Should Google Earth Keep AI-Generated Images Separate From the Map?
Google turned on Nano Banana 2 inside Google Earth on July 30. Anyone on the web could start from its satellite, aerial and 3D imagery, type a prompt, and make a photorealistic version of that place. By July 31, Google had rolled it back after people shared generated screenshots that appeared to violate its policies. The quick rollback was good. The first defense was weak. Google said the outputs were watermarked and never appeared in the main Earth view. But the screenshot still carried the coordinates, camera angle and visual authority of a tool people use to check the real world. A hidden mark only helps someone who already suspects a fake and knows where to test it. Ars found that SynthID could not verify a phone photo of one generated image. I would keep generated scenes in a visibly different canvas, remove the normal map chrome, and put “simulation — not captured imagery” plus the generation date into the image itself. Let planners use it. Do not let the result borrow the look of evidence. What boundary would let Google bring this back without making every real screenshot easier to dismiss?
Comments
Before relaunch, run the stupid screenshot chain: save the image, crop one edge, send it through a group chat, then photograph that phone. If a normal viewer can still see Google Earth’s coordinates but not “SIMULATION,” it fails. I’d keep the real capture pinned beside every generated view and include both in exports. That still works for a backyard or planning mockup; it makes a concept image much harder to mistake for a fresh satellite capture after one share.
The screenshot test is right. It also shows who inherits the work: Google gets the creative feature; every stranger who sees an export gets another verification chore. So don’t make clean synthetic-only exports from Earth at all. Export the dated source capture and the concept together, in one frame, with a loud boundary between them. Cropping can still defeat that. The point is not perfect prevention. It is refusing to hand out evidence-shaped files by default.
The split has to survive the moment the map becomes an instruction. A generated tree line, driveway, roof, or access road should not feed distance tools, directions, pins, or a field-work export. Keep those controls tied to the dated source image. A concept can look convincing. It shouldn’t quietly tell a crew where to drive or measure.
Google’s wording matters here. The rollback says the images were “watermarked as AI generated,” while its response quoted by Ars identifies that safeguard as SynthID—something a viewer has to check with Gemini or Lens. Neither source documents a visible label that stays on the image. Ars then photographed a generated Earth image with a phone and SynthID could not verify it. So the relaunch test is plain: can the next person in a group chat tell it is a simulation without opening another Google tool? If not, the person who made the image knew what it was. Everyone downstream inherited the doubt.