Google has pulled an AI image-generation feature from Google Earth less than a day after launching it, following demonstrations that showed how easily the tool could create convincing fake satellite imagery of real locations.
Among the examples was an AI-generated nuclear plant in Iran created by researcher Henk van Ess, along with fabricated scenes involving refugees, bomb damage, and a fatal crash.
The feature, powered by Google’s Nano Banana 2 image model, allowed users to select a location in Google Earth and use a text prompt to alter its satellite, aerial, or 3D imagery. Google had promoted the feature for uses such as visualising historical sites and redesigning locations.
A Nuclear Plant Appeared In Iran
The misuse became clear almost immediately after the feature became available.
Van Ess, an AI researcher and open-source intelligence investigator, tested the tool by entering prompts for scenarios that could easily be mistaken for real-world events. He generated an image showing a nuclear plant in Iran and also created images showing refugees near the U.S.-Mexico border and a fatal crash on a street in Amsterdam.
He later said that his prompts were not refused or softened during testing. The examples raised particular concern because Google Earth is widely used by journalists, researchers, and open-source intelligence investigators to examine real locations and verify events.
The problem was not limited to one researcher. Other tests produced fabricated bomb craters and disaster scenes, showing that the feature could turn genuine geographic imagery into something that looked like evidence of an event that never happened.
Google Said The Images Were Watermarked
Google initially defended the feature’s safeguards, saying every image created with Nano Banana in Google Earth included its SynthID digital watermark. The company also said users could use Gemini or Google Lens to check whether an image had been generated by AI.
SynthID is designed to embed an invisible digital watermark into AI-generated content so that it can later be identified. Google says the watermark is designed to remain detectable after common modifications such as changes to colour, brightness and compression.
But the presence of a watermark did not eliminate the verification problem.
Van Ess demonstrated that a video generated through Google Earth could fool Hive’s AI detection tool. The Verge also reported that Google’s watermark could be difficult to remove but that generated content could still evade some detection systems when shared outside Google’s own tools.
The reported failure was with downstream detection, not proof that SynthID itself was absent from the generated imagery.
Google Rolled The Feature Back
Google responded within roughly a day, saying it had seen people sharing generated screenshots that appeared to violate its policies. The company said it was rolling back the feature while it worked on stronger safeguards.
“We know that people uniquely trust Google Earth for a reliable view of the world,” Google posted on X. “We’ve seen geospatial professionals using this feature for a range of useful purposes, however we’ve also seen people sharing screenshots of generated imagery that appear to violate our policies. So we’re rolling back this feature in Google Earth while we work on implementing stronger guardrails.”
The rapid rollback highlights a difficult problem for AI image tools. Adding generation directly to a platform that people already use as a source of geographic evidence creates a different verification challenge from a standalone image generator.
For Google, the next step is to determine what stronger safeguards are needed before bringing the feature back.
