In This Article
Follow BestForAndroid
Google pulled an AI feature from Google Earth on July 31, one day after launching it, once users showed it could generate a fake nuclear reactor in Iran, a fake plane hitting a Manhattan skyscraper, and fake refugee crowds at the US-Mexico border, all sitting directly on real satellite imagery.
The feature let anyone zoom into a real place in Google Earth on the web, tap “create image,” and type a prompt using Google’s Nano Banana 2 model to generate an AI image grounded in that actual location. Google framed it as a way to reimagine your street with a dog park or picture the Colosseum in your neighborhood. It’s the same unsettled feeling as opening a group photo and spotting one person who was never actually there, except here the fake wasn’t a person, it was a war.
Within a day, online research expert Henk van Ess tried it and described putting refugees near the Mexican border, a nuclear plant in Iran, and a fatal crash on a street in Amsterdam, all on Google’s own satellite imagery underneath. He said the tool never once pushed back or asked him to try a different prompt.
What the BBC found when it tried the same thing
The BBC ran its own tests and got Russian tanks moving across Ukrainian territory, the Eiffel Tower collapsed on the ground, and the Egyptian pyramids swallowed by a sinkhole.
Researchers also found they could get Gemini to describe these fabricated images as legitimate when asked directly.
Why the guardrails missed this
Most AI image tools are built to catch the obvious problems: nudity, weapons, racist caricatures. A nuclear reactor rendered onto a real map of Iran doesn’t trip any of those filters, but it can still look like satellite evidence of something that never happened. Google’s AI-generated images do carry an invisible SynthID watermark and a visible one, but the BBC found third-party AI checking tools still failed to flag several of the fabricated images.
This isn’t Google’s first brush with Gemini being misused. The company has disclosed that Gemini was abused to help create deepfake terrorism content and child abuse material, and it recently sued a China-linked network over Gemini-powered phishing campaigns.
Google’s response
Google confirmed the rollback in a tweet on X, saying it had seen geospatial professionals using the feature productively, but also saw people sharing generated images that appeared to violate its policies, and that it’s rolling the feature back while it builds stronger guardrails.
AI researcher Henry Ajder told the BBC that a journalist might catch a fake image and report on it, but most people never will. Fact-checkers at the International Fact-Checking Network told a similar story: there’s a wide digital literacy gap, and most people have no real reason to question a convincing image unless something forces them to.
Google has walked this line before with real-world maps. It already shows different territorial borders in Google Maps depending on where a user is located, in disputes like India-Pakistan and India-China. Dropping AI-generated nuclear reactors and plane crashes onto real satellite coordinates during active geopolitical tension is a very different kind of risk, and Google clearly decided it wasn’t worth taking.
