Google Earth Added an AI Image Tool. It Lasted One Day

Google gave the entire planet the ability to fabricate photorealistic “satellite footage” of real places, watched people plant a nuclear facility in Iran and a crater in a Gaza hospital within hours, and pulled the feature the next day. If you missed the story, I covered the product-management side of it in what Google’s Earth AI failure teaches solo builders. That post is about why Google’s launch failed. This one is about what the failure changed for you — because the tool is gone, but the capability isn’t.

The underlying image model, Nano Banana 2, is still free inside Gemini’s image generation, and it’s getting better every month. The only thing that died on July 31 was Google’s willingness to attach it to a map of the real world. So the question stopped being “should this tool exist” and became something more useful: when the next AI image lands in your feed wearing a satellite’s clothing, how do you tell?

What Google actually shipped — in 60 seconds

On July 30, Google Earth web got a “create image” button. Zoom to any spot on the planet, type what you want to see, and Nano Banana 2 would render it grounded in real satellite, aerial, and 3D data. The pitched use cases were genuinely good: teachers showing students what Pompeii looked like in 78 A.D., architects visualizing an empty lot as a retail district, homeowners planning a lakefront cabin.

By July 31, Google’s own update used the phrase that matters: “people uniquely trust Google Earth for a reliable view of the world.” Users had been sharing screenshots of generated scenes — a nuclear plant in Iran, destruction in Gaza, a refugee camp near the Mexican border — all pinned to authentic coordinates, all exportable from the app looking exactly like the satellite imagery newsrooms use to verify events in war zones. Google suspended the feature while it builds stronger guardrails. As of this week it still hasn’t returned.

Why satellite imagery was the worst possible canvas

Here’s the part I keep coming back to. Satellite images aren’t decoration — they’re one of the last widely-trusted verification tools on the internet. When a bombing is reported and journalists can’t reach the site, the first check is often: what does the imagery show? Google attached an imagination engine to the one canvas where imagination is treated as evidence.

A watermark didn’t save it, because the harm never traveled through Google’s product. It traveled through screenshots — a channel Google doesn’t control. That’s the same pattern behind Amazon’s AI-generated product images: the image escapes its original context, and context is where all the safety features live.

How to tell if an image is AI-generated: the 5-step check

This is the practical part, and it works on any image, not just fake satellite shots. It takes about 90 seconds.

1. Reverse image search it. Google Lens, TinEye, whatever you have handy. If an image of a “just-happened” event has been circulating for months or appeared in an unrelated context, you’ll know in one tap. This catches recycled images — still the most common kind of visual misinformation.

2. Interrogate the physics. AI models are better than they were, but they still fumble things like shadows that disagree with each other, text on signs that almost spells words, and reflections that don’t match the scene. Zoom in before you judge. Most AI tells survive at full resolution.

3. Check the source and the timestamp. Who posted it, when, and do they have access to the place? An account you’ve never seen posting “breaking” aerial footage of a conflict zone deserves suspicion before you even look at the pixels.

4. Look for provenance credentials. Content credentials (C2PA) are slowly appearing on images from real cameras and legitimate newsrooms, and Google watermarks its generated images with SynthID. A credential is a good sign; its absence is neutral — but a stripped credential on supposedly professional press imagery is a red flag.

5. Ask what the image would need to be true. This is the habit that scales. A satellite image showing a new building means someone built a building, with construction crews, permits, and materials, without a single other source noticing. If the image requires a silent conspiracy to be real, the image isn’t real. This is the same skepticism I’d apply to any single-source claim — including AI health content, where the stakes are your body instead of your feed (how to use AI health advice without getting burned).

No single step is proof. Together they catch most fakes, because most fakes are lazy.

What this means if you use AI images in your work

Maybe you generate images for a blog, a listing, a pitch deck. Nothing here says stop — I use AI image tools constantly and rank them honestly in which AI image tool actually works. The Google Earth failure isn’t about the model’s capability. It’s about presentation context.

The rule I’ve landed on: use AI images where imagination is expected, and label them where they could be mistaken for records. A fantasy rendering of your city in 100 years is marketing. The same image posted without a label as “aerial view of X” is fabrication. The distance between those two is one checkbox and your credibility.

That’s also why the tool Google actually kept — image generation inside Gemini, free for everyone — survives happily alongside the one they killed. Nobody mistakes a Gemini chat window for a window into the real world. The Earth interface was a window into the real world. Same model, same watermark, completely different contract with the viewer.

The bottom line

Google Earth’s AI tool lasting one day isn’t a story about AI being too dangerous. It’s a story about context being load-bearing. The tool is paused; the capability is everywhere; the verification burden just moved to you. Run the five-step check before you share, and you’re already ahead of the people who will fall for the next one.

Want more beginner-first AI walkthroughs like this one? Start at /start-here/.