Google has rolled back its brand-new AI-powered Google Earth image generator worldwide less than 24 hours after launch.
As I reported on July 30, Google recently added the ability to reimagine any Google Earth imagery based on Gemini text prompts. For example, users could use the feature to visualize the city of Pompeii before the eruption of Vesuvius, or turn famous landmarks into illustrated infographics populated with data pulled directly from Gemini’s multimodal output.
How Google Earth AI Was Exploited
Unfortunately, the launch didn’t go entirely as planned after investigative journalist Henk van Ess and newsrooms including NPR and Futurism proved that the tool could be used to generate potentially misleading content. Van Ess used Google Earth to generate imagery of refugees near the Mexican border and a fictitious nuclear power plant in Iran.
Google Earth’s satellite imagery is relied on by millions as an accurate depiction of the world, meaning creative, or even highly misleading, interpretations can become especially convincing when grounded in Google’s real-world data. Google emphasizes that these AI-generated images were never presented for others to see in the main Google Earth experience and were watermarked as AI-generated. However, it has decided to pause the feature to work on improving safety guardrails.
Why Google Pulled The Feature
Google subsequently announced the rollback of the feature in an update to its original announcement on The Keyword blog:
“Update, 7/31: We know that people uniquely trust Google Earth for a reliable view of the world. 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. It’s important to note that generated images didn’t appear in the main Google Earth experience for others to see and were watermarked as AI generated.”
It’s not immediately clear exactly how these safety guardrails will be improved, but there’s a world of difference in impact between educational infographics about national monuments and potentially believable images of fictitious geopolitical events. Google will have to navigate this carefully if the feature is to remain both useful and safe.
How Google’s AI Watermark Defences Failed
As van Ess discovered, Google’s SynthID system, used to invisibly watermark Google Earth’s AI-generated imagery, was easily defeated simply by taking a photo of a generated image. Furthermore, the Hive AI detection service rated a screen-recording of Google Earth’s AI-generated images as only 1% likely to be AI-generated video. This means SynthID, as it stands, is not enough to prevent simple manipulations of Google’s AI images from being passed off as real.
Google, understandably, wants to maximize opportunities afforded by its generative AI technologies. However, coupling real-world satellite data with photorealistic image generation creates unprecedented misinformation risks that Google must guard against if the feature is to survive. As of publication, Google’s AI image generation in Google Earth remains paused. However, testing Google Earth on the web today reveals that previously generated AI images remain accessible.