Google Bolted an Image Generator Onto Real Coordinates. It Lasted One Day.
Bot Mutiny |
The tool let anyone type a sentence and produce a photorealistic aerial scene layered onto genuine map coordinates. Researchers fabricated nuclear sites, bomb craters and refugee columns within hours. Google pulled it the next day.
On Thursday, a Google product manager named Bryan Horowitz posted instructions for a new Google Earth feature. Zoom in to a place on the web version, tap create image, and type whatever you want to see.
On Friday, the feature was gone.
The tool ran on Google's Nano Banana 2 model. It let anyone open a real location, write a sentence, and get back a photorealistic aerial scene layered onto genuine coordinates. Not a picture of a place that looks like somewhere. A picture welded to the actual latitude and longitude, sitting on top of the authentic satellite view underneath.
What people made in the first day
Henk van Ess, a Dutch open-source researcher, was among the first to test the boundaries and the first to publish. He tried refugees at the Mexican border. A nuclear plant in Iran. A crash in Amsterdam. A hospital with a bomb crater in Gaza. His summary of the result was three words: "Nothing was refused."
NPR generated images of Kharg Island, the Iranian island in the Persian Gulf that holds a strategic oil terminal, on fire. 404 Media added skyscrapers to rural land and dropped a homeless encampment into Los Angeles.
In every case the fabrication sat on real coordinates, and at a glance it was not distinguishable from the map underneath it.
Van Ess made the asymmetry concrete. Removing one genuine building from Google Earth takes a government request. Adding a fake one took a sentence.
Why satellite imagery specifically
The reason this landed harder than an ordinary deepfake story is what satellite imagery has been used for.
Open-source investigators use it to verify events in places reporters cannot reach. It is how bomb damage gets confirmed, how mass graves get located, how a government's account of what happened in a valley gets checked against what the ground looks like. Jake Godin, a senior researcher at Bellingcat, described satellite imagery as a safe bet for verification precisely because it has been hard to fake. The camera is hundreds of miles up, the operators are a small number of companies and institutions, and the provenance chain has historically been short.
Coupling a text-to-image generator to that surface does not add a creative feature to a mapping product. It removes the property that made the product evidentiary.
Henry Ajder, speaking to BBC Verify, made the narrower point that fabricated imagery in a conflict zone is most damaging exactly when events are moving fast and reliable information is thin, which is also when people reach for satellite views.
Google's statement
Google said it had seen geospatial professionals using the feature for a range of useful purposes, but had also seen shared screenshots of generated imagery that appeared to violate its policies. The company said it was "rolling back this feature in Google Earth while we work on implementing stronger guardrails."
Read that sequence again. The guardrails are the thing being worked on after the launch, after the rollback, and after outside researchers had already published a list of what the system would produce on request.
Nobody inside the company appears to have run van Ess's four prompts first. They are not obscure. Refugees, a nuclear plant, a plane crash, a bombed hospital. Any journalist handed this tool for ten minutes would have typed something close to that list, and one did, on day one, in public.
The part that outlasts the rollback
The feature is gone. The problem it demonstrated is not.
Two things persist. The first is that a generally available consumer model produced convincing overhead imagery of conflict scenes on demand, and Google withdrawing one interface does not withdraw that capability from the model or from every other model of its class. At least one tester reported that a commercial AI-detection tool could be fooled by the output.
The second is slower and worse. Verification does not fail only when a fake gets believed. It also fails when a real image gets doubted. BBC Verify made this point directly: even where checks work, not everyone runs them, and persistent uncertainty erodes trust in sources that deserve it. A researcher who now has to prove a genuine satellite image is genuine is doing more work than they were doing on Wednesday, and the person who wants that image disbelieved has been handed an argument for free.
What the record establishes, and what it does not
The record establishes that Google launched an AI image generation feature in Google Earth on Thursday, that researchers including van Ess and outlets including NPR and 404 Media generated fabricated imagery on real coordinates within a day, that Google withdrew the feature on Friday citing policy violations, and that the company has said it is working on stronger guardrails.
It does not establish that any fabricated image produced during that window was used to deceive anyone. It does not establish what pre-launch review the feature received, because Google has not said. And it does not establish that the model refuses these prompts now, or that any other image model does, since nobody has published a systematic test.
What it does establish is the order of operations. The tool shipped, strangers found the floor inside a day, and the guardrails were scheduled afterward.
Sources: Google's rollback statement posted to its media relations account on X, July 31, 2026; Bryan Horowitz's launch announcement; NPR, "Google pauses AI satellite images, after fears of deepfakes in the sky," July 31, 2026; 404 Media's testing of the feature; The Verge; BBC Verify, including comments from Henry Ajder; Henk van Ess, quoted by NPR and The Verge; Jake Godin of Bellingcat, quoted by NPR.