Google launched and then retracted an AI image-generation feature inside Google Earth within a single day after researchers demonstrated it could produce convincing fake satellite imagery of geopolitically charged scenes.
Key takeaways
Google's AI Earth feature let users edit real satellite, aerial, and 3D imagery using text prompts — then removed it roughly 24 hours after launch.
Digital Digging researcher Henk van Ess publicly demonstrated the tool generating fake images depicting refugees near the Mexican border and a bomb crater near a hospital in Gaza.
The feature was pulled after the demonstrations triggered immediate backlash around misinformation risks tied to manipulated geospatial imagery.
The episode illustrates a specific failure mode for AI image tools: combining a real-world geographic anchor with generative edits dramatically raises the believability — and the harm potential — of faked imagery.
Google has not announced a timeline for a revised version of the feature.
What the tool actually did
The feature, which The Verge reported was shut down shortly after its Thursday launch, worked by letting users type a text prompt to modify what appeared to be genuine satellite or aerial views inside Google Earth. The result was an image that looked, at a glance, like authentic geospatial data — because it was layered directly over real map imagery.
That distinction matters enormously. A standalone AI-generated landscape is recognizable as synthetic by most informed viewers. The same generated content composited onto a real-world coordinate, rendered in the visual language of satellite photography, is a different kind of artifact. It carries an implicit geographic claim.
Google's AI Earth feature let users superimpose AI-generated imagery over real satellite maps using text prompts before being pulled.
Van Ess's demonstrations made the risk concrete fast. Images purporting to show refugees massing near the U.S.-Mexico border, or damage near a hospital in a conflict zone, are exactly the kind of content that circulates in geopolitical disinformation campaigns. The fact that a credible Google product was the source would have amplified their apparent legitimacy.
The speed of the reversal signals how seriously this landed
Google's decision to retract the feature in under 24 hours is notable. The company initially responded to early criticism before eventually pulling the tool entirely, according to TechCrunch. That sequence — defend, then reverse — suggests the internal calculus shifted quickly once the van Ess examples were circulating publicly.
For AI-art creators, the episode is a useful case study in how context transforms risk. The same diffusion-style image generation that powers portrait generators and style tools becomes a substantially different product when the output is anchored to real-world geography and presented inside a trusted mapping application. The harm isn't in the generation itself — it's in the framing.
This also connects to a broader pattern of platforms discovering that AI image tools require more upfront harm modelling than text tools. Concerns about AI-generated imagery and detection are already shaping how platforms deploy these features — earlier this year, Pangram launched a dedicated AI image detection model aimed at flagging synthetic content, and independent testing of Google's own SynthID watermark found that technical robustness alone may not stop AI-generated disinformation.
Practical implications for geospatial and reference-image workflows
Creators who use satellite imagery as reference — concept artists building real-world environments, game designers sourcing terrain textures, or anyone using map screenshots as compositional guides — should note that this feature is gone for now. Google has not said whether a guardrailed version is planned.
More broadly, the episode is a reminder that AI image tools embedded inside platforms with geographic or documentary authority face a much higher bar than standalone generators. A feature that would be unremarkable in an art tool becomes a liability when the host platform is trusted as a source of factual visual data about the real world.
For creators building workflows around AI-generated environments or landscapes, the practical path forward remains using dedicated image generation tools where the synthetic nature of the output is unambiguous — rather than waiting on a product that Google may not re-release in its original form.
