Google shut down the launch of its AI image generator based on the Nano Banana model, integrated into Google Earth, in record time. The tool lasted less than a day — such a swift rollback points to serious flaws in content moderation during the testing phase.

The essence of the feature was to use text prompts to draw any objects or events directly on the map — from historical scenes to hypothetical disasters. However, within just a few hours of the release, generated images violating the service's rules began spreading online. Among the most viral examples were fake images of an explosion crater in Los Angeles and a flooded U.S. Capitol building.

Particularly alarming is the fact that the model did not filter even the most sensitive requests. Open data researcher Henk van Ess confirmed through his own testing: prompts related to depicting refugees at the Mexico border or nuclear facilities in Iran went through without any blocks. This is direct evidence that the safety system was either disabled or had not undergone proper calibration before the public launch.

The developers explained the removal of the generator by saying that users were mass-publishing created content that violated platform policy. However, such a formulation looks like an attempt to soften reputational damage: the problem is not user behavior, but the lack of built-in restrictions at the model level itself.

This incident is another signal for the entire industry. The rush to bring AI products to market without strict censorship mechanisms produces the opposite effect: instead of attracting an audience, companies face a wave of negativity and are forced to roll back releases. Google should learn a lesson from this failure and, in the future, implement generative features with multi-level filtering before they ever reach users' hands.

My comment: The situation is telling: even a giant with enormous resources cannot guarantee the safety of AI generations without preliminary stress-testing on real-world scenarios. The market expects not speed, but reliability — and this case will become a study subject in all laboratories working with generative models.