Histoire
août 10, 2026
Google Earth’s AI experiment lasted a day before fake-world fears won
Google promoted Nano Banana 2 as a creative way to reimagine places in Google Earth. Critics showed how quickly the same tool could manufacture plausible scenes, forcing a rapid rollback and a promise of tougher safeguards.
Google’s bid to make Earth more imaginative ran headlong into the reason people trust it: the map is supposed to show what is there, not what a prompt can invent.
On Thursday, Google introduced Nano Banana 2 inside Google Earth’s web experience, pitching the image-generation feature as a way to visualize a city’s past or imagine additions such as a neighborhood basketball court. Demis Hassabis amplified that launch message, which framed the technology as a creative extension of mapping rather than a replacement for reality.
1
Critics immediately saw the other use case. Because prompts could superimpose almost any generated scene over real geography, journalists and researchers warned that a platform long treated as visual evidence could become a pipeline for convincing geographic misinformation. One BBC journalist captured the alarm sarcastically: “There’s no way that this new AI image generation feature on Google Earth … could possibly be abused to spread misinformation online.”2
Testing made the concern concrete. Digital Digging’s Henk van Ess generated scenes involving refugees near the Mexican border and a bomb crater by a Gaza hospital; he said, “Nothing was refused, nothing was softened, and nothing suggested I try a different prompt.”3 Google initially stressed that images carried an AI watermark and that harmful requests were blocked. But van Ess said generated material could still evade an AI-detection tool, weakening the practical value of those labels.3
Within a day, Google withdrew the feature. The company said professionals had found useful applications, but screenshots of generated imagery “appear to violate our policies,” adding: “We’re rolling back this feature in Google Earth while we work on implementing stronger guardrails.”2 It also maintained the images were not inserted into the main Earth experience for other users and were watermarked.3
A separate account described a fabricated Pope Francis image spreading online and characterized the pause as temporary, with verification measures planned before any return.4 The broader lesson is sharper: safeguards may label synthetic imagery, but they do not automatically preserve public trust once a trusted map can be made to show a fiction.