Google Earth’s AI Experiment Lasted One Day
By Suad Seferi ·
Google added an AI image generator to Google Earth on Thursday. By Friday, the feature was gone. The short-lived experiment allowed people to select a real location and modify the surrounding satellite, aerial or 3D imagery with a written prompt. Within hours, users had produced fictional scenes involving refugee camps, damaged hospitals, military infrastructure and disasters at recognisable locations. The images did not replace the official imagery seen by other Google Earth users. They were private generated outputs, marked as AI-created and protected by Google’s SynthID watermarking system. But they could still be captured as screenshots, removed from the original interface and shared elsewhere as if they showed a real place or event. That possibility was enough to alarm researchers and open-source intelligence specialists. On July 31, less than a day after the wider release, Google said it was rolling back the feature while it worked on stronger safeguards. The company acknowledged that people place particular trust in Google Earth and said it had seen screenshots of generated imagery that appeared to violate its policies. The speed of the reversal made this more than another technology launch that needed minor adjustment. It exposed a problem that should probably have been considered before the tool was made available: Google Earth is not an ordinary creative platform. People use it because they expect it to show the physical world. A real location with an invented history Researcher Henk van Ess demonstrated how the feature could be used to place fictional events over genuine coordinates. His examples included refugees near the US-Mexico border and a bomb crater close to a hospital in Gaza. Other reports described generated scenes involving a nuclear facility in Iran and disasters at public locations. None of these images became part of Google Earth’s shared map, but that distinction would not necessarily travel with a screenshot posted on social media. This is where the design became risky. A fabricated image created from nothing may still deceive people, but an image carrying the visual language of satellite mapping has an additional layer of authority. It appears tied to a real place. It may include familiar terrain, buildings and coordinates. To the average viewer, it can look less like an illustration and more like documentation. Satellite-style imagery already plays an important role in reporting on wars, border movements, environmental damage, construction and humanitarian emergencies. Journalists and investigators regularly compare images taken at different times to establish what changed at a particular location. Introducing freely generated material into that environment makes verification harder, especially once the original labels and controls are no longer visible. Google initially pointed to SynthID and other verification tools as safeguards. Users who doubted an image could, in theory, check it through Gemini or Google Lens to determine whether it carried an AI watermark. Those protections have value, but they depend on people suspecting that an image may be false in the first place. Most users do not submit every satellite image they encounter to a separate verification service. Van Ess also said he was able to create material without the system refusing or softening his prompts. He demonstrated that at least one external AI-detection system could be misled by generated media originating from Google Earth. The idea itself was not useless There are legitimate reasons to generate possible futures over real locations. Architects could visualise a planned building in its surroundings. Urban planners could test how a redesigned neighbourhood might look. Environmental organisations could illustrate the possible effects of drought, flooding or deforestation. Emergency services could use simulated disasters for training and preparedness. Google’s wider work on Earth AI is already being used for practical analysis rather than fictional image generation. Its geospatial models can help identify flood risks, monitor wildfires, examine water systems and combine satellite information with population and infrastructure data. Google says organisations are using these systems for disaster response, environmental monitoring and public health. That work is not the problem. There is an important difference between using AI to analyse the physical world and using it to create a convincing image of something that never happened there. The removed feature crossed that line too casually. A planning visualisation can be useful when it is permanently and unmistakably presented as a proposal. A disaster simulation can help emergency teams when it remains inside a controlled exercise. The danger begins when the same image can be detached from its purpose and circulated as evidence. Google said geospatial professionals had already found useful applications for the feature. Its decision to withdraw it suggests that usefulness alone was not enough to justify the risks of releasing it in its original form. Why the platform matters Generative image tools have been capable of producing fictional wars, disasters and political events for years. What made this case different was the platform hosting the generator. Google Earth has built its reputation by helping people inspect places they may never visit. It is used in classrooms, investigations, court cases, property searches, planning and journalism. Its imagery may not be live, and it may come from different collection dates, but users generally understand it as a record of something that physically existed. That expectation is part of the product. Adding an image generator did not automatically destroy the reliability of Google Earth’s official data. The generated scenes were separate and watermarked. But the feature borrowed the authority of the platform while making it unusually easy to create images that contradicted what the platform normally r…