Google pulls its Earth AI image tool one day after launch over fake satellite imagery

Google on Friday pulled a new AI image generation feature from Google Earth, one day after launching it, because users were creating fabricated satellite imagery that appeared to violate company policies. The tool, powered by the company’s Nano Banana 2 model, let anyone zoom into a real location, type a text prompt, and generate a photorealistic aerial scene layered onto genuine coordinates.

Dutch intelligence researcher Henk van Ess was among the first to demonstrate the problem. In a blog post titled “How to plant a nuclear plant in Iran,” he showed that single-sentence prompts could insert a nuclear facility into Iranian territory, place refugees on a road near the Mexican border, and add a hospital with a bomb crater in Gaza, each in seconds, all rendered on top of authentic satellite imagery.

He was not alone. BBC Verify produced a toppled Eiffel Tower, a sinkhole swallowing the Great Pyramid of Giza, and Russian tanks in Kyiv. AFP generated explosions in Paris and an Islamic State training camp in Syria.

The news agency 404 Media added skyscrapers to rural land and inserted a homeless encampment into Los Angeles. In each case, the fabricated imagery was layered directly onto authentic coordinates, indistinguishable at a glance from the real satellite view underneath.

The concern is not that AI can generate fake images, which it has been able to do for years, but that Google was welding fabricated scenes directly onto real coordinates inside a platform that journalists, investigators, and intelligence analysts treat as a primary verification tool. Van Ess noted that it takes a government request to remove a single real building from Google Earth, but only a sentence to add a fake one.

Google said in a post on its media relations account on X that it had seen geospatial professionals using the feature for useful purposes but was rolling it back to implement stronger guardrails. The company added that none of the AI-generated images appeared in the main Google Earth platform, meaning the fabricated scenes were confined to the generation tool rather than mixed into the base imagery layer.

The generated images did carry Google’s SynthID digital watermark, an invisible signature designed to survive screenshots and compression. But watermarking only helps after the fact, once someone knows to check. It does nothing to prevent a fabricated satellite image from circulating on social media as if it were real, particularly in fast-moving conflict situations where verification lags behind virality.

The episode is the latest in a series of rapid launches and retreats for Google’s generative AI products. The company paused Gemini’s image generation of people in 2024 after it produced historically inaccurate depictions, and has since rebuilt those safeguards. The Google Earth rollback follows the same pattern: a feature ships broadly, edge cases surface within hours, and the company pulls it back to add the guardrails that were missing at launch.