Alphabet's Google rolled back an AI image-generation feature in Google Earth on July 31, just one day after launch, after users created fake satellite imagery of floods, refugee camps, and military attacks that appeared to violate the company's policies.
The feature shipped on July 30 with no waitlist and no enterprise-only tier: a "create image" button in Google Earth on the web, powered by the Nano Banana 2 model, let any user zoom to any location on the globe, type a text prompt, and receive a photorealistic image built from Google's satellite, aerial, and 3D data of that exact spot. Global availability from day one. It lasted roughly 24 hours.
By July 31, AI researcher Henk van Ess had already used the tool to generate fake satellite imagery of a refugee camp along the US-Mexico border and a nuclear plant in Iran. Other users created images depicting a nonexistent flood and what appeared to be a 9/11-style attack. More troubling still, fake images of a damaged US base in Bahrain circulated in Iranian media, apparently generated using Google Earth imagery and Google's AI. Google's official statement, posted to X by its News account, confirmed the decision: "We've seen geospatial professionals using this feature for a range of useful purposes, however we've also seen people sharing screenshots of generated imagery that appear to violate our policies. So we're rolling back this feature in Google Earth while we work on implementing stronger guardrails."
Deepfake tools are not new. What made this particular rollback urgent is the platform it happened on. Google Earth has spent two decades earning a specific kind of trust: journalists, researchers, human rights investigators, and open-source intelligence analysts use it to verify whether photographs are real. Bellingcat, the investigative outlet whose methodology depends on satellite imagery, has built entire investigations on the assumption that Google Earth shows you what is actually there. A senior researcher at Bellingcat told NPR the feature represented a change of speed rather than of kind, warning that streamlining fake satellite imagery production "is going to make it proliferate more." That framing cuts to the problem. The tool didn't create a new threat. It automated an existing one and embedded it directly inside the verification layer.
Fake satellite imagery already existed before July 30. It required skill, access, and time. Nano Banana 2 inside Google Earth removed all three barriers simultaneously. No application, no expertise, no delay. That combination, grounded photorealism plus the implicit credibility of the Google Earth interface, is precisely what made screenshots of the generated imagery spread as if they were real. The platform's authority became its vulnerability.
This isn't Google's first AI image rollback #
Alphabet has been here before. In early 2024, Google's Gemini image generation was suspended after the model produced historically inaccurate images of people, and CEO Sundar Pichai called the results "unacceptable." The company spent months rebuilding guardrails before restoring the feature. The Google Earth incident follows a structurally identical arc: a feature ships broadly, real-world misuse surfaces within hours rather than weeks, and the product is pulled while the team retrofits the safety layer that should have shipped first.
The Engadget headline called the tools "needless," which is probably too blunt, but it points at something real. Geospatial professionals do have legitimate uses for AI-assisted visualization: urban planners, environmental scientists, disaster response teams all stand to benefit from prompt-based imagery grounded in real satellite data. Google's statement acknowledged as much. The problem is the deployment sequence. Rolling the feature out to every user on the planet before testing whether the guardrails held under adversarial conditions is the decision that needs explaining, not the feature itself.
Google says it's working on "stronger guardrails" before restoring the feature. It's also worth noting, as the company pointed out in its statement, that generated images were labeled as AI-created and not stored in Google Earth itself. That distinction matters legally and technically. It didn't matter much in practice, because screenshots strip labels, and the imagery spread anyway.
The broader lesson isn't that AI image generation is incompatible with mapping platforms. It's that trust is asymmetric. Google Earth built its credibility over 20 years of showing the world as it is. A single viral cycle of faked disaster imagery can erode a meaningful portion of that credibility in a news cycle. Frankly, a 24-hour rollback is the right call. The harder question is why a product this obviously dual-use, layering generative AI onto the internet's most trusted geographic reference, cleared launch without a more adversarial red-team process first.
Also read: NXP Semiconductors is in talks to buy Ambarella for $3.3 billion to own the camera chips powering self-driving cars • The EU's AI deepfake labeling law takes effect Sunday with a 38-person enforcement squad already watching • Perplexity charges $200 a month for an AI agent that does Microsoft Copilot's job on Microsoft's own Windows