Putting generative imagery inside a trusted mapping product created useful visualization tools—and an immediate credibility problem.
Google introduced an AI image generator inside Google Earth on July 30. By the following day, the company had removed it.
The feature allowed users to select a location in the web version of Google Earth and enter a text prompt. Google’s Nano Banana 2 model would then create a custom image grounded in the product’s satellite, aerial and three-dimensional imagery.
In its original Google Earth announcement, Google suggested using the tool to reconstruct ancient Pompeii, visualize a real-estate development, place a proposed house on an empty lot or imagine how a familiar location might look in the future. The company launched it globally to Google Earth web users.
Those examples positioned the feature as a planning and visualization tool. Users quickly demonstrated that it could also generate realistic-looking scenes of events that had never occurred.
Google rolled back the capability after people shared screenshots that appeared to violate its policies. The company said it would work on stronger safeguards before making the feature available again. Reuters reported on the rapid rollback.
The episode lasted only a day, but it illustrates a larger problem for businesses introducing generative AI into products that customers use as sources of truth.
The Tool Did Not Corrupt Google Earth
An important distinction is that the generated images were not added to Google Earth’s public geographic record.
Other users would not encounter a fabricated scene while browsing the product. Google also said the images were watermarked as AI-generated.
That makes the incident different from secretly replacing authentic satellite imagery with synthetic material. The feature operated as a private creation tool layered onto Google Earth’s real-world data.
But the distinction could disappear once an image left the product.
A user could capture a screenshot, crop it, repost it or place it inside another document. The resulting image might retain the appearance and authority of Google Earth even when the surrounding interface, warning or watermark was no longer obvious.
The Verge reported that researcher Henk van Ess used the feature to produce fabricated scenes involving recognizable locations, including refugees near the U.S.-Mexico border and damage near a hospital in Gaza. He also demonstrated that an altered Google Earth video could fool one AI-detection service.
The problem was therefore not limited to what Google stored. It involved what the product made easy to create and export.
Trusted Products Lend Their Authority to Their Outputs
Google Earth is not merely an image editor. People use it to inspect places, understand terrain and obtain a visual representation of the physical world.
That reputation changes how viewers interpret content produced inside it.
A clearly fictional image created in a general-purpose art tool begins with a presumption of invention. A realistic image carrying the visual conventions of a mapping or satellite product may begin with a presumption of evidence.
This is the central business risk. A generative feature can inherit the credibility of the system containing it, even when the generated output is technically separate from the system’s official records.
The same issue can arise when businesses place creation tools inside document repositories, property databases, inspection platforms, financial systems or medical applications. The generated item may look like the records around it without having the same origin, review process or evidentiary value.
An AI-produced summary may resemble an approved policy. A simulated property image may resemble a current site photograph. A generated incident report may resemble a verified account of what occurred.
The interface alone can make those differences difficult to see.
Watermarks Are Necessary but Incomplete
Google said the Earth images included its SynthID digital watermark and could be checked with Google tools. That is a meaningful safeguard, particularly when the original file remains intact.
Watermarking does not, however, eliminate the need for product-level controls.
A recipient must know that verification is necessary, have access to a compatible verification tool and receive a version of the content in which the identifying information has survived. Screenshots, video recordings, compression and republication can complicate that chain.
A label also addresses only one part of the risk. It may tell a viewer that AI was involved without explaining which elements were generated, what real information served as the foundation or whether a qualified person reviewed the result.
AskMaisy’s analysis of the EU AI Act’s transparency requirements explains why both visible disclosure and machine-readable provenance are becoming business responsibilities rather than optional product features.
For high-trust business systems, provenance needs to travel with the content. The record should show what generated the item, when it was created, which source material was used and whether someone approved it for operational use.
Businesses Need a Boundary Between Simulation and Record
The lesson is not that organizations should avoid generative visualization.
A construction company could use synthetic images to help a client imagine a proposed renovation. A city planner could compare development concepts. A disaster-response team could create hypothetical scenarios for training. Those uses can be valuable precisely because they show something that does not yet exist.
The danger begins when the product does not maintain a strong boundary between what might be and what has been verified.
Businesses adding AI creation features to trusted systems should use visibly different workspaces for simulations and official records. Generated material should carry persistent labels and provenance data. Exported files should preserve those disclosures, and systems should log who created and approved them.
Organizations should also decide where generation should be prohibited. An insurer might allow AI to illustrate a hypothetical repair but not alter photographs submitted as evidence for a claim. A property firm might generate design concepts while preventing synthetic images from entering an inspection record. A compliance team might use AI to draft a report but require verified attachments and human approval before filing it.
Similar governance questions arise when employees introduce unapproved tools into ordinary workflows. AskMaisy’s article on shadow AI in small businesses examines why organizations need clear boundaries around approved tools, information and outputs.
These controls are not merely technical. They define which information the business is willing to stand behind.
Google’s rapid reversal showed that a product can be functioning as designed and still create an unacceptable trust problem. The generated images were separate, watermarked and potentially useful, yet the connection to a respected view of the real world made misuse more consequential.
Generative AI can help people imagine alternatives. Systems of record help them establish what is true.
Businesses that combine the two must make certain their customers can always tell the difference.
