Twenty-four hours
Google had moved its Nano Banana 2 image generator inside the Google Earth map interface: a user writes a prompt and the generated image lands on top of the real map. The feature opened on Thursday and was removed on Friday. In its statement the company said it had seen geospatial professionals using the tool for a range of useful purposes, but had also seen people sharing screenshots of generated imagery that appeared to violate its policies, and that it was rolling the feature back while working on stronger guardrails.[1]
The job Google Earth does is narrow and clear: look at what a place actually looks like. No setup, no account, nothing to learn; the friction budget of that job is close to zero. When an editable layer is placed on top of a task that easy, what changes is the assumption about what the result corresponds to, while the tool's capability stays where it was.[1]
The two halves of the company's own sentence
The statement says two things at once. Geospatial specialists used the feature usefully; at the same time, screenshots of generated imagery went outside policy. The rollback is not proof that the tool had no legitimate use. When the useful and the harmful version of a product sit behind the same button, the problem lies in where that capability is opened rather than in the capability itself.[1]
The break happens not in the generation but after the image leaves the app. Inside the app the user knows what they did, because they wrote the prompt. The moment a screenshot is taken that context drops away: what remains is an image that looks like Google Earth, and whoever sees it has nothing to tell it apart. That was the basis of the criticism, since Google Earth counts as one of the most reliable sources of visual evidence for journalists and researchers.[1]
What to look at if it comes back
Stronger guardrails is an intention for now, not a specification. The version a user can actually check comes down to three questions: is the generated layer visually distinguishable inside the app, does generation move to a surface separate from the map itself, and does an image that leaves the app carry on its face what it is. The third matters more than the first two, because that is where the problem appeared.[1]
The signal to watch is therefore narrow. If the feature returns by 31 December 2026 and a generated image still carries no visible marking that survives a screenshot, this week's failure returns in the same form. If it does carry one, the rollback has turned into a product decision.[1]