Google Earth’s AI deepfake tool only lasted one day

Google moved fast on this one, but not in the way it wanted. Thursday it rolled out a Google Earth feature powered by Nano Banana 2 that let anyone type a prompt and reshape real satellite imagery. By Friday, the feature was gone.

The problem showed up almost immediately. Henk van Ess, who writes the Digital Digging newsletter, tested the tool by generating fabricated scenes near real, sensitive locations — refugees appearing near the Mexican border, a bomb crater placed next to a hospital in Gaza. These weren't abstract art projects. They were AI-altered versions of real coordinates tied to ongoing conflicts, the kind of images that could easily be mistaken for actual satellite intelligence if stripped of context.

Google's first line of defense didn't hold up. When van Ess flagged the outputs, the company pointed to a digital watermark embedded in generated images and said it blocked prompts on "harmful topics." That answer aged badly within 24 hours, since the examples he produced clearly slipped past whatever filters were supposed to catch them. A watermark doesn't stop a screenshot from spreading without it, and it does nothing to prevent someone from believing what they're looking at in the first place.

By Friday, Google walked the feature back entirely, effectively conceding that dropping a generative editing tool onto real-world satellite maps was riskier than it first appeared. The company hasn't laid out a timeline for bringing it back, or what guardrails might change if it does.

What's notable here isn't that an AI tool got misused — that happens constantly. It's how quickly a major company folded once someone showed, with minimal effort, that its safety claims didn't match reality. Mapping tools carry a specific kind of trust; people assume satellite imagery reflects something that actually exists. Nano Banana 2 blurred that line in a single afternoon.