AI 安全與產品工程
Google Earth Pulled Its Image Generation Feature After One Day as SynthID Failed to Stop Fake Satellite Images From Spreading
After Google connected Nano Banana 2 to real-world geospatial imagery, users quickly generated and shared screenshots of fabricated explosions, airstrikes, and other scenes. The company withdrew the feature the next day. Although generated images carried watermarks and were not added to shared maps, the incident shows that provenance labels cannot replace input restrictions and interface isolation.
On July 30, Google added “Create image” to the web version of Google Earth, allowing users to use satellite, aerial, or 3D imagery from their current view as a scene and have Nano Banana 2 generate images such as historical reconstructions, architectural plans, or cities of the future. Less than a day after the feature launched globally, social media users and news outlets showcased fabricated airstrikes, exploding buildings, and crowds at borders. Google promptly rolled back the feature, saying it would add stronger safeguards.
This did not mean that generated content directly contaminated Google Earth’s shared basemap. Google said results were visible only to their creators, would not appear in the main Earth experience for other users, and included an AI watermark. The problem was that the product placed a trusted geospatial basemap, precise locations, and a photorealistic image generator within the same workflow. Users could move generated results outside the original interface simply by taking screenshots. The original UI, interaction state, and machine-readable provenance data could disappear when images were reshared. SynthID also requires recipients to actively verify content using Gemini or Lens, so it cannot automatically prevent misidentification in social media feeds.
Of greater concern to engineering teams is the inconsistency of the safeguards. Testing by The Atlantic found that some violent scenes were rejected in the standard Nano Banana editing workflow but could be generated through Google Earth’s location-based interface. This suggests that safety classifiers, prompt rewriting, or contextual policies did not maintain the same boundaries after product integration. A model passing safety tests does not mean every interface that uses it is equally safe.
If the eventual fix merely expands keyword blocking, users may still bypass it with euphemistic descriptions. A more comprehensive design should account for high-risk crisis locations and real-time events, classification before and after generation, prominent visual labels that cannot be cropped out, C2PA-style provenance credentials on export, and watermarks that remain verifiable after screenshots are taken. Google has not disclosed when the feature will return, its policy-violation rate, or the safeguard test suite it uses. For now, it is only possible to confirm that the feature was withdrawn—not that the underlying problem has been resolved.