Google Pulls AI Image Tool From Earth After 1 Day

Google's experiment with generative AI inside one of the internet's most trusted mapping tools lasted barely 24 hours before it was pulled back.

On Thursday, Google launched Nano Banana (also referred to as Nano Banana 2), an AI image editing and generation feature built into the web version of Google Earth. The tool let users create prompt-based images superimposed over real maps, which Google pitched as a way to visualize historical sites or real estate projects. One day later, amid mounting criticism that the feature could be used to spread misinformation, Google rolled the feature back.

What went wrong

Google Earth has long been treated by journalists and researchers as a reliable source of visual evidence — a reputation that made the misuse risk especially sensitive. Henk van Ess, a researcher with Digital Digging, tested the tool by intentionally generating fabricated images and a video to demonstrate how easily it could be abused. According to van Ess, the tool did not refuse, soften, or redirect any of his prompts during testing — a sign of weak content moderation guardrails at launch.

More alarming: van Ess said he was able to use a video generated in Google Earth to fool Hive's AI detection system, one of the tools relied on to flag synthetic media. A BBC journalist publicly questioned how a feature attached to a source as trusted as Google Earth could ship without stronger abuse protections.

Google has said every image generated with Nano Banana carries a SynthID digital watermark, that it blocks image creation on harmful topics, and that generated content did not appear in the main Google Earth experience for other users to see. The company also said geospatial professionals were using the feature for legitimate purposes. However, Google acknowledged that screenshots of AI-generated imagery were shared publicly in ways that appeared to violate its policies — despite the watermarking and containment measures in place.

Sources differ on what happened next

Accounts of Google's response diverge. One characterization is that Google "initially rolled back" the feature while it implements stronger guardrails — implying a temporary pause. Another describes the move as Google "shutting down" or "scrapping" the feature outright. It remains unclear whether Nano Banana in Google Earth will return, and if so, when or with what changes. Google has not detailed what specific "harmful topics" or policy violations triggered the rollback, nor has it explained how a video generated with the tool managed to bypass Hive's detection system despite watermarking.

Why founders should care

This episode is likely to be read as a cautionary tale for any startup building generative AI features on top of trusted data sources or brands. A few takeaways stand out:

Pre-launch misuse testing may be non-negotiable. The speed of the rollback — one day — suggests that reputational damage from a poorly-guarded generative AI feature can materialize almost immediately once external researchers start probing it.
Detection tools are not a silver bullet. The fact that AI-generated content reportedly bypassed Hive's detection system suggests founders building or relying on verification products should treat third-party detection as one layer of defense, not a guarantee.
Watermarking claims deserve independent scrutiny. Google's SynthID watermarking and containment measures were in place, yet content still allegedly evaded detection and spread beyond intended boundaries. Founders relying on vendor safety claims may want to verify those claims independently rather than assume they hold up under adversarial testing.
Rapid response plans matter. Given how quickly this went from launch to rollback, teams shipping generative AI features tied to sensitive or trusted domains (maps, identity, verification) may benefit from having a fast-reversal playbook ready before launch, not after criticism hits.

What's still unknown

Several details remain unresolved: the scope and volume of misuse before takedown, what "stronger guardrails" Google plans to implement, any timeline for a possible relaunch, and how exactly the watermarking system failed to prevent the Hive detection bypass. Google also has not detailed the scope of the "useful purposes" it says geospatial professionals found for the tool before it was pulled.