Google Earth AI gained an image generator on Thursday and lost it on Friday. The company rolled out Nano Banana 2, its AI image model, inside Google Earth, letting anyone type a prompt and paste fabricated scenery over real satellite maps. Twenty-four hours of public pushback later, the feature was gone.
I get why Google wanted it. Being able to reimagine a city block or drop a fictional landmark into a map is a fun demo, and the announcement post framed the Google Earth AI feature as a way to get creative with geography. Geospatial pros apparently found genuinely useful things to do with it.
The problem is what else it enabled. Google Earth is one of the last places on the internet where a screenshot still reads as evidence. Journalists, researchers, and aid workers treat satellite imagery as ground truth. Bolting a deepfake generator onto that and calling it a feature was asking for trouble, and trouble showed up before the weekend.
The Google Earth AI backlash was instant
A BBC journalist posted that there was no way a tool like this, aimed at one of the most trusted sources of visual evidence around, could possibly be abused to spread misinformation. The sarcasm did not take long to age into prophecy.
Henk van Ess of Digital Digging put the tool through real testing. He generated refugees near the Mexican border, a bomb crater next to a hospital in Gaza, and a nuclear plant in Iran. Google’s initial response noted that images carried a digital watermark and that the company blocked harmful topics. Van Ess said that in his testing nothing was refused, nothing was softened, and nothing suggested he try a different prompt.
He also showed why the watermark defense was weak. He fooled Hive’s AI detection tool with a video generated in the feature, which means generated media could spread even with the label attached. Watermarks only help when people can’t strip them and detectors can’t be tricked.
What Google did next
Google’s first instinct was to defend the feature. Its second, a day later, was to kill it. The company said it had seen geospatial professionals using the feature for useful purposes, but that people were also sharing generated imagery that appeared to violate its policies. It said the feature was being rolled back while the team works on stronger guardrails.
The statement also reminded everyone that generated images never appeared in the main Google Earth experience for other people to see, and that they were watermarked as AI generated. That’s technically true and completely beside the point. The screenshots were the product. People who wanted to fake a place could generate the image, screenshot it, strip or crop the watermark, and post it anywhere.
Why the trust hit matters
Here’s what I keep coming back to. Google Earth AI was a test of whether a company will protect a product’s trust when that trust is its entire value. Google Earth works because people believe the imagery is real. Every fake satellite image shared this week, even the obviously intentional ones from researchers, adds a layer of doubt to a resource that depends on zero doubt.
Google has been shipping AI updates at a brutal pace this year, from Gemini 3.6 Flash to one experimental feature after another. It’s also the same month OpenAI’s model hacked Hugging Face during a security evaluation, so the trust question is suddenly the whole industry’s. Most of these features are harmless. Some of them, like this one, collide with a product where the stakes are different. An image generator inside a mapping tool isn’t the same as an image generator inside a drawing app. The output of the first one looks like evidence.
The honest read is that Google did the right thing on Friday. It looked at the backlash, realized the feature was damaging the thing it was attached to, and pulled it. That’s a better outcome than doubling down. But the rollout itself was the mistake, and it was a predictable one. You don’t need to be a disinformation researcher to know that putting a prompt-to-fake-scene tool on a satellite map ends badly.
Nano Banana 2 still exists in other Google products, and the company says it wants to bring the Earth integration back with stronger guardrails. I’d rather it not come back at all. Google Earth AI was a neat demo bolted onto a trust instrument, and the 24 hours it lasted were enough to prove the point. Some features should stay dead.
