Google quietly discontinues its Earth AI feature a day after its rollout after users made no-no images

Google had a rough 24 hours. The company baked its Nano Banana image-generation tech into Google Earth, letting anyone overlay AI-made scenes onto real satellite and 3D imagery of actual places. Within hours, users weren't mocking up urban planning concepts or historical recreations, the uses Google had in mind. They were faking images of a nuclear facility in Iran and manufacturing scenes of refugees at the US-Mexico border. Google yanked the feature before the first full day was out.

The company's defense leans hard on its safeguards: generated images stayed separate from the core Earth experience, carried visible watermarks, and were tagged with SynthID, Google's cryptographic system for flagging synthetic content. That's also the mechanism Google uses to satisfy the EU AI Act's transparency rules for AI media. But NYU's Emily Black, who studies how AI outputs shape user trust, points out the obvious hole: SynthID is invisible to the human eye. A screenshot or a re-upload strips the metadata entirely, and the person looking at a fabricated nuclear plant three shares later has no idea it started as a Gemini prompt.

Black's broader research on AI hallucinations suggests the problem runs deeper than watermarking anyway. Even with blunt disclaimers, like "AI can make mistakes, check important information," test subjects who saw incorrect AI-generated advice still reported being more likely to act on it. Slap a label on something and people often trust it anyway once it looks convincing enough. That's a rough finding for a company betting its content-labeling strategy will keep synthetic disaster imagery from spreading as fact.

What makes this different from, say, a weird Midjourney image going viral is the twenty years of credibility Google Earth has built as a literal record of the planet. Nobody opens Earth expecting fiction. Henk Van Ess, an online-research expert, framed the real damage well: once people know a tool like this exists, an official can point at a genuine photo of a genuine atrocity and simply claim it's AI-generated. The tool doesn't even need to be used against a specific image. Its existence alone gives bad actors a plausible-deniability card for real events.

Google says it's rebuilding the guardrails and hasn't said when, or if, the feature comes back. Given how fast this went sideways, and how thin the current defenses turned out to be, a quiet return sometime later seems more likely than a quick fix.