Google pulls Google Earth AI tool that faked satellite scenes

Google spent about a day shipping a Google Earth feature that let anyone type a sentence and drop a fabricated scene onto real satellite imagery, then pulled it. The tool used Google's Nano Banana 2 image-generation technology and was rolled out on Thursday, and 404 Media reported that testers quickly added skyscrapers to a rural area, a homeless encampment to a politically contested part of Los Angeles, a bomb blast and crater to another part of the city, and protesters outside Google's own corporate buildings. AFP's own testing on Friday produced explosions in Paris, nuclear sites in Iran, bomb craters in Russia, and ISIS training camps in Syria. After publication, Google said it was 'rolling back this feature in Google Earth while we work on implementing stronger guardrails.'

The reason this landed harder than a normal generative-AI stumble is what Google Earth is used for. Newsrooms, human-rights researchers, and open-source intelligence analysts have long treated its imagery as the reference layer against which other claims get checked. Henk Van Ess, a digital investigations expert, put it directly: 'Google spent 20 years building the reference the world checks against. Today it added a button that makes things up.' Ross Burley of the Centre for Information Resilience called the launch irresponsible and warned that 'trust in satellite imagery has taken decades to build and could be irrevocably damaged overnight.'

Google's mitigation was SynthID, an invisible watermark it says is embedded in every generated image and detectable through its Gemini chatbot, along with a claim that it prevents 'image creation on harmful topics.' The block clearly did not hold. Brady Africk, a satellite imagery analyst at the American Enterprise Institute, told AFP the feature 'makes it easier to generate convincing fake satellite imagery that can spread quickly online and mislead the public.' A watermark only helps if a viewer bothers to run the check, which is not what happens when a screenshot goes viral during a breaking event.

The honest caveat is that the generated images did not replace the underlying Google Earth tiles for other users, so the tool was less structurally dangerous than it first read. What the reporting does not give you is why no pre-launch review flagged the obvious red-team inputs, or what happens to the screenshots already in circulation.

The forward-looking piece is that provenance infrastructure just got a mainstream test case. Cryptographic image sourcing, SynthID detection wired into browsers and social platforms, and verification services from firms already inside the OSINT ecosystem move from academic proposals to concrete asks. The rollback bought Google time; whether the feature returns with real guardrails, or quietly stays shelved, is the part worth watching.