I’m starting to understand where Google’s visual AI model gets its name, because after playing around with it for a couple of days, that’s how I’d sum it up: bananas. The images it generates are so realistic it’s bananas. I feel like I’m going bananas after staring at them for too long. And if I had to pinpoint one reason why Nano Banana Pro’s images look so much more realistic than the AI slop that came before them, it’s this: They look like photos taken with a phone camera.
Sure, the tells are there if you look for them. Take the image at the top of this article of the (not real!) couple on the city sidewalk. The streetlight in the background doesn’t look quite right to me, and some of the building facades — especially farther into the background — look a little strange and blocky. But if I was just scrolling past this photo on social media? No way I’d clock it as AI. The subjects look realistic, but I think the fact that the image doesn’t look too perfect is what sells it.
The bright, flat exposure, the generous depth of field, the slightly crunchy details: It all screams phone camera to me. Ben Sandofsky, cofounder of the popular iPhone camera app Halide, agrees. In the AI-generated image of the ferry boat above, he noted the “aggressive image sharpening you encounter on smart phone photos. It’s a visual trick that helps image ‘pop.’” Another hallmark of photos taken with a phone? Noise. “Most AI generated photos feel far too clean. The texture in these photos feel like they came from a tiny smart phone sensor.”
So where is Google’s AI getting its notions about phone photos from? Google Photos would seem like an obvious — and deeply problematic — place to go, but Elijah Lawal, the global communications manager for the Gemini app, says that “for Nano Banana we don’t use Google Photos.” He also tells me that Nano Banana Pro hasn’t been specifically steered toward producing a phone camera look. “One of the huge improvements is that it can connect to Google Search,” he says. If you prompt it to create an infographic about today’s weather, it can go look up the temperature — previously, you would need to include more of that information in your prompt.
According to Lawal, this is limited to text search and not image search. But being able to go get real-world information on its own might be a key ingredient here. Nano Banana Pro is especially good at adding things to images that make sense in that context — even if you never specifically asked for them. It can add historical elements like period-appropriate clothes and cars without being told expressly to do so. It even added a watermark for the Northwest Multiple Listing Service when I asked it to create a fake Zillow listing for a fake house in Seattle. It’s getting a lot better at understanding the assignment and adding those little details without being prompted to.
Image: Nano Banana Pro
I asked Gemini for a Zillow listing for a craftsman-style house with white paint and black trim in West Seattle. It gave back a wordy text-only listing describing the place, but with another prompt, I used Nano Banana Pro to create an image to go with the description. I hadn’t specifically asked for it, but included in the image is a 2023 copyright, which is deeply funny, and a watermark like the one that’s on basically every real estate photo you find in the greater Seattle area. Interestingly, it’s not the current logo — it’s the previous version, which is the same one on every picture of the house I bought in 2018.
I asked Google where Nano Banana could possibly come up with that, and DeepMind product manager Naina Raisinghani suggested it was a hallucination, offering this statement: “Nano Banana Pro provides major upgrades to character consistency, image generation, and search-grounded accuracy. While this is our most precise image model to date, AI hallucinations can occur. If an image isn’t quite right, we encourage you to retry, as a subsequent attempt often yields a result more in line with your intention.” The thing is, adding the watermark for a real estate listing service seems like the model working exactly as intended.
Watermark or no, I guess the small print on the “for sale” sign might give this away as AI, or maybe that the potted plants on the front porch look a little too perfect, but honestly? I’m having a hard time believing this house is not real, even though I know in my bones 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 help sell it as genuine. If AI is getting this good at imitating the things that signal a photo is real, then guys: We are cooked.
That’s what’s most concerning to me: The AI tells are getting harder to spot, and Nano Banana is getting better at mimicking little details that make the image seem real. We gave it some vague prompts to depict a Verge reporter covering a live event; it added details like a microphone with the Verge logo in the reporter’s hand and a chyron in the lower part of the screen. No misspellings or alien-looking letters. No hands with six fingers. Nothing that would obviously tip it off as 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 sense that there was a day coming in the future, a day when it wouldn’t be wise to believe any photo or video I saw online from an unfamiliar source unless proven otherwise. This exercise has convinced me that that day isn’t in the future; it’s here now. Tune your AI radar appropriately, and don’t be surprised if it drives you a little bit bananas.
