Google has walked back an AI feature that allowed users to generate artificial images inside Google Earth, after a predictable flurry of deepfakes.
Google switched on the AI image generation feature inside Google Earth’s web version on July 30. It was available to everyone.
The system used Google’s Nano Banana 2 image generator to create its images. That tool can already generate images from simple text input, but the advantage of doing it in Google Earth is that it can use the real satellite images as the basis for its deepfake versions. That makes it easier to make AI pictures with real, accurate building and landscape details.
In its initial blog post on the launch, it said that students could use it to “bring history to life”, while realtors could use it to produce professional real estate plans. However, others warned that the system could be used to mislead people.
Within hours, researchers and press outlets demonstrated that the tool would happily produce photorealistic satellite imagery of things that did not happen in places where they did not happen.
Dutch open source intelligence researcher Henk van Ess explained: “I tried 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”.
Demonstrating what was possible with the new capability, NPR fabricated an image of fires in Iran, along with a deepfake of Washington, D.C. underwater. The BBC ran images of a collapsed Eiffel Tower and the Great Pyramid of Giza swallowed by a sinkhole. Even though the service had some guardrails in place, the BBC’s anti-disinformation Verify service was able to circumvent them by tinkering with basic AI prompts.
Google acknowledged the failure in a statement on X:
“We’ve seen geospatial professionals using this feature for a range of useful purposes, 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.”
It didn’t commit to never re-introducing the idea.
Users were apparently unimpressed. “There is 0 chance that no one on your development team didn’t raise exactly this concern,” commented one. “You guys are living in a complete bubble,” accused another.
Google’s fallback safeguard was a SynthID watermark and the fact that generated images didn’t appear in the main Google Earth experience for others to see. SynthID is Google DeepMind’s watermarking system, an invisible signal baked into the pixels of AI-generated images so that a compatible detector can spot them later. Google positions it as one half of its provenance stack, sitting alongside the C2PA metadata standard the wider AI industry has settled on.
On paper, the signal is meant to hold up through compression and even social media re-uploads. However, these claims collapsed on contact with reality. The watermarks are detectable by Google’s AI services like Gemini. They are not visible to users, who can screenshot the images and share them anywhere. It’s unlikely that everyone will know to check for the provenance of an image. Researchers have also reported that Gemini could not reliably identify AI-generated images with a SynthID watermark.
What this means for you
Content creators were already producing fake AI images showing events that didn’t happen. Traditionally, satellite imagery has been a key component of open source journalism. Fake it convincingly and you are attacking the reference layer reporters use to check whether something actually happened.
This also comes at a time when trust in AI is measurably eroding. According to our own research, released in June, 88% of people said it’s becoming harder to tell what content online is genuinely human or real, with 84% saying that even “convincing video evidence” no longer feels like proof.
The practical advice is to take a breath whenever a disturbing satellite image of a disaster, a weapon strike, or a border crossing hits your timeline. Check whether a wire service with a named reporter has published it. Look for a caption identifying the imagery provider. If the source is an anonymous account posting a single dramatic frame, assume you might be looking at something assembled in a browser tab last night.
For more information on how to identify AI images, check out our guide.
The industry’s shipping model now apparently treats users as the test group. Critical media literacy is now the only reliable tool that readers and viewers have.
From reporting threats to removing them.
Cybersecurity risks should never spread beyond a headline. Keep threats off your devices by downloading Malwarebytes today.
