Google Pulls AI Image Tool From Earth After Users Fabricate Satellite Photos in a Day

Google's experiment with generative AI in Google Earth lasted barely 24 hours before the company shut it down. The feature, launched Thursday, July 31, 2026, embedded the Nano Banana 2 image generator directly into Google Earth, letting anyone superimpose AI-fabricated scenes onto real satellite imagery using nothing more than a text prompt.

The rollback came fast because the abuse case was obvious almost immediately. Researcher Henk van Ess, of the open-source investigation outlet Digital Digging, showed how easily the tool could manufacture convincing fake events: refugees crowded at the Mexican border, a bomb crater carved into the ground near a hospital in Gaza. Neither scene happened. Both looked real enough to pass for satellite evidence.

Google had built in some safeguards. Generated images carried an AI watermark and were kept out of the shared Google Earth experience seen by other users, meaning the fakes weren't broadcast publicly by default. Van Ess's testing found that safeguard didn't go far enough anyway.

He demonstrated that watermarked Nano Banana 2 content could still slip past Hive, a widely used AI-detection service, undermining the core justification Google had offered for shipping the feature in the first place.

The watermarking Google pointed to as its safety net didn't stop detection tools from being fooled.

Google's own explanation for the retreat cited a need for "stronger guardrails" before any return, according to statements given to TechCrunch and The Verge. The company framed the decision as a pause rather than a permanent cancellation, though no relaunch timeline has been given.

How long did it actually last?

Reporting on the exact window between launch and removal isn't perfectly aligned. TechCrunch and The Verge both describe the pull as happening "one day after launch," while The Decoder characterizes it as roughly two days. Given that all outlets agree on the Thursday launch date, and Google's own comments to The Verge reference action taken "one day later," the one-day timeline appears to be the more accurate account.

Why this matters beyond one feature

Satellite imagery carries an unusual amount of default credibility: it's used in journalism, conflict verification, disaster response, and legal disputes precisely because it's assumed to be captured, not conjured. A generator that can convincingly counterfeit that specific format, embedded inside the platform millions already trust as a geographic reference tool, raises the stakes well past typical image-generation controversies.

The episode also lands at a moment when AI detection infrastructure is being treated as a primary defense against synthetic misinformation. If tools like Hive can be fooled by watermarked outputs from a major lab's own model, that defense looks considerably shakier than vendors have been suggesting, regardless of how quickly any single feature gets pulled.