Google Earth AI Disaster Image Tool Pulled Within 24 Hours

A One-Day Experiment in Manufacturing Reality

On July 30, 2026, Google switched on a feature in the web version of Google Earth that let users type a prompt and generate photorealistic AI imagery layered directly onto the platform's authentic satellite, aerial, and 3D data. The tool was powered by Nano Banana 2, the image-generation model in Google's Gemini family, and product manager Bryan Horowitz pitched it as a way to "generate custom images using Google Earth's satellite, aerial, and 3D imagery" — concepts "grounded in the real world"2. Google's promotional framing leaned toward visualization: sketch a garden on an empty lot, or turn the ruins of Pompeii into a hyper-realistic view of the town as it looked in 78 CE12.

It took the internet roughly one day to turn that pitch into a falsified planet. By July 31, Google announced it was pulling the "create image" capability while it worked on "stronger guardrails"13. The rollback was not a gradual deprecation — Ars Technica checked the interface shortly after the announcement and found the feature simply gone2.

What Users Actually Made

The most alarming demonstrations came from open-source intelligence researchers, the very community that has long relied on Google Earth as an evidentiary baseline. Dutch investigator Henk van Ess reported that with single sentences he placed refugee crowds near the US-Mexico border, planted a nuclear facility in Iran, staged a fatal crash on an Amsterdam street, and rendered a hospital with a bomb crater in Gaza — all of it welded to genuine satellite coordinates underneath29. His verdict: "What on earth is Google doing?"2

Other testers pushed the same envelope. Futurism's own experiments produced a Pennsylvania town stricken by drought and wildfires, a nuclear meltdown turning a river into toxic sludge, and a White House "warzone" prompt that the model happily answered9. NPR generated flooding in Washington, D.C. and fires at an Iranian oil terminal, events that never happened86. Business Insider noted users created a nonexistent flood and a 9/11-style attack scene11. India Today reported images circulating on social media included a plane crashing into the World Trade Center12. Bellingcat founder Eliot Higgins responded to the launch with a sarcastic Bluesky post and a fabricated image of a giant golden Trump statue looming over the White House2.

The pattern across every outlet is consistent: nothing in the safety layer refused these prompts at the point of generation, which is precisely the problem Google's rollback statement gestures at without quite admitting.

Why Watermarks Didn't Save It

Google's first line of defense, offered publicly before the retreat, was SynthID — an invisible watermark embedded in every image its models produce. The company's News From Google account posted that users unsure about an image could ask Gemini or use Lens in Search to check for the mark, and that Google "prevents image creation on harmful topics"29. Both claims frayed under testing. Van Ess's Gaza hospital image sailed through the "harmful topics" filter entirely2. Tal Hagin of the OSINT platform Golden Owl fed that same fabricated image back into Gemini's SynthID checker and got no definitive answer — "no reliable signals were detected," per his screenshot9. Ars Technica found the same in its own testing, and when it photographed an AI-altered image with a phone camera, SynthID verification failed outright2. India Today and other outlets confirmed the same failure mode: images made in Google Earth were not reliably flagged as AI-generated12.

There is also a structural point van Ess made that watermarks cannot answer. "Fakes do not travel as clean files with their credentials intact," he wrote. They travel as screen recordings, re-encodes, screenshots of screenshots, filmed off somebody's phone in a hurry2. A watermark that survives compression is worthless once the image is stripped of any interface a checker can point at. Rest of World's retrospective on the episode drew the same conclusion from a different angle: the fabricated frames from this year's Iran conflict are still being screenshotted into new claims, while the authentic record now carries an asterisk it never earned16.

The Multimodal Trap: Grounding as a Liability

The deeper issue is architectural, and it is one the industry's multimodal push is racing straight into. Google's own marketing was the indictment: "grounded in the real world." Van Ess unpacked exactly what that means in a forgery context. The invented object is welded to genuine coordinates, drawn on genuine imagery, in the same colors, light, and angle as the picture beside it. "The forgery does not have to look convincing on its own. It inherits the credibility of the map it was born on"9.

This is the inversion that makes the Google Earth case categorically different from a standalone image generator. The Conversation's technical analysis made the same point: a from-scratch AI satellite image of a flooded city contains telltale misshapen buildings and wonky roads that a verifier can catch by cross-checking against real Earth imagery. But when the generation happens inside the trusted source — editing genuine satellite pixels rather than replacing them — the surrounding context is perfectly aligned, and the usual forensic anomalies disappear18. BBC Verify demonstrated this by collapsing the Eiffel Tower via the tool and finding almost nothing to flag18.

In other words, the multimodal grounding that Google sells as a feature for AI assistants and image tools becomes a disinformation multiplier the moment it is attached to an evidentiary instrument. Google Earth was not a canvas. It was the fact-checker's reference library.

A Familiar Cycle, and a Real Cost

This is not Google's first rapid AI reversal. In February 2024 the company paused Gemini's ability to generate images of people after the model produced historically inaccurate depictions814. The pattern — ship broadly, watch users find the obvious stress cases within hours, pull back, promise better testing — is now well-established, and observers noted the Google Earth launch cleared internal review in that form at all14.

The cost lands on the OSINT and journalism communities. Rest of World reported that Google Earth has been the platform on which open-source investigators are formally trained for conflict work in Syria, Yemen, and Sudan, and that since the February strikes on Iran, commercial satellite imagery has been the closest thing to ground truth for the region16. Van Ess argued the episode both undermines that trust and hands a ready-made excuse to anyone who wants to dismiss a genuine satellite photo as AI-generated — the "liar's dividend" that Index on Censorship documented concretely during September's Nepal floods, where even authentic CCTV footage was being assumed fake16. Evan Hill of the Washington Post's visual forensics team questioned whether Google had done adequate internal red-teaming, calling the abuse opportunities "almost limitless"6.

Google has not said when or in what form the feature might return, only that stronger guardrails are coming8. Based on what the coverage shows, labels and watermarks alone won't be enough. The credible fix has to assume what the last week of testing proved: that people will immediately test any such tool against disasters, wars, and borders — and that a feature which failed that test in public should not pass it in private the second time.

Where sources diverge: Google maintains its safeguards were meaningful (SynthID, harmful-topic blocking, no persistence in the shared Earth layer13); researchers uniformly report those safeguards failed at every point of testing, from prompt filtering to detection2912. Google's framing that this was a creative visualization tool with a few bad actors; nearly every independent account treats it as an inevitable outcome of putting a generator inside the world's most trusted geospatial reference.