Google Earth’s AI Image Tool Lasted One Day Before Misinformation Fears Forced a Retreat

Google has pulled an artificial intelligence image tool from Google Earth only one day after launch, after users showed how easily it could turn genuine maps into convincing fictional scenes. The reversal matters because Google Earth is widely treated as a reference for what a place actually looks like, even though the generated pictures never replaced its public map imagery.

KEY NUMBER
The feature remained available for about one day before Google withdrew it to build stronger guardrails.

Why this one-day experiment matters

On July 30, Google introduced image generation on the web version of Google Earth. A user could choose a location, press “create image” and ask Nano Banana 2 to show a historical reconstruction, a proposed building or another imagined scene grounded in the platform's real-world imagery. Google's original launch post and subsequent rollback update presented useful cases for education, architecture and planning.

The same grounding that made those examples useful also made fabricated scenes more persuasive. An invented building or crater could inherit the viewing angle, terrain and visual authority of a familiar mapping service. NewTqnia's reading is that the failure was not simply that an image model could make a fake, which was already possible. Google reduced the work needed to anchor that fake to a real place inside a product people routinely use as evidence.

What users demonstrated

Open-source researcher Henk van Ess tested prompts involving a refugee camp, a fictional nuclear facility, a road crash and damage near a hospital. His documented tests of the Google Earth feature said those prompts were not refused. He also reported that an altered screen recording was not identified reliably by an outside detector.

Those examples do not prove the tool would create every harmful request or that every watermark could be defeated. They do show a gap between Google's stated protections and one researcher's real-world testing. The Verge's examination of the generated scenes also stressed that believable flyover-style images could travel outside Google Earth, where viewers might not see their original context.

What Google did, and what it did not say

Google updated its launch page on July 31, saying people had shared screenshots that appeared to violate its policies. The company rolled back the feature while working on stronger guardrails. It did not specify the exact policy violations or give a date for a return.

Google emphasized two boundaries: generated images did not become part of the main Google Earth experience for other users, and outputs were marked as AI generated. Reuters confirmed the one-day launch and withdrawal, while noting the broader misinformation concern around fake scenes placed over real landmarks.

Why a watermark is useful but insufficient

A digital watermark can help a compatible verification service identify an image made by a particular model. It is valuable provenance, especially when the original file remains intact. It is not the same as a visible warning that survives every screenshot, crop, video capture or repost.

The harder problem is social rather than purely technical. A dramatic image can spread before anyone checks it, and later uncertainty can also be used to dismiss genuine evidence as synthetic. A safer design therefore needs several layers: refusal rules for sensitive scenarios, conspicuous labels, durable provenance, friction around exporting risky content and independent ways to compare claims with dated imagery from other satellites.

What happens next

The important test will be whether a relaunched feature changes more than its prompt filter. Google could restrict transformations involving conflict, disasters and sensitive infrastructure, retain stronger origin information through common editing and sharing steps, and make synthetic status obvious without requiring a separate Google check.

The episode offers a wider lesson for consumer AI: trust inherited from the host product can amplify risk. A creative tool placed inside a map, health app or financial service is not judged only by what its model can generate, but also by the authority users attach to the surrounding interface. Google reacted quickly; whether it solves that deeper design problem will matter more than the speed of the rollback.