Google’s AI Satellite Feature Lasted One Day Before Misinformation Fears Forced a Retreat

Google moved fast. On a Thursday in late July it rolled out an experimental feature inside Google Earth. Users on desktop could zoom to any spot on the planet, tap a button and type a prompt. The company’s Nano Banana 2 model would then generate an image blending real satellite, aerial and 3D data with whatever the user imagined.

Dream home in Central Park. Nuclear facility in Iran. Fires raging across a strategic oil terminal. The tool delivered. And then, less than 24 hours later, it vanished.

The reversal came after an immediate and vocal backlash. Researchers, journalists and open-source investigators tested the feature and recoiled. They saw the obvious. Anyone could create convincing satellite-style imagery of events that never happened. Screenshots would spread. Trust would erode. Google pulled the plug.

“We’ve seen geospatial professionals using this feature for a range of useful purposes,” the company said in a statement posted on X. “However we’ve also seen people sharing screenshots of generated imagery that appear to violate our policies. So we’re rolling back this feature in Google Earth while we work on implementing stronger guardrails.” The post came from the official News from Google account on July 31, 2026. (NPR)

That statement revealed the tension. Professionals found value. The rest of the internet found opportunity. Henk van Ess, an open-source researcher, ran the tool through a gauntlet of sensitive scenarios. Refugees at the Mexican border. A nuclear plant in Iran. A crash in Amsterdam. A hospital with a bomb crater in Gaza. Nothing was refused, he reported. (CNET)

Van Ess had already published a newsletter post demonstrating the risk. He generated an image of a nuclear facility planted in Iran. The headline captured the damage: Google spent twenty years building the reference the world checks against. Today it added a button that makes things up. His work went viral. It forced the conversation into the open. (NPR)

NPR itself tested the feature before the rollback. Staff generated an image of multiple fires on and near Kharg Island, an Iranian island in the Persian Gulf home to a strategic oil terminal. Another showed significant flooding around the U.S. Capitol complex. Neither event occurred. Both would have qualified as major breaking news. The images looked plausible enough to fool casual viewers. Annotations added by NPR made the point unmistakable. (NPR)

Evan Hill, a visual forensics investigator at The Washington Post, captured the collective shock. He posted on X that he was curious how the idea survived internal review. “The opportunities for abuse and disinfo are literally boundless,” he wrote. The post included a screenshot of the new button sitting inside Google Earth. (NPR)

The speed of the backlash surprised few who follow these debates. Satellite imagery has long served as a rare constant in a sea of questionable online content. Since the Cold War it has offered governments, media outlets and researchers a view from above that felt objective. Hard to fake. Controlled by institutions or companies with expensive hardware in orbit. That assumption is cracking.

Jake Godin, a senior researcher with Bellingcat, the open-source investigative group, explained the shift. “Satellite imagery has been kind of a safe bet when it comes to verifying an event because it’s hard to fake,” he told NPR. Images come from cameras hundreds of miles above the surface. Their provenance has historically been clear. (NPR)

But the new tool changed the equation. It did not replace the base imagery for other users. Generated pictures stayed in the creator’s private view. They carried a SynthID watermark designed to flag them as artificial inside other Google products such as Gemini. Yet the ease of screenshotting and sharing bypassed those controls. Once an image leaves the platform, context disappears. Belief takes over.

Godin noted that fake satellite imagery had already appeared in limited cases. Early in the U.S.-Israel conflict with Iran, fabricated pictures circulated in Iranian media showing damage to a U.S. base in Bahrain. They appeared to combine Google Earth views with AI generation. Real satellite data later confirmed actual damage, but not in the exaggerated form claimed. The precedent existed. Google’s feature made creation trivial. (NPR)

Concerns stretch beyond immediate headlines. Brady Africk, an open-source intelligence analyst and author of the newsletter Views From Above, mapped the broader pattern in a October 2025 piece for Time. He described how AI-generated satellite fakes had already surfaced in multiple 2025 conflicts. Ukraine’s Operation Spiderweb saw exaggerated claims of destroyed Russian bombers accompanied by fabricated overhead images. After U.S. and Israeli strikes on Iranian nuclear sites, fake pictures depicted crowds around a downed Israeli F-35 and misleading missile footage. India and Pakistan both circulated doctored imagery during their brief clash to inflate reported damage. (Time)

Africk argued these images rarely fool well-resourced militaries with their own satellite fleets. Yet they shape public opinion with speed and scale. More than half the globe uses social media. A single convincing fake can travel faster than any correction. One hoax image falsely showing a fire near the Pentagon last year triggered a noticeable stock market dip until authorities stepped in to clarify. (Time)

The technology enabling these fakes has improved dramatically. Earlier models produced blurry, low-quality results. Current systems need nothing more than free software and a well-written text prompt. Quality has risen. Barriers have fallen. And the images now mimic the precise visual language that analysts and journalists have come to trust.

Google’s initial blog post had pitched the feature as creative and grounded. Product manager Bryan Horowitz wrote that users could simply zoom in, tap create image, and type whatever they wanted to see. The company highlighted uses such as reimagining infrastructure or exploring historical changes. It insisted generated images remained watermarked and isolated. Those assurances failed to calm the experts. (CNET)

CNET’s own test captured the seductive simplicity. A writer searched for Central Park, zoomed, hit the button and requested a dream home. Minutes later the image appeared. Three stories tall. Slightly unrealistic yet plausible at first glance. The feature allowed refinement and saving to projects. It felt playful until the implications sank in. (CNET)

Van Ess captured the deeper irony. Google Earth had become one of the primary references people use to check whether pictures on social media are true. Now the company offered a way to generate fakes from inside that same trusted platform. “The fake is made inside the thing people use to check whether pictures are true,” he said. The contradiction stung.

This episode fits a larger pattern. Technology companies race to ship new AI capabilities. They tout creative and professional applications. Then the adversarial tests arrive. Misinformation researchers, not product managers, often surface the worst cases first. Guardrails prove inadequate. Rollbacks follow. The cycle repeats.

Google emphasized it takes misinformation seriously. It pointed to ongoing updates to protections and the SynthID watermark. Yet the open-source community sounded unconvinced. The feature’s availability, however brief, demonstrated how quickly synthetic satellite imagery can be produced at scale. Proliferation seems inevitable even without Google’s direct help.

Existing detection tools lag behind generation capabilities. Researchers have published papers on identifying AI-generated satellite images by spotting terrain inconsistencies and structural artifacts. One arXiv paper from late 2025 examined models like StyleGAN2 and Stable Diffusion and warned of growing threats to scientific and security analysis. Commercial providers, media organizations and governments all face pressure to adapt.

Africk called for a society-wide response. News outlets that publish satellite imagery should explain their verification steps. Companies like Google could offer public tools to authenticate images claimed to come from their platforms. Governments might expand public education on disinformation, following examples set by Sweden and Finland in their crisis preparedness materials. The U.S. defense department’s guides, he noted, still fall short on this specific risk. (Time)

Meanwhile the incident leaves practical questions unanswered. How will future versions of this technology incorporate stronger controls without killing legitimate uses? Can watermarks survive screenshots and re-uploads? Will dedicated bad actors simply move to open-source models that lack any corporate oversight?

One thing is clear. The genie is out. Satellite imagery no longer carries the same presumption of truth it once did. The view from above has been democratized. And with that openness comes vulnerability. Google learned the lesson in a single day. The rest of the information environment will wrestle with the consequences for years.