it’s AI, and plenty of little details to sell it as the real deal.
A year ago, or even a few months ago, I had a feeling that there would come a day in the future, a day when it would be unwise to believe any photo or video I saw on the Internet from an unknown source unless proven otherwise. This exercise convinced me that this day was not in the future; is here now. Tune your AI radar accordingly and don’t be surprised if it drives you crazy.