Google Pulls AI Image Generator From Google Earth Just One Day After Launch, Following Fake Refugee and Nuclear Site Images

Timeline: Why Was the Feature Pulled After Just One Day?

According to a report from Inside (inside.com.tw), Google announced on July 31, 2026 that it was withdrawing the AI image-generation feature it had launched inside Google Earth just one day earlier. Inside describes the turnaround as "withdrawn less than 24 hours after launch."

Per the same Inside report, the feature — built on Google's Nano Banana 2 model — first opened to Google Earth web users worldwide on July 30, 2026, and was pulled the very next day, meaning the entire public run lasted under 24 hours.

How Did the Google Earth AI Image Generator Work?

Inside reports the tool ran on Google's Nano Banana 2 model. A later report from Infosecu (infosecu.technews.tw) describes the exact user flow: in the web version of Google Earth, a user zoomed in to any location, clicked the "create image" button, and typed a text description of what they wanted to see — and the tool generated it.

What Safety Measures Did Google Say It Had in Place?

In an official statement quoted by Inside, Google said it "takes misinformation seriously," noting that every image created with Nano Banana in Google Earth carries a SynthID digital watermark, so that anyone unsure whether an image is real could check it through the Gemini app or Lens in Search. Google's statement also said the system was designed to block requests involving harmful topics.

What Did Investigative Journalists' Tests Reveal?

Multiple named testers, cited across both Inside and Infosecu, described the opposite result from what Google's stated safeguards would predict:

• Henk van Ess, a Dutch investigative journalist and AI researcher, told Inside he used short prompts to generate fictional refugees near the US–Mexico border, a fabricated new nuclear facility inside Iran, and a fake bomb crater next to a hospital in Gaza. He said none of these prompts were rejected, softened, or flagged for a prompt change.
• Joseph Cox of 404 Media, in testing cited by Engadget and relayed by Inside, generated a fictional explosion crater in Los Angeles and a nonexistent protest crowd outside Google's Mountain View, California campus.
• NPR, per Infosecu, tested the feature and generated an image of Kharg Island — which Infosecu notes handles 90% of Iran's oil exports and is described as central to Iran's economy — shown engulfed in flames, as well as an image of the US Capitol flooded. Neither event occurred.
• Van Ess also told NPR, as reported by Infosecu, that his tests covering migrants at the Mexican border, an Iranian nuclear plant, a plane crash in Amsterdam, and a bombed-out hospital in Gaza were rejected "not a single" time by the system.

Why Are the Fabricated Images Hard to Detect?

Van Ess converted Google Earth-generated content into a video and ran it through the third-party AI-detection tool Hive, according to Inside. Hive's own analysis, as quoted in the report, returned the following:

Separately, The Verge — cited by Inside — advised that verifiers checking a suspicious image can go beyond Gemini or Google Lens and cross-reference satellite data sources such as Sentinel-2 and Landsat to confirm the actual capture date, satellite source, and real terrain.

Questions About Google's Internal Safety Testing

Evan Hill, a visual-forensics investigative reporter at The Washington Post, questioned Google's internal review process in remarks carried by Infosecu: "I am very curious how this idea was red-teamed internally at Google — or whether it was red-teamed at all."

Google's Official Response and the Decision to Pause

In its statement, quoted by Infosecu, Google said: "We saw professionals in the geospatial field using this feature for a range of practical purposes. But we also saw people sharing screenshots of generated images that appeared to violate our relevant policies. As a result, we are pausing this feature in Google Earth while we work to improve our safety protections."

What This Means

The gap between Google's stated safeguards and the test results is the central tension in this episode. Google said every generated image carries a SynthID watermark and that harmful-topic requests are blocked (per its own statement, cited by Inside) — yet Henk van Ess told both Inside and NPR (via Infosecu) that his refugee, nuclear-facility, plane-crash, and hospital-bombing prompts were never once rejected, softened, or flagged. Separately, footage built from the feature's output scored just 1% on Hive's AI-video detection despite carrying a SynthID watermark, according to Inside — meaning the labeling Google pointed to as its verification path did not register the content as synthetic in that test. Evan Hill's public question about whether Google red-teamed the feature at all, reported by Infosecu, sits directly alongside Google's own admission — in the statement announcing the pause — that it saw shared screenshots "that appeared to violate" its policies. Google's fix, per its own statement, is to keep the feature paused while it works on improved safety protections; no timeline for that work is given in either report.