Google Turned Earth Into a Deepfake Canvas — and Pulled It Within a Day

Last week Google gave the entire planet a tool to fabricate it, then took it back twenty-four hours later. The speed of the retreat is the reassuring part of this story. The reason it was necessary is the part worth thinking hard about.

What Google Launched, and Unlaunched

On Thursday, 30 July, Google added an AI image-generation feature to Google Earth on the web, powered by its Nano Banana 2 model. The pitch was benign and even charming: zoom to any location, type a prompt, and see it reimagined — Pompeii before Vesuvius, a landmark turned into an illustrated infographic, a neighbourhood visualised for real-estate planning. Within hours, the open-source investigator Henk van Ess and newsrooms including NPR and Futurism showed the other use of the same button. He generated a fictitious nuclear power plant in Iran and refugee camps at the US–Mexico border, both rendered convincingly onto real geography. Others produced explosions, military sites, and floods that never happened.

By Friday, 31 July, the feature was gone worldwide. Google's statement was unusually candid about the stakes: people uniquely trust Google Earth for a reliable view of the world, it said, and while geospatial professionals had found legitimate uses, others were sharing generated imagery that appeared to violate its policies. So it was rolling the feature back while it built stronger guardrails. Launch to retreat: about one day.

Why the Watermark Wasn't Enough

Google's defence rested on two safeguards. Every generated image carried a SynthID watermark marking it as AI-created, and the images did not appear inside the shared Google Earth experience that other users see. Both claims are true, and neither solved the problem.

A watermark travels with a file; it does not travel with a screenshot. The moment someone captures a fabricated satellite image and posts it to social media, the SynthID marker and the surrounding context fall away, and what remains is a photorealistic picture that looks exactly like the tool people have trusted for two decades to show them the real world. The safeguard was engineered for a world where images are viewed inside Google's own products. Disinformation does not live in that world. It lives in the crop, the repost, and the caption — precisely where the watermark cannot follow.

The Deeper Damage: The Liar's Dividend

The most serious harm here is not any single fake. It is what open-source analysts immediately named: the erosion of Google Earth as a shared reference for truth. For years, investigators verifying atrocities, troop movements, or disaster damage in places journalists cannot reach have leaned on satellite imagery as a fixed point — something closer to evidence than to opinion. A convincing deepfake layer over that imagery does not just enable false claims. It enables the opposite move, the one disinformation researchers call the liar's dividend: once believable fakes exist, anyone can dismiss a genuine, damning image as "probably AI." A government caught by a real satellite photo now has a ready defence. The fake does not need to be deployed to do its work; its mere possibility is enough to make every real image arguable.

That is why the reaction from the verification community was not proportionate to any one image that was actually made. It was a reaction to a category of trust being put at risk by a product launch that treated planetary reference imagery as just another canvas for a text prompt.

Claim and Counter-Claim

Google's defenders can make a fair case. The company moved in twenty-four hours, faster than almost any regulator could compel. The safeguards it had built were real, the abuse was the work of people deliberately stress-testing the tool, and pulling the feature rather than defending it shows a functioning safety reflex. On this reading, the system worked: a risk surfaced, and the company responded before real harm spread.

The counter-case is harder to dismiss. This risk did not require a genius to foresee; that a photorealistic geographic image generator could fabricate sensitive locations was obvious before launch, not after. The safeguards Google leaned on were the ones it must have known were weakest against screenshots. And the damage a retreat cannot undo has already been done: the proof of concept exists, the headlines exist, and the seed of doubt over every future satellite image is planted. A company does not get to un-ring that bell by disabling a button. The honest verdict sits in between — a genuine and creditable reflex to withdraw, wrapped around a launch decision that should never have cleared review.