Google has rolled back a newly launched AI image generation feature in Google Earth after researchers demonstrated that it could produce fabricated satellite imagery layered onto real-world coordinates. The feature, introduced on Friday, was pulled just one day later following a wave of public examples showing how easily the tool could create convincing but entirely artificial aerial scenes.
The tool relied on the company's Nano Banana 2 model, an AI system designed to generate photorealistic imagery from text prompts. It allowed users to zoom into any location on the globe, enter a short description, and receive a synthetic aerial view that matched the underlying geography and coordinate system. While the output was clearly artificial to the model's creators, the images were rendered directly onto authentic satellite views, making them difficult to distinguish from genuine reconnaissance or mapping data.
Dutch researcher Henk van Ess was among the first to highlight the potential for abuse. In a post titled "How to plant a nuclear plant in Iran," he showed that a single sentence could place a nuclear facility on Iranian territory, put refugees along a road near the Mexican border, or add a hospital with a bomb crater in Gaza. Each image was generated in seconds and appeared on real coordinates, as though the scene had been captured by an orbiting satellite.
Van Ess's demonstrations quickly spread across social media and prompted other news organizations to test the tool. BBC Verify produced images of a toppled Eiffel Tower, a sinkhole swallowing the Great Pyramid of Giza, and Russian tanks in Kyiv. The French news agency AFP generated explosions in Paris and an Islamic State training camp in Syria. The digital culture publication 404 Media added skyscrapers to rural areas and inserted a homeless encampment into Los Angeles. In every case, the fabricated scenes were placed on real coordinates, visually seamless enough to fool a casual observer.
The core concern is not that AI can create fake images—that capability has existed for years—but that Google chose to integrate such generative power directly into a platform widely used for verification. Journalists, humanitarian organizations, intelligence analysts, and disaster response teams rely on Google Earth as a primary reference for ground truth. The tool effectively allowed anyone to pollute that reference layer with synthetic content that looked like official satellite imagery.
Van Ess noted the asymmetry: it can take a formal government request to remove a single real building from Google Earth, but only a sentence to add a fake one. That asymmetry matters in conflict zones, where misinformation can have life-or-death consequences. A fabricated image showing military movements, bomb damage, or refugee flows could spread rapidly on social media before fact-checkers have a chance to intervene.
Google responded through its media relations account on X, saying it had seen geospatial professionals using the feature for legitimate purposes, but acknowledged the need for stronger guardrails. The company confirmed that none of the AI-generated images appeared in the main Google Earth platform, meaning the synthetic scenes were confined to the generation tool itself and did not alter the underlying base layer of satellite imagery. That distinction is important, but it does not address the broader risk: users could capture the generated images, share them outside the platform, and present them as authentic overhead footage.
The generated images did include Google's SynthID digital watermark, an invisible signature designed to survive screenshots and compression. SynthID is a promising technical measure, but it only works after the fact. Someone who sees an image on social media must know to look for the watermark and have the tools to detect it. In fast-moving situations, such as an active conflict or a natural disaster, verification often lags behind virality. By the time a misinformation researcher flags an image as synthetic, it may have already been viewed thousands or millions of times.
This episode is not an isolated incident for Google's generative AI products. The company paused Gemini's image generation of people in 2024 after the model produced historically inaccurate and racially inappropriate depictions. At the time, Google apologized and said it would fix the issues before relaunching the feature. The Google Earth rollback follows a similar pattern: a feature is released broadly, problems surface within hours, and the company retreats to retrofit safeguards that should have been in place from the start.
The pattern raises questions about Google's product development culture and its willingness to ship AI tools first and patch them later. In the case of image generation, the risks are not limited to factual inaccuracies. When AI-generated imagery is attached to real geographic coordinates, the damage can extend to international relations, military decision-making, and public trust in the platforms that are supposed to provide objective information.
For geospatial professionals, the potential harms are especially acute. Mapping agencies use satellite imagery as a baseline for constructing maps, monitoring land use, and planning infrastructure. If synthetic imagery becomes mixed with authentic scenes, the integrity of those maps could be compromised. Even if the compiled imagery is not distributed, the existence of a tool that can produce such realistic fakes could undermine the credibility of legitimate satellite photographs. Anyone who wants to dismiss genuine evidence of human rights abuses or war crimes could claim it was generated by an AI model.
Google's decision to pull the feature is a recognition of these risks. But the underlying technology remains in the company's arsenal, and it is likely to return with additional restrictions. The challenge is that no set of guardrails can anticipate every malicious use. The same text prompts that create absurd images like a toppled Eiffel Tower can also be used to generate realistic-looking damage to critical infrastructure, military bases, or contested borders. As AI models become more powerful, the gap between generated and real imagery will continue to narrow.
In the short term, researchers and journalists have called on Google to maintain a public log of AI-generated images and to embed more robust authentication mechanisms in its platforms. Some have suggested that any AI-generated aerial image should carry a visible watermark or be denied access to real coordinate data altogether. Others have argued that Google should require verified credentials for users who want to access geospatial AI features. None of these ideas are foolproof, but they reflect a growing expectation that technology companies must build responsibility into their products from the beginning.
For now, the Google Earth AI tool is offline. The episode serves as a reminder that generative AI does not exist in a vacuum; when it is connected to spatial data, the consequences become physical. A synthetic image is no longer just a digital artifact—it becomes a claim about a place on the planet. That claim can influence how people perceive conflicts, disasters, and political changes. If the underlying maps can be altered with a single sentence, then the ground truth itself is up for grabs.
Google has not announced a timeline for the feature's return. The company's initial statement emphasized that it was rolling back the tool to implement stronger guardrails and would take time to evaluate how to prevent misuse. In the meantime, the bar for trust in satellite imagery may have been permanently raised. Anyone viewing an overhead image must now ask an uncomfortable question: was this scene really there, or was it placed there by a machine?