Did Google pull Nano Banana too late to stop AI misinformation?

It took less than 24 hours for Google's newest AI experiment to turn into a case study on how fast image-based misinformation can spread.

The technology was integrated into Google Earth by Google on Thursday, using its Nano Banana 2 model. People had an opportunity to enter the instruction of their choice and receive a photorealistic image superimposed upon actual satellite, aerial, and 3D maps, with potential applications including imagining one's dream house or what a particular area will look like decades down the road.

Within a couple of hours, people started creating much scarier pictures, including artificial catastrophes, staged terrorist attacks, and imaginary sinkholes, as well as images involving refugees or protest crowds located in places where that has never taken place.

A researcher created a blast hole in LA, while another person used the tool to show protesters marching on Google's own California headquarters. The use of actual Google Earth maps meant that those pictures were very convincing.

Google initially defended the tool, pointing to SynthID watermarking and noting that Gemini or Lens in Search could confirm whether an image was AI-generated.

Critics weren't convinced, arguing most people encountering a screenshot online never bother to check and that the fabricated images never needed to appear inside Google Earth itself to do damage; screenshots circulating on social media were enough.

By Friday, Google confirmed it was rolling back the feature while building stronger guardrails, adding that generated images had been watermarked and hadn't shown up in the shared Earth experience.