Google Earth’s AI Editor Collapsed After One Day of Deepfake Alarm

Google Earth’s AI Editor Collapsed After One Day of Deepfake Alarm
Google wanted to turn its mapping platform into a creative canvas. Instead, its experiment exposed how quickly a trusted record of the physical world can become a factory for plausible fakes.

Google launched Nano Banana 2 in the web version of Google Earth with aspirational use cases: revisiting cities as they looked decades ago or visualizing a proposed community basketball court. The company pitched the feature as a way to combine its satellite, aerial and 3D imagery with generative AI—useful, it argued, for projects such as real-estate visualization and historical reconstruction.

Researchers and verification specialists saw a more immediate use. Henk van Ess demonstrated that a simple prompt could place “refugees near the Mexican border,” a nuclear plant in Iran and a fatal crash in Amsterdam over Google’s real imagery. “What on earth is Google doing?” he wrote. Critics argued that embedding image generation within a service routinely used to authenticate locations lowered the barrier to manufacturing persuasive misinformation.

Google’s first response stressed that every generated image carried its SynthID watermark and that harmful prompts were blocked. But van Ess said his tests told a different story: “Nothing was refused, nothing was softened, and nothing suggested I try a different prompt.” He also showed an AI-altered Google Earth video could fool an outside detector, reinforcing concerns that watermarks may not survive the screenshots, re-encodes and screen recordings through which false images often spread.

Within a day, Google rolled the feature back. The company acknowledged that professionals had found legitimate uses, but said screenshots of generated imagery “appear to violate our policies” and that it would implement “stronger guardrails” before reconsidering the tool. Google also emphasized that generated images did not alter the main Earth experience for other users. Yet the episode left the central problem unresolved: a platform valued as evidence had briefly made the evidence itself editable.

Continue reading https://foxvector.com/stories/019fd199-0c9d-0dbb-7265-3b9af270a4ee