Google rolls back AI-generated satellite imagery in Google Earth

Google briefly enabled an AI feature inside Google Earth that let users generate altered versions of satellite and aerial imagery, then rolled it back within a day after examples of abuse spread online. Ars Technica reports the tool combined Google Earth’s real-world imagery with Google’s image generator, letting prompts rewrite scenes that many people treat as a reference record. Google said the feature was being rolled back after criticism highlighted how easily the output could be used to fabricate “evidence” tied to specific places.

The episode lands in a familiar gap between what generative tools can do and what the product is for. Google Earth is not just a map; it is routinely used by journalists, researchers, and investigators to corroborate locations and timelines, and its authority comes from the assumption that the underlying imagery is documentary. When the same interface offers a button to “generate” a new version of that imagery, the burden shifts to the viewer to prove what is real, not to the publisher to keep the record clean. That matters most in the exact cases Google’s own marketing leaned on, such as visualising proposed real-estate projects or reconstructing historical sites: the tool’s selling point is plausibility.

Critics focused on how the feature compresses a multi-step forgery into seconds. Ars Technica cites Bellingcat founder Eliot Higgins sharing an AI-modified image depicting an outsized statue over the White House, and independent investigator Henk van Ess describing scenarios such as staging refugees near the Mexican border or placing a nuclear plant in Iran. Those are not just generic deepfakes; they are fabrications anchored to recognisable coordinates in a product people already use to check claims. Google responded by pointing to its SynthID watermarking system and said images generated in Google Earth would carry the mark, with verification available through Google tools such as the Gemini app or Lens.

Watermarking, however, assumes the audience will check, and that the checking tools will be used by the people most likely to be misled. It also assumes the image remains intact as it travels across social platforms, screenshots, crops, and re-uploads—exactly the distribution chain where context is routinely stripped. The rollback suggests Google judged that adding provenance metadata was not enough to protect the core utility of Earth as a quasi-public reference library.

Google’s product manager announced the feature in a blog post on July 30. By the time Ars Technica tested it after the rollback notice, the button was already gone.