Google Quietly Killed Its AI Satellite Image Editor After Everyone Immediately Misused It
Google launched and then rapidly buried a feature that let users generate AI-modified satellite imagery directly inside Google Earth. The turnaround took less than 24 hours. That should tell you something.
The feature, built around Google's Nano Banana 2 image generator, was announced on July 30 by Google Earth product manager Bryan Horowitz. The pitch was straightforward enough: users could take real satellite, aerial, and 3D imagery from Google Earth and use AI to remix it based on text prompts. Google framed this as a way to visualise completed construction projects or, somewhat ambitiously, reconstruct ancient Pompeii in photorealistic detail.
Researchers had a different reaction.
Eliot Higgins, founder of investigative outfit Bellingcat, posted about the feature on Bluesky the same day it launched. His sarcasm was barely restrained. To illustrate the point, he included an AI-modified Google Earth image showing a giant golden statue of Donald Trump towering over the White House. "No way this could be abused," he wrote.
Independent investigator Henk van Ess went further, documenting in a blog post exactly what he managed to generate with a single sentence of input each time: refugees near the Mexican border, a nuclear plant in Iran, a fatal crash on an Amsterdam street. All sitting on top of Google's own authentic satellite imagery underneath.
The specific concern here isn't just that AI image tools exist. They're everywhere. The problem is that Google Earth carries implicit authority. Journalists and open-source researchers regularly use it to verify footage and photographs of real locations. Bolting an AI image editor onto that creates an obvious and immediate trust problem.
Van Ess pointed to an earlier incident where someone used a Google Earth screenshot of the US Navy's Fifth Fleet base in Bahrain and prompted Gemini to depict it as having been struck by Iranian drones. That process previously required multiple steps. With the new built-in tool, it could apparently be done in seconds.
Google's initial defence centred on SynthID, its digital watermarking system for AI-generated images. Every image created through the feature would carry an invisible watermark, Google said, and users could run verification checks through Gemini or Google Lens.
That defence had holes. Ars Technica tested SynthID and found that while the watermark does survive significant editing and compression in many cases, checking is capped at roughly ten images per day. More critically, when Ars photographed an AI-altered image with a smartphone camera, SynthID returned nothing. No verification at all.
Van Ess made the same point bluntly. Fabricated images don't circulate as pristine files with metadata intact. They get screenshotted, re-encoded, filmed off someone's phone, shared twelve times, and posted again. A watermarking system that breaks the moment an image leaves its native format is not a serious misinformation control.
Google also claimed its content policies would block image generation on harmful topics. Van Ess tested that too, producing an image of a hospital with a bomb crater in Gaza without apparent difficulty.
By July 31, Google announced on X that the feature was being rolled back. Its statement acknowledged that "geospatial professionals" had been using it for legitimate purposes but admitted that people were generating and sharing images that violated its own policies. The company said it was working on "stronger guardrails" before any potential relaunch.
The damage, such as it is, extends beyond this specific tool. By introducing an AI image generator into a platform trusted for authentic imagery, Google handed ammunition to exactly the kind of actors who benefit from public confusion about what's real. Any government official or propagandist wanting to dismiss a genuine satellite photograph now has a slightly more credible "it could be AI" defence.
Google pulled the feature fast. But the fact that it shipped at all, apparently without anyone seriously stress-testing the obvious misuse scenarios first, is the more interesting story.