MOUNTAIN VIEW, 3 AUG 2026 — On 30 July, Google Earth shipped a feature that let users type a description and place generated imagery onto real satellite coordinates. Google withdrew it on 31 July, after news organisations showed within hours that it would produce convincing pictures of events that had not happened.
The tool was powered by Nano Banana 2, Google's image-generation model. A user could zoom to a location in the web version of Google Earth, choose to create an image, and describe what they wanted to appear there.
NPR reported generating flooding in Washington, DC and a fire at an Iranian oil terminal. Researchers quoted in the same coverage raised the prospect of fabricated nuclear facilities placed on real sites.
What Google said
The company's statement is worth reading closely, because it concedes the design problem while framing the withdrawal as a policy-enforcement matter.
"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," Google said. "We're rolling back this feature in Google Earth while we work on implementing stronger guardrails."
The operative word is screenshots. A guardrail inside the product governs what the product will generate. It does not govern what happens to an image once it has left the product as a picture file, and a screenshot strips whatever context the interface supplied.
Why satellite imagery is a different case
Generated faces have been around for years, and the public has learned some scepticism. Overhead imagery has not been through that, and it carries an assumption a portrait does not: that it is a measurement.
Satellite pictures are used as evidence precisely because they are understood to be captured rather than composed. They appear in reporting on conflicts, in monitoring of industrial sites, in disaster assessment and in litigation. The credibility comes from provenance — a sensor, on a platform, at a known time, over known coordinates.
| A generated portrait | A generated satellite view | |
|---|---|---|
| What the viewer assumes | Someone composed this | A sensor recorded this |
| Anchored to a real place | No | Yes — real coordinates supplied by the host map |
| Public scepticism | Substantial, after several years | Largely undeveloped |
| Used as evidence | Rarely | Routinely — reporting, monitoring, disaster response |
| Cost of a convincing fake | Low | Previously high; briefly, near zero |
RECATOOLS comparison of the two categories. The contrast is our framing of why the same generative capability carries different consequences depending on the genre of image it produces.
The significance is not that fabricated overhead imagery became possible — that was always possible with image editing and time. It is that the cost collapsed, and the fabrication arrived pre-registered to genuine coordinates inside the interface people use to check where something is.
The question nobody has answered
The detail that determines how much this mattered is whether generated images carried a provenance marker, and whether that marker survived being screenshotted. Google operates a watermarking scheme for AI imagery, but we could not confirm from sources we were able to read whether it was applied here, or whether it persisted through the screenshot path the company itself identified as the problem.
That gap is the entire question. If the marker survives, this is a labelling and detection problem. If it does not, pulling the feature was the only control available, and the next product with the same capability re-opens the same exposure.
A one-day product lifetime is the finding
The feature was reviewed, approved and shipped by a company with an established AI safety apparatus and a specific institutional memory of image-generation incidents. It survived a day in public.
The failure was not that the abuse was unforeseeable — journalists found it in hours, with no special access or expertise. That is precisely the failure mode internal review exists to surface. What the episode measures is the distance between a review process that asks whether a model produces prohibited content, and one that asks what a capability becomes when placed in a particular context — here, a map that people consult specifically to establish what is real.
A withdrawal within a day shows the system working at the far edge. It is also a great deal more expensive than not shipping.
Why this matters in the region
Southeast Asia is a natural setting for the harm this feature made cheap. The South China Sea disputes are argued substantially through overhead imagery of reclamation and construction, with the Philippines, Vietnam, Malaysia and Brunei all parties to overlapping claims. Fabricated imagery of an installation on a contested feature would not need to persuade a defence ministry, which has its own collection. It would need to persuade a news audience for a day.
The region also runs on disaster imagery. Indonesia and the Philippines both use satellite assessment in response to floods, eruptions and typhoons, where the pictures inform where resources go. We have reported previously on synthetic media reaching ASEAN banking fraud; the same capability aimed at emergency response has a shorter path to physical consequence.
The guidance has not changed and does not hinge on this product. Overhead imagery presented as evidence should carry its provenance — the operator, the acquisition time, the coordinates — and where it does not, its absence should be treated as a material gap rather than a formatting omission.
View author profile → · Editorial policy
