Google Earth Pulls Satellite Image Editing Tool After Just One Day

Google has disabled a new Google Earth feature that allowed users to modify satellite imagery with text prompts after researchers demonstrated it could generate convincing deepfakes of sensitive locations.

AI & Culture Desk
Updated 5d ago 2 min read
Google Earth Pulls Satellite Image Editing Tool After Just One Day

Google has deactivated a Google Earth feature launched on Thursday that allowed users to manipulate satellite imagery via text prompts. The tool, which relied on the company's Nano Banana 2 model, was intended to let users modify geographic landscapes. However, the company pulled the capability just one day after its release, citing instances where users shared screenshots of generated imagery that violated company policies. According to The Verge, Google acknowledged the unique trust users place in the platform for a reliable view of the world, prompting the decision to roll back the feature while the company works on implementing stronger guardrails.

The vulnerability of the system was highlighted by Henk van Ess of Digital Digging, who used the tool to generate synthetic images of refugees near the Mexican border and a bomb crater adjacent to a hospital in Gaza. While Google initially defended the feature by noting that images included digital watermarks and that the system was designed to prevent "image creation on harmful topics," van Ess reported that his testing faced no refusals or content softening. He further demonstrated that the generated media could bypass Hive’s AI detection tools, undermining the effectiveness of the company's watermarking strategy.

Before the rollback, Google maintained that the generated images were isolated from the main Google Earth experience, meaning they were not visible to other users. The company stated that it had observed geospatial professionals utilizing the feature for productive purposes before the policy violations occurred. Despite these safeguards, the ease with which the model could be prompted to create sensitive, fabricated scenarios suggests a significant gap between the tool's intended creative use and its potential for misuse in generating geospatial misinformation.

The rapid reversal underscores the friction between deploying generative AI and maintaining the integrity of platforms that function as authoritative records of physical reality. By allowing users to alter satellite data, the tool created a pathway for deepfakes that could be easily disseminated, regardless of whether the images were hosted on the official platform. The incident highlights the difficulty of applying standard content moderation to generative models, particularly when the output is designed to mimic verifiable, real-world data.

Google has committed to implementing stronger guardrails before potentially reintroducing the capability. The situation serves as a case study in the risks of embedding generative AI into legacy mapping software without sufficient threat modeling. As developers continue to integrate these models into consumer-facing platforms, the challenge remains to create boundaries that prevent the fabrication of convincing, misleading representations of the physical world.

Sources

  1. 01 Google Earth’s AI deepfake tool only lasted one day — The Verge
#google #geospatial #deepfakes #trust-and-safety #generative-ai