Google Put An AI Image Generator Inside The World's Most Trusted Map. It Lasted One Day.

BBC Verify's demonstration took a few minutes to build and did more damage to Google's credibility than a month of bug reports. Researchers opened Google Earth's new AI tool, pointed it at the real coordinates of the Eiffel Tower, and generated a photorealistic image of the landmark lying collapsed in rubble, sitting directly on top of Google's own satellite base map. Nothing about the interface said this was fiction. The tool had been live for less than a day.

The feature was called Create image, built on Google's Nano Banana 2 image-generation technology and folded into Google Earth's web version on Thursday, July 30. Google's pitch, according to the company's own description, was that users could "visualize history, create real estate plans and more," typing a prompt and watching an AI-rendered scene snap onto real coordinates over real satellite photography. By Friday, July 31, Google had rolled it back. The company said in a statement that "we've seen geospatial professionals using this feature for a range of useful purposes, however we've also seen people sharing screenshots of generated imagery that appear to violate our policies. We're rolling back this feature in Google Earth while we work on implementing stronger guardrails."

What Testers Built In Under 24 Hours

The Eiffel Tower image was not the outlier. Wire service AFP ran its own tests during the tool's brief public window and generated a fabricated explosion in Paris, a fake nuclear site in Iran, a fabricated bomb crater in Russia, an invented Islamic State training ground in Syria, staged flood scenes across Bangladesh, a fictional migrant caravan at the U.S.-Mexico border, and a suspicious warehouse in Georgia that does not exist. The Washington Post separately documented users generating fake bombings, riot scenes, and fire and smoke damage layered onto real neighborhoods. None of this required technical skill. It required a text box and a location search, both of which Google Earth already had.

The Watermark Was Never The Problem

Google was not blind to the misuse risk. Images generated through Nano Banana 2 carry SynthID, an invisible digital watermark meant to let platforms and researchers identify AI-generated content after the fact. That safeguard describes exactly why the rollback happened anyway: SynthID can tell you an image was synthetic once someone thinks to check for it, but it does nothing to stop a screenshot from circulating first. A fabricated satellite image of a collapsed landmark does its damage in the seconds before anyone runs a provenance check, not after. By the time a fact-checker confirms the watermark, the screenshot has already been captured, stripped of metadata, and reposted somewhere the watermark cannot follow it.

Ross Burley, executive director of the Centre for Information Resilience, put the stakes in terms that go beyond one product recall. "Trust in satellite imagery has taken decades to build and could be irrevocably damaged overnight," Burley said. Digital investigations researchers including Henk Van Ess and Brady Africk of the American Enterprise Institute flagged the same concern once the tool went live: Google Earth has functioned for two decades as one of the few visual references that journalists, human rights investigators, and open-source analysts at groups like GeoConfirmed treat as a neutral starting point in disputes over war zones, disasters, and border incidents. That reputation is the product. Google spent it to ship a feature that lasted about a day.

This Is Not A New Warning, Just A Faster One

The industry had a preview of this exact failure mode in 2023, when a fabricated image of an explosion near the Pentagon briefly moved financial markets before it was debunked. That image took real effort to fake convincingly and still spread fast enough to matter. Nano Banana 2 collapses that effort to a text prompt and a search bar, run through the one company whose base maps are treated as closest to ground truth. Google did not build a tool that occasionally hallucinates a wrong fact, the failure mode this site has spent two years documenting in chatbots. It built a tool that manufactures convincing false evidence on demand, attached to the platform people use specifically because it is supposed to show them what is actually there.

Google's own explanation, that "geospatial professionals" were using the feature productively while others violated policy, treats the misuse as an edge case rather than the predictable outcome of putting a generative image tool on top of a trust-based reference product. A real estate agent mocking up a renovation and a bad actor faking a border crisis were never going to be filtered apart by a policy statement. They were separated, if at all, by whoever happened to try the feature first. In this case, that turned out to be researchers actively looking for the worst use of the tool, which is the best-case version of this story. The rollback took a day. Rebuilding confidence that a Google Earth screenshot means what it has always meant will take longer than that, and Google has not said when, or whether, Create image comes back.