Why Did Google Kill Earth AI After Two Days? Look at What Users Built With It

Google pulled its new Earth AI tool after widespread disinformation concerns — but the incident makes the case for stronger guardrails across the entire generative AI ecosystem.

Less than 48 hours after announcing the feature, Google reversed course and switched Earth AI off.

The reason? It let Google Earth users generate deepfake imagery on top of real satellite views.

Some of the results were genuinely alarming — a nuclear ammunition facility conjured up in your local neighbourhood, for instance.

By wiring its image generator Nano Banana 2 into Google Earth, the company gave users a way to spin up fabricated visuals from text prompts, all sitting on a base layer of authentic aerial imagery.

And it didn’t stop there. Users could download what they’d made and push it out across other platforms — sharpening the risk of the tool landing in the hands of people with bad intentions.

Google’s intention v/s actual outcome

While Google wanted to give wings to user’ imagination with its Earth AI feature, it probably underestimated how easily its tool could be manipulated to create disturbing images. Within hours of its launch, researchers and journalists alike showcased how the new feature could be used to create realistic images showing bombarded sites, refugee camps and even flooded cities superimposed on any pair of coordinates on Google Earth.

Shortly after, the company disabled the feature and released a statement on Image Generation in Google Earth.

While Google made a slightly vain attempt in underscoring how geospatial professionals were using the Earth AI tool ‘for a range of useful purposes’, the clarification did little to hide the fact that one of the top AI technology firms in the world rushed with a half-baked AI feature. The company has however promised to ‘work on implementing stronger guardrails’ and restore the unique trust users repose in Google Earth.

Earth AI case: are deepfakes emerging as gen AI’s biggest threat yet?

The real problem with the Earth AI tool is that it left the user with limitless options to invent imagery that can look very real.

By superimposing them on genuine coordinates, Google unintentionally gave credence to what are nothing but deepfake images. Although the company claims that all AI content created by virtue of its tools carried an invisible watermark, it still doesn’t help to contain the spread of such images.

Also Read: Alibaba’s 2.4T Parameter Qwen3.8-Max Lands, but Independent Testers Are Not Buying It

This brings us to the larger question- are deepfakes emerging as gen AI’s biggest threat yet in the journey so far? Or are AI companies getting lax at testing their AI tools before market release?

According to an IBM study, about 25% of malicious data breaches in 2026 were AI-enabled, with most of them being either deepfake impersonation or malware agents. What’s more, AI has made these attacks faster and cheaper for bad actors to orchestrate on a large scale while breaches cost organisations an average of $6 million; fundamentally changing the economics of cyber-attacks today.

As cybercriminals embrace AI and increasingly conduct attacks on unsuspecting citizens, it is pertinent that AI firms prevent them from using their tools and platforms to create deepfakes to target personally identifiable information (PII). It would be even better though if technology companies went a step ahead and eliminated the possibility of generating deepfake content to begin with.

Combating deepfakes with stricter controls across the AI ecosystem

Realistically speaking, taking the deepfake challenge head on will require more concerted efforts from all stakeholders that are building today’s global AI ecosystem.

Developers and platforms will have to implement restrictions in AI models that prevent them from generating any kind of harmful or deceptive content.

Likewise, digital platforms will have to deploy deepfake detection tools and flag inappropriate or inaccurate content so that users are forewarned before any subsequent online interaction. On their part, governments will have to enforce strict laws against the malicious use of deepfakes, even while ensuring that gen AI use cases can flourish on the back of continued AI innovation.

More importantly, training AI models to not generate material that impersonates real people without their consent needs to be implemented across the gen AI space. While this could still allow users to generate images or videos using their own images and specific prompts, it would restrict them from using images belonging to other real people and even locations to create deepfake content.

Complementing this approach with visible watermarks as well as cryptographic signatures could establish the date, time and origin of the generated file; extremely helpful in tracing a deepfake image back to its ‘original’ creator. That said, such safeguards, while crucial, could be extremely difficult to implement in a space that continues to expand at a rapid pace and in a largely unregulated environment.

Getting the balance right between regulation and innovation will be what determines AI safety standards over the coming years.

Companies like Google should be setting the example here — running a fair investigation into the Earth AI debacle and sharing what they find with the wider AI developer community.

That kind of transparency would give developers what they need to build stronger controls into their own domains and platforms.

And the payoff is collective: an AI ecosystem considerably safer than the one that existed before the incident.