Google Removes Google Earth AI Feature Over Misinformation Fears

The feature launched globally for Google Earth web users allowed anyone to select a location, press a “create image” button and enter a text prompt describing how they wanted the place to be transformed. On July 31, Google announced that it was rolling back the capability while it developed “stronger guardrails.” rsal was unusually fast even by the standards of experimental AI products. Within hours of its release, researchers had used the tool to fabricate refugee camps, military facilities, disaster damage and scenes of violence on top of Google Earth imagery.

A creative Google Earth experiment quickly became a misinformation concern

Google introduced the feature as a way for people to explore history, test architectural ideas and imagine how familiar locations might change in the future.

The company demonstrated five proposed uses. These included reconstructing Pompeii as it may have appeared in 78 A.D., creating historical infographics about the Statue of Liberty, redesigning an empty Tokyo lot as a commercial district, adding a sustainable cabin to a real landscape and turning Google’s Mountain View campus into a futuristic city.

To use it, a person only had to open Google Earth on the web, zoom in on a location, select “create image” and type a prompt. Nano Banana 2 would then combine the request with Google Earth’s existing satellite, aerial and three-dimensional imagery. escribed the resulting pictures as concepts “grounded in the real world.” That connection to real coordinates and recognisable geography, however, became the feature’s biggest risk.

Unlike a completely fictional AI landscape, the generated images could inherit familiar roads, buildings, lighting conditions and terrain from an actual location. A screenshot taken out of context could therefore resemble satellite evidence of something that had genuinely happened.

The feature’s brief timeline

rchers showed how easily realistic crises could be fabricated

Dutch investigative researcher Henk van Ess was among the first to publicly test the feature’s limits. He reported using short prompts to add refugees near the US-Mexico border, place a nuclear facility in Iran, stage a fatal traffic accident in Amsterdam and create a hospital beside a bomb crater in Gaza.

According to van Ess, the prompts were not rejected, softened or redirected, despite Google initially saying that harmful image creation was restricted. He argued that the danger came from combining synthetic elements with genuine imagery from exact real-world coordinates. ers demonstrated fabricated flooding and tornado damage, while another example recreated imagery resembling a major terrorist attack. One weather-focused user said he was able to spoof flood and tornado damage in approximately 20 seconds. ern was not that these generated scenes replaced the official imagery that every Google Earth user could see. They did not. The problem was that screenshots and screen recordings could be exported, reposted and stripped of the surrounding interface that identified them as generated concepts.

Google initially pointed to SynthID protection

Google’s first response focused on its provenance technology. The company said every image generated through Nano Banana 2 in Google Earth included SynthID, its invisible digital watermark for identifying content made with Google AI.

Users who questioned an image’s authenticity could reportedly check it through the Gemini app or Google Lens. Google also said it prevented generation involving harmful topics and regularly updated its protections. ana 2 already supports SynthID alongside C2PA Content Credentials, an industry standard intended to preserve information about how digital media was created or edited. Google said in February that the SynthID verification feature in Gemini had been used more than 20 million times since its introduction. the researcher’s testing illustrated a weakness common to invisible watermarking systems: online images rarely circulate as untouched original files.

Van Ess uploaded a screen recording of one generated scene to X, where the third-party detection service Hive reportedly returned a probability of just 1% for AI-generated video and 0% for a deepfake. The service also assigned a 36% probability of AI-generated music, even though the clip contained no audible music. These were the results of one researcher’s test rather than a controlled independent audit, but they demonstrated how screenshots, recordings and re-encoded files can complicate automated detection. e acknowledged that trust in Earth made the risk different

One day after defending the watermarking and safety measures, Google changed course.

The company said people “uniquely trust Google Earth” as a reliable view of the world. Although geospatial professionals had found useful applications for the feature, Google said it had also seen screenshots of generated imagery that appeared to violate its policies.

Google therefore rolled back the feature while working on stronger safeguards. It stressed that generated images had never appeared in the main Google Earth experience for other users and were watermarked as AI-generated. id not identify the exact images or policies involved. It also did not provide a date for the feature’s possible return.

This means the capability appears to be paused rather than permanently abandoned, although any future version will likely need tighter restrictions around military locations, disasters, public safety incidents, hospitals, borders and other politically sensitive places.

Nano Banana 2 made the images particularly convincing

Nano Banana 2, also known as Gemini 3.1 Flash Image, was introduced in February as a faster alternative to Google’s higher-end Nano Banana Pro model.

Google said the model combines Gemini Flash’s speed with improved world knowledge, photorealism and instruction following. It supports resolutions ranging from 512 pixels to 4K and can preserve the consistency of as many as five characters and 14 objects during an image workflow. It can also use information and images from web search to generate more accurate representations of specific subjects. pabilities make the model useful for design, advertising, education and architectural visualisation. Inside Google Earth, however, the same strengths allowed fictional events to be integrated into recognisable physical environments with relatively little effort.

The controversy was therefore not simply about another AI image generator. It was about placing a powerful generator inside a product commonly treated as a reference source by journalists, researchers, investigators, emergency managers and ordinary users.

Fake satellite imagery was already spreading before the launch

The risks raised by researchers were not hypothetical.

In March 2026, fact-checkers documented an AI-manipulated image that supposedly showed destroyed US radar equipment at a base in Qatar. Researchers found that it was actually based on an older Google Earth image of a US facility in Bahrain.

Rows of vehicles remained in identical positions in both the genuine and altered versions, helping investigators identify the manipulation. Despite those clues, the fake image accumulated millions of views after spreading across several languages and social platforms. ted satellite images have also appeared during conflicts involving Russia and Ukraine, India and Pakistan, and the Middle East. Such content can influence public opinion, complicate assessments of military damage and potentially affect financial markets when people react before verification is complete. arth has historically helped investigators expose this kind of manipulation. In 2015, Bellingcat used dated Google Earth and DigitalGlobe imagery to challenge satellite photographs presented by Russia’s Ministry of Defence following the downing of Malaysia Airlines Flight MH17. The investigation compared vegetation, fields and other landscape details to show that the claimed dates did not match the available imagery. tory helps explain the reaction to Google’s experiment. A platform used to verify visual evidence had temporarily gained a built-in method for producing fictional evidence from the same underlying geography.

The rollback arrives as AI transparency rules tighten

The timing is also notable because the European Union’s AI Act transparency requirements begin applying on August 2, 2026, just two days after Google launched the feature.

Article 50 includes transparency and machine-readable marking requirements for certain AI-generated or manipulated content. Providers that fail to comply can face fines of up to €15 million or 3% of their worldwide annual turnover, depending on the violation and circumstances. id not connect its decision to the EU rules, and its use of SynthID suggests it had already built provenance measures into the product. Still, the incident shows why regulators increasingly view technical watermarking as only one part of the solution.

Invisible labels may help investigators inspect an original file, but they cannot guarantee that viewers will recognise a screenshot, cropped image or screen recording as synthetic before it spreads.

A warning for future AI mapping tools

Google Earth’s short-lived Nano Banana 2 integration shows the tension between creative visualisation and factual trust.

A tool that helps an architect show how a proposed building could look can also fabricate damage to a hospital. The same technology that reconstructs an ancient city for students can create a military facility that never existed.

Google’s decision prevented the feature from remaining widely available while those risks were still unresolved. The harder challenge will be redesigning it without removing the flexibility that made the tool useful in the first place.

Possible safeguards could include visible labels embedded directly into exported images, restrictions around sensitive locations, stronger prompt filtering, separate visual styling for fictional concepts and preservation of tamper-resistant provenance information when users take screenshots or recordings.

For now, the main Google Earth imagery remains unchanged and publicly generated pictures were never inserted into the shared map. But the one-day experiment demonstrated how quickly the distinction between geographic visualisation and synthetic evidence can disappear once a highly capable image model is placed directly on top of the real world.