Google Put Image Generation Into Google Earth and Pulled It a Day Later

Google added generative image editing to Google Earth on Thursday and removed it on Friday. In between, researchers and journalists demonstrated that the feature would produce convincing fake satellite imagery on demand.

What people made

The circulated examples were not subtle failures of realism. They included blast craters beside a hospital in Gaza, columns of people at the US-Mexico border, and nuclear facilities that do not exist. The researcher Henk van Ess demonstrated the failure mode publicly; he did not find a bug, he used the feature as built.

Why the watermark did not help

Google's first line of defence was that every generated image carried SynthID provenance marking and an AI-generated label in the interface. That defence answers a different threat. The images that spread are screenshots, and a screenshot of a labelled image carries neither the label nor the watermark. Provenance embedded in a file does not survive the way this content actually travels.

What was not affected

The generated images never altered the shared Google Earth basemap. Nobody edited the map everyone else sees. The framing that Google Earth itself was corrupted is wrong — the risk was that a fabricated view could be captured and passed off as the real product's output, which carries the credibility of satellite imagery.

A pause, in Google's words

The company says it is rolling the capability back while it works on implementing stronger guardrails before enabling image generation inside Google Earth again. That is an explicit intention to return, not a withdrawal. No timeline has been given, and Google has not said what a sufficient guardrail would look like for a tool whose entire appeal is making the map look like something it is not.

Twenty-four hours is the interesting number

The feature shipped on Thursday and was gone by Friday. Nobody needed weeks to find the failure — the Federation of American Scientists published on it, NPR covered it, and the demonstrations were running within hours of launch. A rollback that fast is not evidence of a subtle flaw slipping through review. It suggests the obvious question was either not asked or was answered with the watermark, which is the answer that did not hold.