Google Earth's AI makeover lasted less than 24 hours

Turning satellite maps into custom AI scenes sounds incredible on paper—until users realize they can conjure bomb craters and border crises with a single sentence.

Google rolled out a new image generation feature inside Google Earth web, powered by Nano Banana 2. By zooming into any coordinate, clicking "create image," and typing a prompt, users could reshape real-world satellite, aerial, and 3D imagery into custom visuals.

Just one day later, on July 31, Google abruptly pulled the plug and rolled the entire feature back.

A 24-hour deepfake playground

Google originally pitched the tool as a breakthrough for education, urban planning, and casual creativity.

In its launch announcement, the company highlighted use cases like rendering hyper-realistic street views of Pompeii in 78 A.D., drafting 3D real estate concepts over empty lots in Tokyo, pulling factual infographics via Gemini, or turning its Mountain View campus into a sci-fi utopia.

Geospatial professionals welcomed the speed of on-canvas ideation. But the open-world sandbox immediately clashed with real-world geopolitics.

Digital Digging researcher Henk van Ess quickly stress-tested the system by generating fake refugees near the U.S.-Mexico border and a convincing bomb crater right next to a hospital in Gaza.

According to van Ess, the guardrails completely failed. He noted that during his testing, nothing was refused, nothing was softened, and the system never suggested alternative prompts. He even fooled Hive's AI detection tool with video generated from the platform.

Google rolls it back

Initially, Google responded by pointing to its invisible digital watermarks and safety filters that supposedly blocked harmful topics.

By Friday, however, the tone shifted. After screenshots of policy-violating imagery circulated online, Google announced it was pausing the feature entirely.

In an update to its blog post, Google explained that users place a unique trust in Google Earth as an accurate view of our planet. While noting that generated imagery was watermarked and never inserted into the public map view for others to browse, Google admitted it had to step back and implement stronger safeguards.

Why it matters

Anchoring generative AI directly to real-world coordinates was undeniably clever, but launching it without airtight safety boundaries was an unforced error.

For architects and teachers, being able to paint ideas directly on real satellite topography is immensely valuable. But satellite imagery carries an intrinsic stamp of factual truth that regular synthetic canvases simply don't have.

We think this incident highlights a growing blind spot among frontier tech giants: applying standard chatbot moderation to hyper-specific spatial canvases doesn't work. Watermarks mean very little once a low-resolution crop hits social media during an active geopolitical crisis.

Until Google can reliably sandbox sensitive locations and enforce zero-tolerance filters on sensitive geopolitical areas, turning Google Earth into a generative canvas remains too dangerous to deploy at scale.