Google Earth AI tool exposes geospatial disinformation risk

Google has withdrawn an experimental image-generation feature from Google Earth after users demonstrated how easily it could fabricate convincing scenes of wars, disasters and sensitive security installations.

The tool allowed people to select a location in the web version of Google Earth and transform its satellite, aerial or three-dimensional imagery through written prompts. Within seconds, ordinary neighbourhoods could appear bombed, flooded or occupied by military forces, creating pictures that closely resembled authentic geospatial evidence.

Powered by Google’s Nano Banana image model, the “create image” function was designed for visual experimentation. Users could redesign buildings, imagine alternative urban landscapes or depict how locations might have appeared during another historical period. The generated material was confined to individual user sessions and did not replace imagery in Google Earth’s public database.

However, demonstrations showed that those safeguards did little to prevent screenshots from being detached from their original context and circulated online. Test images included fictional bomb craters, damaged cities, refugee camps near the United States-Mexico border and a fabricated nuclear facility in Iran. Other prompts turned quiet towns into battlefields filled with smoke, ruined buildings and armoured vehicles.

Google disabled the function less than a day after its launch, saying it had observed people sharing screenshots that violated company policies. The company said it was taking misinformation risks seriously and would strengthen safeguards before considering whether to restore the capability.

The rapid reversal has highlighted the growing tension between generative creativity and the evidentiary value of mapping platforms. Google Earth imagery has long been used by journalists, researchers, human-rights organisations and open-source intelligence specialists to investigate military activity, environmental destruction, infrastructure projects and damage caused by natural disasters.

Placing an image generator directly beside trusted geographical material risks blurring the distinction between recorded reality and manufactured scenarios. Even when an image is created for entertainment or design purposes, a cropped screenshot can remove warnings, labels and other visual indicators that identify it as synthetic.

Google said generated images contained SynthID, an invisible digital watermark designed to identify material produced or altered by its AI systems. The company also pointed to verification tools available through Gemini and Google Lens. Yet digital watermarks are not a complete defence because online audiences often judge an image before checking its provenance, while reposting, compression and further editing can complicate detection.

Research into synthetic-image verification has found that highly realistic generated pictures can be difficult for people to distinguish from genuine photographs. Academic work has also demonstrated that watermarking systems may be vulnerable to removal or manipulation, reinforcing calls for multiple layers of protection rather than reliance on a single technical marker.

The risk is particularly acute during armed conflicts and humanitarian emergencies, when misleading imagery can spread faster than investigators can verify it. Fabricated satellite views could be used to support false claims about troop movements, attacks on civilian areas, clandestine construction or damage to strategic facilities. Authentic evidence may also be dismissed as artificial once audiences become accustomed to realistic geospatial fakes.

Open-source investigators usually authenticate satellite material by comparing dates, coordinates, shadows, weather conditions, building layouts and imagery from multiple providers. Generative tools can raise the cost of that work by producing large volumes of plausible but false material that demands examination.

The episode also illustrates a broader problem facing technology companies as image models become embedded in mainstream products. Features developed for architects, planners, educators and creative users can be repurposed almost immediately for propaganda, fraud or harassment. Safety testing conducted before release may not anticipate the speed or ingenuity with which users discover harmful applications.

Google has continued expanding Earth AI for purposes including disaster forecasting, public-health planning, building-damage assessment and environmental monitoring. Its geospatial models can analyse satellite data to identify patterns that would take human researchers far longer to detect. The withdrawn generator differed because it created fictional visual content on top of recognisable real-world locations, rather than analysing existing geographical evidence.