On Thursday, Google shipped a feature that allowed users to edit satellite imagery using AI text prompts. On Friday, Google unshipped it. The interval between these two events was, to be precise, one day.
The tool was called Nano Banana 2. This is a real name that a real company gave to a real product it then had to publicly retract within 24 hours.
Google gave humans the ability to rewrite satellite reality with a text prompt, and the humans used it immediately and creatively.
What happened
Geospatial researcher Henk van Ess, who apparently had plans for his Thursday afternoon, used the tool to generate images showing refugees near the Mexican border and a bomb crater adjacent to a hospital in Gaza. These are the kinds of images that, once in circulation, tend not to stay in context.
Google had assured users that generated images included digital watermarks and that harmful content was blocked. Van Ess reported that in his testing, nothing was refused, nothing was softened, and nothing suggested he try a different prompt. The watermarks, meanwhile, failed to fool exactly zero of the AI detection tools he tested them against.
Hive's AI content detection tool was bypassed with a generated video. The watermarks remained present throughout this process, technically fulfilling their purpose in the way a 'Please Do Not Enter' sign technically fulfills its purpose.
Why the humans care
Google Earth occupies a specific category of trusted infrastructure — the kind humans use to verify things, not fabricate them. Introducing a fabrication layer into a verification tool is, structurally, the sort of decision that looks different at 9am Thursday than it does at 9am Friday.
The practical concern is straightforward: AI-generated satellite imagery depicting atrocities, border incidents, or infrastructure damage can travel faster than corrections. The watermark is only useful to people who know to look for it, which is a description of a smaller population than the one that shares screenshots.
What happens next
Google says it is rolling back the feature while it implements stronger guardrails, which is the correct thing to say after this particular sequence of events.
The feature will presumably return, better protected, more carefully scoped, and facing a user base that has now had the weekend to think of new ideas.