Google suspended, less than 48 hours after its launch, the integration of its AI image generator Nano Banana 2 into Google Earth. The tool allowed any user to produce entirely fictitious scenes overlaid on real satellite, aerial, or 3D backgrounds with a simple text instruction — and then download and share them elsewhere. A swift retreat, but it raises a fundamental question: what is the value of trust in a map when AI can append anything to it?
A launch on Thursday, a suspension on Friday
The feature had been announced on a Thursday. By the following Friday, Google stated it was pausing it “while we work on implementing stronger safeguards.” In its statement, the company acknowledged having found that “people were sharing generated images that appear to violate our rules.”
The gesture is remarkable for its speed, but also for the implicit admission it contains. Google itself noted: “We know people trust Google Earth particularly for reliable views of the world.” It is precisely this credibility — historically acquired and largely deserved — that forms the crux of the problem raised by experts.
Tests that send shivers down the spine
The investigation by BBC Verify, published on July 31, 2026, highlighted the extent of the risk. Journalists managed to generate several convincing fictitious scenes:
- A collapsed Eiffel Tower;
- A chasm swallowing the Great Pyramid of Giza;
- Russian tanks rolling through Kyiv.
Even before the suspension, AI and disinformation specialist Henk van Ess had tested the limits of the system on his end. He produced:
- A nonexistent nuclear power plant in Iran;
- A refugee camp on the US-Mexico border;
- A fake hospital in Gaza accompanied by a bomb crater.
His analysis succinctly summarizes the danger: “Falsification doesn’t need to be convincing on its own. It inherits the credibility of the map on which it was born.” In other words, a mediocre image anchored in a recognized cartographic environment can mislead much more effectively than an isolated deepfake.
Workarounds to the safeguards
Google claims that contents generated by its AI tools contain invisible watermarks intended to signal their artificial nature. The company also states that someone with doubts can verify an image via the Gemini app or the Lens function of its search engine.
BBC Verify's tests showed that these mechanisms generally worked… but not always. Journalists managed to bypass these checks and get Gemini to present false Google Earth images as authentic. External AI-generated content detection tools also failed, in some cases, to recognize the productions created from Google Earth.
The rules embedded in the model that are supposed to prevent the production of images on “harmful subjects” also proved porous. BBC Verify found that slight modifications in the wording of instructions were enough to breach protections:
- A request to create a “raised platform to hang traitors” near the British Parliament was denied. A less precise formulation was then accepted, and the image was generated.
- The system refused to write the word “Trump” on the White House lawn, invoking its rules. However, it accepted to install a community garden at the same location.
This pattern — refusal of explicit requests, acceptance of twisted formulations — is now well known among AI security researchers, but its application in a cartographic context amplifies it significantly.
Increased risk in conflict zones
AI detection researcher Henry Ajder, interviewed by BBC Verify, warned that creating false images in conflict zones could be particularly “destabilizing” when events are rapidly evolving and reliable information is scarce.
“Journalists may be able to verify this image… but not everyone will,” he emphasized.
For Ajder, the continued proliferation of artificial content increases the value of areas still perceived as “free from AI”, especially for journalists and people working in the field during rapidly evolving crises. The issue goes beyond simple falsification: it's the perpetual doubt that sets in. Every geolocated image becomes suspect, undermining the entire information chain.
Satellite imagery, a historically trusted source
Bill Greer, geospatial analyst and co-founder of the nonprofit satellite service Common Space, recalled that satellite imagery has long been considered a particularly trustworthy data source. For many, Google Earth is the entry point to freely accessible satellite imagery, used to gain insights into hard-to-reach areas.
“This is potentially particularly damaging to public trust, because satellite imagery has historically been seen as a particularly reliable data source,” he stated.
Greer adds a concerning dimension: governments are increasingly restricting access to satellite data, which could lead to “more false images and fewer real images,” further reinforcing the erosion of trust. Falsified satellite images have already been used to spread misinformation in the past.
Implications for cyber intelligence and OSINT
Beyond the media observation, this episode raises concrete issues for professionals in cyber intelligence, open-source intelligence (OSINT), and verification. Several lessons emerge:
- A geolocated image is no longer enough. Provenance, verification chain, and the resilience of detection systems become essential elements of any analysis.
- Watermarks are not 100% reliable. Bypassing Google’s checks — including by Gemini itself — shows that current signaling mechanisms remain insufficient in the face of images “fused” with real coordinates.
- Uncertainty zones are the most exposed. Conflicts, natural disasters, and political crises — where real images are scarce and information demand is high — are the most conducive grounds for exploiting this type of tool.
- Public data restrictions worsen the problem. Less access to authentic satellite images means more room for fakes, and an asymmetry that benefits malicious actors.
A precedent, not an end
Google has not provided a specific timeline for a potential reintroduction of the feature. The company says it is working on "stronger safeguards," but the episode illustrates a fundamental trend in generative AI: the difficulty of aligning creative models with robust safety rules, especially when the output medium has existing credibility.
The case also serves as a reminder that OpenAI offers a tool for verifying images and sounds generated by its models — allowing users to upload a file to detect signals indicating an AI origin. However, no universal, multi-platform detection system exists to date.
In the meantime, one thing is certain: the map is no longer the territory. It can now be a fiction.