Google Pulls AI Image Generator From Google Earth One Day After Launch

Google launched and partially pulled a new artificial intelligence feature in Google Earth within a single day. The Nano Banana 2 image generator, which lets users create realistic visuals embedded in real satellite imagery, rolled out to all users of the browser version of the service on July 30, 2026. Just one day later, the company announced it was withdrawing the feature, citing violations of its policies.

A Rapid Reversal

The launch of Nano Banana 2 in Google Earth looked like one of the service's most striking updates in years. The Gemini-based AI model didn't generate images detached from reality, as classic generators do, but built them on real satellite data, aerial photos, and 3D models of a specific location. Users simply had to zoom into any point on the map, click "create image," and type a prompt to get a visualization embedded in real spatial context.

The company promoted the feature as a tool for reconstructing historical sites, planning urban greenery, or visualizing architectural projects. Just 24 hours later, Google added an update dated July 31 to the same blog post, announcing it was suspending the feature's availability. The company wrote that it had seen geospatial professionals use the tool in many useful ways, but had also noticed users sharing screenshots of images that violated its policies.

A Researcher's Test

Before Google announced the feature's withdrawal, independent researchers and journalists had already flagged the risks tied to the tool. Henk van Ess showed that creating a fake image took just one sentence and a few seconds. He used it to place a nonexistent nuclear power plant in Iran, refugees at the Mexican border, and a hospital with a fresh bomb crater in Gaza on the map. The results preserved correct lighting, camera angle, and coordinates, making them look like authentic satellite photos.

Van Ess pointed out that removing a real object from satellite maps usually requires government intervention, while adding a fake one, after Google Earth's new feature, required nothing more than typing a prompt. He also recalled that satellite imagery has long served as evidence in investigations, for instance when verifying Russia's fabricated photos from the MH17 crash, where researchers compared them against archival Google Earth footage.

The Bahrain Precedent

The risk isn't purely theoretical. During an earlier period of tension between the US and Iran, fake before-and-after photos circulated online, purportedly showing a destroyed American military base. In reality, it was an AI-generated image overlaid on an authentic satellite photo of a facility in Bahrain. The material was viewed by roughly 950,000 people before fact-checkers managed to debunk it.

That incident was created without any built-in Google Earth generator, using only external AI tools and image editing. Critics of the new feature argued that giving every user the ability to do the same thing inside the mapping service itself, one that enjoys exceptional trust as a source of reliable geographic data, drastically lowers the barrier to this kind of disinformation.

Google's Safeguards

Google defended the feature with two layers of protection: every generated image carried an invisible SynthID digital watermark, and generated graphics did not appear in the main Google Earth view seen by other users, only on the creator's own account. A user unsure of an image's origin could check it through the Gemini app or the Lens feature in Search.

We know people place exceptional trust in Google Earth as an accurate depiction of the world. We've seen geospatial data professionals use this feature in many helpful ways, but we've also noticed users sharing screenshots of generated images that appear to violate our policies. That's why we're pulling this feature from Google Earth while we roll out stronger safeguards - Google, July 31, 2026 blog post update

The company did not specify exactly which images violated its policies or how many reports it received during the feature's one day of operation. Nor did it disclose a specific timeline for restoring the tool, saying only that it was working on "stronger safeguards" before the generator returns to the service.

What's Next

The episode shows how quickly enthusiasm for new generative AI features can collide with real risks of abuse when a tool is connected to a service perceived as an objective source of data. For investigative journalists, open source intelligence analysts, and institutions that use Google Earth as a reference point, the episode signals that even watermarked AI images require additional verification, especially since tools for removing watermarks like SynthID already exist.

Google did not specify whether or when the image generator will return to Google Earth in a modified form, or whether restrictions will target specific categories of subjects, such as military installations or critical infrastructure. Until it relaunches, the feature remains unavailable to users of the browser version of the service.