Google has temporarily removed a new artificial-intelligence feature from Google Earth after researchers demonstrated how easily it could be used to create convincing but entirely fabricated satellite images. The tool, which allowed users to select a real location and describe what they wanted to appear there, was withdrawn only a day after its introduction as fears grew that it could become a powerful new source of wartime propaganda, financial misinformation and fabricated evidence.
The feature combined Google Earth’s real satellite, aerial and three-dimensional imagery with Google’s Nano Banana image-generation technology. A user could zoom in on a location, select “create image” and enter a prompt describing a scene. Although Google promoted the technology for creative and professional uses—such as visualizing historical landmarks, urban-development projects or future real-estate designs—it could also generate realistic-looking disasters and geopolitical events that never occurred.
Open-source intelligence researcher Henk van Ess tested the system by creating scenes involving refugees near the U.S.-Mexico border, a nuclear facility in Iran, a vehicle crash in Amsterdam and a bomb crater near a hospital in Gaza. NPR reporters were also able to produce fictional images showing fires around Iran’s strategically important Kharg Island and severe flooding around the U.S. Capitol. According to the researchers, the tool did not initially reject those prompts, even though the resulting images could easily be mistaken for evidence of major breaking-news events.
Google said it was rolling back the feature because screenshots of generated images appeared to violate its policies. The company acknowledged that geospatial professionals had used the technology for legitimate purposes, but said stronger guardrails would be needed before it could safely return. The pause does not mean Google is abandoning AI-powered geospatial technology altogether. Instead, it reflects an effort to reconsider how image generation can be separated from the trusted reference imagery that millions of people associate with Google Earth.
The controversy is particularly serious because satellite images have traditionally been treated as relatively dependable evidence. Journalists, human-rights organizations, military analysts and open-source investigators use them to confirm damage after attacks, locate military equipment, document mass graves, track natural disasters and verify videos uploaded from areas that reporters cannot safely reach. Google Earth is not always current, but its broad global coverage frequently serves as a reference point against which newer photographs and videos are compared.
By placing an AI generator directly inside that trusted environment, critics argued that Google blurred the line between observation and imagination. A fabricated satellite image created from an unrelated image generator may attract suspicion. A fake built around the correct geography, terrain, road network and building layout from Google Earth may appear far more credible because many of the surrounding details are real. This reduces the obvious visual errors that often expose conventional AI images and makes the fabrication harder for an ordinary viewer to recognize.
The potential consequences extend beyond misleading social-media posts. False imagery showing an attack on an oil terminal, damage to a military base or flooding around a government building could briefly move financial markets, inflame international tensions or force news organizations and government agencies to spend valuable time disproving a fictional event. During a fast-moving conflict, even a fake that is exposed within hours may already have reached millions of people.
Another danger is the so-called “liar’s dividend.” As synthetic imagery becomes more common, governments and other interested parties may dismiss authentic evidence as AI-generated. Real satellite photographs documenting military activity, civilian casualties or environmental destruction could therefore become easier to deny. The result would not simply be more false images, but a broader decline in confidence in genuine ones.
Google had already introduced several protective measures. Every image created through the feature included SynthID, the company’s invisible digital watermark for AI-generated content. Google said suspicious images could be checked through Gemini or Google Lens to determine whether the watermark was present. The company also said it blocked harmful categories of image generation and continually updated its protections.
Those protections, however, did not fully reassure investigators. Most social-media users do not stop to upload an image into a verification service before sharing it. Screenshots, cropped images and videos created from altered Google Earth scenes may also separate the content from the context and labels that originally identified it as synthetic. Researchers therefore recommend comparing suspicious images with independent sources such as the European Sentinel-2 satellite system, NASA and U.S. Geological Survey Landsat imagery, while also checking the claimed capture date, satellite name, weather conditions and shadows.
Before relaunching the feature, Google may need more visible labeling, stricter restrictions around military sites, disasters, borders and active conflicts, and clearer separation between generated scenes and authentic Google Earth imagery. Other possible safeguards include permanent on-image warnings that survive screenshots, limits on photorealistic output in sensitive locations, stronger detection of attempts to remove watermarks and built-in links allowing viewers to compare a generated scene with the original imagery.
Google’s decision illustrates a larger problem facing the AI industry. A technology may have legitimate uses in planning, education and design while still being capable of producing harmful material with very little effort. In this case, the risk was amplified because the generator was connected to a platform people have long used to establish what places actually look like.
The pause is therefore more than a correction to one Google Earth feature. It is an acknowledgment that generative AI can weaken the credibility of the very tools used to expose misinformation. Unless companies can make synthetic geospatial imagery unmistakably different from authentic observations, “deepfakes in the sky” could make it increasingly difficult for the public to determine not only what is happening on the ground, but whether photographic evidence can be trusted at all.
