Google Pulled an AI Tool That Let Anyone Fake Satellite Photos of Real Places
A Nano Banana 2-powered Google Earth feature let users generate fabricated disaster imagery over real coordinates before Alphabet rolled it back within days.
Declan Moss 7 min read
Alphabet rolled back an AI image-generation feature in Google Earth after it was used to produce imagery that violated the company's own policies, according to reporting by Reuters between 31 July and 1 August 2026. The tool, built on Alphabet's Nano Banana 2 generator, allowed users to create fabricated scenarios layered onto real satellite and aerial imagery of actual locations. Ars Technica documented one example: a fabricated image showing fire, smoke and a burn scar at the Googleplex, Alphabet's own headquarters, generated using the company's own tool.
What the tool actually let people do
Unlike a general-purpose image generator producing an entirely synthetic scene, this feature operated on top of genuine geospatial imagery of real, identifiable coordinates. A user could take an actual satellite view of a specific address and instruct the model to add a wildfire, flooding, structural damage or other events that never occurred there. The result inherits the visual credibility of real satellite photography while depicting something that did not happen.
Why geospatial imagery is a uniquely dangerous category
- Satellite and aerial imagery carries institutional trust that ordinary photographs do not, because the public associates it with scientific measurement rather than composed photography.
- Real, verifiable coordinates give a fabricated image a specificity that generic AI-generated scenes lack, making it far more usable in disinformation about a particular place or event.
- Emergency response, insurance claims and news verification all rely on satellite imagery as a comparatively trustworthy source, a reliance the tool directly undermined.
- Geolocation of a fake image to a real address is trivially checkable by the tool itself, unlike synthetic imagery of an invented location.
A convincing fake photo of an unspecified place is a curiosity. A convincing fake photo of a specific, real address is a weapon.
The speed of the walk-back
Alphabet withdrew the feature within days of the problematic outputs surfacing publicly, a fast turnaround by the standards of major platform policy responses. That speed suggests the company recognised the category of harm as unusually severe rather than treating it as an ordinary content-moderation edge case requiring lengthy review. It also suggests the risk was foreseeable enough that it should arguably have been caught before launch rather than after.
The broader pattern
This episode fits a recurring problem across generative AI products released in 2026: capability shipped ahead of a clear policy for the specific misuse that capability enables. Consumer-facing image generators have absorbed years of scrutiny over deepfakes of people. Geospatial generation over real coordinates raises an adjacent but distinct set of harms, around disaster misinformation, property fraud and geopolitical provocation, that the industry has had comparatively little practice policing.
What to watch
Watch whether Alphabet reintroduces the feature with restrictions such as watermarking, coordinate-blocking near sensitive sites, or a ban on disaster-related edits specifically, and watch whether competing mapping and satellite-imagery products with generative features face similar scrutiny before their own equivalents ship.
Was this helpful?
Declan Moss
Security Editor, Lonic
Declan spent a decade in security operations, including four years running incident response for a multinational bank, before writing about the field full time.
- Cybersecurity
- Incident response
- Threat intelligence
Read our editorial standards or send a correction.
Related articles
Anthropic Says AI Orchestrated a Real Cyberattack. Here Is What That Actually Means
The disclosure describes a model executing most of an intrusion campaign with limited human direction. Separating the genuinely new part from the familiar part matters.
Defending at Machine Speed: How Security Teams Have to Change
If attackers automate the middle of the kill chain, defenders who keep a human in every loop will simply be outrun. A practical framework for what to automate first.
Agentic Orchestration Explained: What It Is and Why Enterprises Are Buying It
Search interest is up sharply and vendors have moved fast. Underneath the terminology is a genuinely useful architecture — and a set of failure modes worth understanding first.
Antidetect Browsers: What They Are, Who Uses Them, and Why Search Interest Is Rising
A tool built for legitimate multi-account operations sits one step away from fraud infrastructure. A clear-eyed look at the technology and its consequences.