Google pulls Earth AI image tool one day after launch
Google rolled back a Nano Banana 2-powered generator that let users paint fabricated imagery onto satellite maps, citing policy-violating outputs being shared online.
Google has withdrawn a generative image feature from Google Earth roughly a day after switching it on, after researchers and journalists warned it would make fabricated satellite imagery trivially easy to produce and share. The tool, driven by Google's Nano Banana 2 image model, let users generate synthetic imagery and place it directly onto Earth's mapped view.
In a statement, the company said users had been circulating screenshots of generated imagery that appeared to breach its policies, and that it was rolling the feature back while it built stronger guardrails. Google did not give a timeline for reinstatement or say what those guardrails would consist of.
The objection raised most consistently was not that the model could make convincing pictures, but where it was placed. Google Earth functions as a reference layer for open-source investigators, newsrooms, conflict monitors and environmental researchers, many of whom cite it as corroborating evidence. Embedding a generator inside that surface risked collapsing the distinction between the imagery archive and freehand illustration, and did so inside a product whose screenshots carry implicit provenance from Google itself.
The rollback is unusually fast for a shipped Google AI feature and follows a pattern of the company pushing image generation into general-purpose products, then retreating when the outputs escape their intended framing. Critics also noted the limits of the fix: models capable of altering aerial and satellite imagery are widely available from multiple vendors, so removing the in-product path narrows convenience rather than capability. What it restores is Google Earth's status as a surface that does not itself manufacture the pictures it displays.
Why it matters
The episode marks a boundary that a major platform has now conceded exists: some product surfaces derive their value from being treated as evidence, and generative editing is incompatible with that role regardless of labelling. As image models get folded into maps, search results, documents and camera rolls, this is the first high-profile case of a vendor reversing that integration on provenance grounds alone — a precedent other platform teams weighing similar features will be asked about.