Google Earth’s AI deepfake tool only lasted one day
Google has shut down Google Earth feature it launched Thursday that allowed users to edit satellite images with text prompts using AI. The tool essentially let users create AI deepfakes of the real world using text prompts; Digital Digging's Henk van Ess, for example, intentionally generated images adding things like refugees near the Mexican border and a bomb crater by a hospital in Gaza.
Google launched a feature on Thursday and killed it by Friday. That is the entire story, and it is the only part that matters.
The feature in question lived inside Google Earth. It let users type a prompt and have an AI model — Nano Banana 2 — rewrite satellite imagery of the real world. Refugees near the Mexican border. A bomb crater beside a hospital in Gaza. The kinds of edits that, if mistaken for genuine satellite footage, could move markets, inflame conflicts, or feed into the already saturated pipeline of war propaganda circulating on social media.
Google's first defense was procedural. The company pointed to a digital watermark embedded in generated images and claimed the system blocked "image creation on harmful topics." Neither safeguard addressed the actual problem. A watermark is invisible to anyone scrolling a feed at speed, and a blocklist of "harmful topics" is a confession that the model was trained to recognize the very categories of imagery it was about to produce. The journalist who demonstrated the failure, Henk van Ess of Digital Digging, did not need to outsmart the system. He simply asked.
The second statement, issued a day later, acknowledged the gap and pulled the feature. No timeline for a revised version. No commitment to a new architecture. Just a quiet withdrawal.
The underlying issue is structural, not editorial. Generative image models are built to follow instructions, and satellite imagery is, by definition, a record of physical reality. Combining the two creates a product whose failure mode is indistinguishable from its success mode. A user who asks for a parking lot and receives a refugee camp cannot, from the output alone, tell which one they are looking at. The watermark exists; the trust does not.
For Google, the cost of the experiment was reputational rather than financial. Earth is a free product, and Nano Banana 2 is a research preview. The real exposure sits elsewhere — in the credibility of any Google surface that touches geospatial data, from Maps to Cloud imagery pipelines used by enterprise customers. A one-day feature does not damage that on its own. It does, however, confirm that the competitive pressure to ship generative tools into every product surface is now outrunning the company's ability to scope what those tools should not do.
The feature is gone. The lesson is not.