Google launched an image generation tool inside Google Earth on Thursday, called Nano Banana 2: it allowed users to alter satellite photos with a text prompt. Researcher Henk van Ess demonstrated that a single phrase was enough to create credible images of refugees at the Mexico border and a bomb crater next to a hospital in Gaza. By Friday, Google had withdrawn the feature.

The problem lay in the tool’s intended use case, applied to a domain where trust matters more than creativity. Google Earth is the instrument that journalists, NGOs, and OSINT analysts use to verify what’s happening on the ground. If anyone can overwrite real satellite imagery with invented elements, the cost of verification rises for everyone.

Google had embedded the SynthID watermark on every image and claimed to block harmful topics. It didn’t work: van Ess generated conflict scenes and migrant imagery without a single prompt being rejected. A video produced with the tool even fooled Hive, an external AI content detector.

The practice that remains solid is cross-verification. Van Ess recommends comparing images against independent platforms like Sentinel-2 or Landsat and checking orbital metadata. It’s the same logic as the lesson on how much to trust AI output: you can’t verify everything, but you can know where to look before using a result.

In detail

Nano Banana 2 was the image generation model integrated into the web version of Google Earth. The launch had been presented on Google’s blog as a tool for visualizing historical sites and real estate projects: the idea was to let users imagine how certain places might look with new construction or restoration work. The official demo showed Pompeii reconstructed.

Henk van Ess, who runs the Digital Digging blog, tested the tool with different intentions. He asked the model to show refugees near the Mexico-US border, and a bomb crater next to a hospital in Gaza. The resulting images were credible enough to circulate as false documentation in journalistic or political activism contexts.

Google’s initial response focused on the watermark. Every image generated with Nano Banana 2 carries a SynthID marker, the invisible watermarking technology developed by Google DeepMind. SynthID modifies pixels in ways imperceptible to the human eye but detectable by Google’s tools. The problem is that SynthID requires someone to look for it: if someone pastes the image in an article or shares it on social media, the average reader doesn’t know they should verify it. Van Ess made the point by creating a video from generated images and submitting it to Hive, a commercial AI content detector. Hive didn’t recognize it as synthetic.

Google had also claimed to prevent image creation on “harmful topics.” Van Ess reported the opposite: no rejected prompts, no modifications, no suggestions to rephrase the request. The policies existed, but the filters weren’t enforcing them.

After van Ess’s posts and journalist inquiries, Google issued a new statement. The key phrase: “We know that people rely on Google Earth for a reliable view of the world.” They acknowledged that although generated images didn’t appear in Google Earth’s main experience (they stayed in the user’s session), screenshots circulated out of context. The rollback came as they worked on “stronger guardrails.”

The point for those augmenting with AI lies at the boundary between generation and monitoring. Generation tools are becoming good enough to produce images that blend with real data. When the base is real satellite imagery and AI adds plausible elements, falsification becomes hard to detect by eye. SynthID and similar watermarks are useful for traceability, but they require that whoever receives the image knows to check it and has the tool to do so.

Cross-verification remains the most solid practice. Van Ess suggests three checks: compare the image against independent satellite platforms (ESA’s Sentinel-2, NASA’s Landsat), verify orbital metadata (which satellite captured it, when), and seek confirmation from third-party sources on the ground. It’s a protocol OSINT practitioners already know, but one that now needs to extend to images apparently coming from Google Earth.

What remains open: Google hasn’t said when or if the tool will return, nor what guardrails it intends to implement. This story is also a textbook case of predictable harmful use. A tool generating images on top of real satellite imagery, accessible to anyone on the web, was bound to be used for disinformation. The 24-hour withdrawal shows Google understood that, but the fact it made it to launch suggests risk assessment was insufficient.