Storia
agosto 4, 2026

Google Earth’s AI experiment buckled under a trust test

Google pulled Nano Banana 2 from Google Earth within a day after researchers showed it could convincingly fabricate events at real locations. The company says useful professional applications remain, but stronger safeguards are needed.

Google’s attempt to make Earth imagery more creative ran straight into a harder question: what happens when a widely trusted reference tool can manufacture convincing scenes of reality? Within a day of launch, the company withdrew the feature.

Nano Banana 2 had been added to the web version of Google Earth to let users generate images from its satellite, aerial and 3D imagery. Google and its supporters framed the capability as a way to visualize historic places, proposed developments and community projects; its launch messaging asked users to imagine seeing a city as it looked a century ago or visualizing a new local basketball court.

But researchers and journalists saw a sharper risk. The feature joined generative AI to imagery that investigators routinely use to establish where photos and videos were taken. Henk van Ess demonstrated that a single prompt could place “refugees near the Mexican border,” a nuclear plant in Iran and a fatal crash in Amsterdam over Google’s underlying imagery. Other tests produced a bomb crater near a Gaza hospital, reinforcing fears that false claims could borrow credibility from a familiar mapping platform.

Google initially pointed to SynthID watermarks and said it blocked harmful image requests. Yet critics argued that screenshots, re-encoded clips and recordings could shed the contextual cues needed for verification. Van Ess said a video made with the tool fooled Hive’s AI detector, while his prompt testing found that “Nothing was refused, nothing was softened, and nothing suggested I try a different prompt.”

The company ultimately acknowledged the central issue: users “uniquely trust Google Earth for a reliable view of the world.” It said geospatial professionals had found useful applications, but screenshots of generated imagery appeared to violate policy, prompting a rollback while it develops “stronger guardrails.” The retreat addresses the immediate exposure, not the broader problem: AI tools already make manipulation easier across the web, while verification now increasingly requires cross-checking independent satellite sources and metadata.

Copertura della storia