Google Earth briefly made satellite-photo fakes a text prompt away.

A researcher typed a text prompt into Google Earth this week and got back a satellite-looking image of refugees near the Mexican border. The request was a single sentence. The output was indistinguishable, at a glance, from the kind of image a satellite pass would produce for a breaking news event.

Nano Banana 2, Google's image-generation model, launched in Google Earth earlier this week, according to Google's product blog. The Verge had flagged the rollout in advance and positioned it as a way to "reimagine" locations: an aerial view of a historical site, a flyover of a planned real estate project. The same week, a researcher showed what else that pipeline can produce.

Henk van Ess posted the demonstration on the site: fake "refugees near the Mexican border," a bomb crater near a hospital in Gaza, and a Sphinx crossed with the Statue of Liberty. The first two are not novelty prompts. They are the shape of a breaking news photo, and they come out of Google Earth looking like the real thing.

Google's response, posted by the official @NewsFromGoogle account on X in reply to van Ess, leans on two safeguards. The company says it "prevent[s] image creation on harmful topics and are continually updating our protections." The Verge, which broke the story, notes that van Ess's prompts produced those images anyway, on launch day.

The week before, The Verge had reported that researchers had reverse-engineered SynthID and could strip the mark from generated images while keeping the image itself intact. Van Ess took a different route. He used Google Earth to generate an AI-altered video, then ran it through Hive, a commercial AI-content detector. Hive flagged the video as human-made. The SynthID watermark was not the only verification layer the ecosystem depends on, and it was not even the one that caught this attempt.

Google's own description of SynthID, as quoted in The Verge's piece, is that the watermark is "difficult to remove, but not impossible to miss or bypass." That sentence is the gap. A watermark that survives a careful adversary is one thing; a watermark the company itself calls bypassable is a courtesy signal, not a verification system.

After The Verge's article published on July 31st, Google announced it was rolling the feature back. Futurism confirmed the rollback shortly after. The product decision is live. The underlying pipeline is now part of the visual vocabulary of news events, with or without a Google Earth-branded button: text prompt, to satellite-looking image, to flyover video.

So what does a reader do when the next "satellite photo" lands in a feed? Two moves that work today, both built around the gap the watermark leaves.

First, check the watermark. Google's recommendation is to ask Gemini via the @verifyai tag, or use Lens in Search, to see whether an image carries the SynthID mark. A "no" is not a verdict. The watermark can be stripped, as the reverse-engineering work showed. But a clean image with a confident provenance claim is the first thing to verify.

Second, and more important, cross-check the scene against an independent satellite pass. Sentinel-2 and Landsat are the two public constellations that publish medium-resolution imagery on a regular cadence. If a "satellite photo" of a specific event claims to be from a specific day, those passes can confirm or contradict the scene. The metadata question (when the image was taken, by which sensor) is the one that AI generation cannot fake without an entirely different forgery.

Nano Banana 2 may come back to Google Earth in some form. The verification gap it exposed will not close on its own. A generative tool that produces evidence-shaped output is a new misinformation surface. The defenses available (watermarks, AI-content detectors, on-the-fly topic filters) are not yet at the level of "this image was taken by this satellite on this day." Until they are, the reader's job is to ask which of those three an image can actually claim.