Google shipped a tool for generating satellite imagery on Thursday and withdrew it on Friday [1]. One day is fast for a rollback. It is also roughly how long it took reporters to demonstrate that the thing would render an active war zone on fire on request.
The feature lived inside Google Earth. Product manager Bryan Horowitz described the workflow as zooming in to a place, tapping "create image," and typing "whatever you want to see" [1].
NPR generated images of Iran's Kharg Island on fire and of a flooded U.S. Capitol complex [1]. Henk van Ess, an open-source researcher, catalogued his own attempts: "I tried refugees at the Mexican border, a nuclear plant in Iran, a crash in Amsterdam, a hospital with a bomb crater in Gaza. Nothing was refused" [1].
Google says the tool prevents image creation on harmful topics [1]. Van Ess says nothing he asked for was refused [1]. Both statements appear in the same article, and they do not fit together.
This newsroom has a specific stake in that gap, and it is worth naming rather than gesturing at. Overhead imagery is what a small outlet uses when it cannot put a reporter on the ground: a place we cannot reach, photographed by something with no opinion about what it sees. Jake Godin, a senior researcher at Bellingcat, stated the old assumption plainly: "Satellite imagery has been kind of a safe bet when it comes to verifying an event because it's hard to fake" [1].
Godin's second observation is the structural one. Faking overhead imagery has always been possible for someone with time, skill and a compositing workflow. "It's just streamlining the process, which I think is going to make it proliferate more," he said [1]. The change is not capability. The change is unit cost, and unit cost decides how much of anything gets produced.
Evan Hill, a visual forensics investigator at the Washington Post, aimed at the process instead of the output: "Very curious how or if this idea was red teamed internally because the opportunities for abuse and disinfo are literally boundless" [1].
The scenario is not hypothetical, and NPR reports a case that already ran. Images that surfaced in Iranian media purporting to show damage at a U.S. base "appeared to use Google Earth images and Google AI to generate fake damage at the base" [1]. Kharg Island, the site NPR picked for its own test, sits in the theatre this desk has covered continuously through the Gulf shipping crisis. A convincing fake of a burning Iranian facility does not have to persuade an analyst. It has to move for six hours.
Google's handling of the withdrawal deserves the credit it earned, and this is the part that belongs in any fair account. The company caught the problem in roughly twenty-four hours, which is faster than most platforms have moved on most things. Every generated image carried a SynthID watermark, which NPR reports ensures the image is flagged as AI-generated in other Google tools such as Gemini [1]. In a statement posted on X, the company said: "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." The company added that it is "rolling back this feature in Google Earth while we work on implementing stronger guardrails" [1].
The watermark is where that good faith meets the physics of distribution. Google's own stated trigger was people sharing screenshots. A screenshot is the standard unit of circulation for this category of image, and it travels through platforms that never query a Google tool. NPR does not report whether SynthID survives a screenshot, or whether anyone outside Google can read it. On the record as it stands, the watermark is a guarantee inside Google's products rather than a guarantee about an image once it has left them.
The verification substrate is a shared asset that nobody owns and nobody funds. It runs on the assumption that certain categories of evidence are expensive to fake, and that assumption has been quietly doing the work in every story about a place a reporter could not get to. Pulling a feature restores it for a week. The model that made the images still exists, the demonstration that it works is now public, and the next build of this capability will not necessarily come from a company that answers a reporter's email in a day.
