Google Earth: twenty years to build trust. Twenty-four hours to put it at risk

Google Earth has existed since 2005. In twenty years it has become the reference visual source for journalists, researchers, investigators and courts. Even the United Nations uses the tool, which has earned a fundamental evidentiary role in global affairs.

After all, a satellite image is credible because it cannot be invented. And over the years, that is what set Google Earth apart from every other platform.

In twenty-four hours, however, Google put all of this at risk by integrating a generative artificial intelligence system into Google Earth.

Why did it do this? To let users give free rein to their imagination, selecting any point on the planet and altering it. A creative outlet that was meant to let people see the Colosseum at the height of its splendour, or Pompeii during the eruption. Yet that is not how it turned out.

Within a few hours, across the globe, there appeared: King Kong; nuclear power plants in Iran; Mexican refugees along the US border; nuclear craters in Gaza. And, of course, planes flying into New York skyscrapers.

No sensitive prompt had been refused. A clear AI Governance failure, which forced Google to withdraw the feature in under twenty-four hours, putting its own reputation at risk.

The underlying error

The decision that produced this outcome is a clear category error.

Whoever approved the integration of Nano Banana 2 into Google Earth treated Google Earth as a creative platform, which Google Earth is not. Adding an AI image generator to a creative platform would have been an extension of functionality. Adding an AI image generator to an evidentiary standard, by contrast, is an attack on its core function, and therefore on its entire architecture of value.

Is it possible that Google did not know the real function of its own tool? Or did it genuinely believe people would not use the feature in ways like this?

The defence that does not defend

Google’s initial response to the criticism was a watermark: every generated image contained SynthID, the invisible marker for the company’s AI products. A pointless response, given that during a conflict any fake satellite image can still destabilise. In the heat of the moment, nobody checks the watermark (as we had already seen in early 2026 in the war between the United States, Israel and Iran). Nano Banana 2 would have made that process trivial.

The watermark depends on “proper” use. But what about those who want to spread disinformation? Using it as a safety argument, therefore, shows how little the problem has been understood. And Google, clearly, did not understand much of it.

No testing, no safeguards

Clearly, no one asked before launch: “What would happen if a user generated a nuclear power plant in Iran?”

And yet adversarial tests exist precisely for this: to deliberately seek out the scenarios in which a system can fail, before users discover them in production.

Eliot Higgins, founder of Bellingcat, wrote that “Google launched a tool that appears not to have passed the most basic adversarial testing before going into production. That testing was then carried out by strangers, for free, at Google’s expense.”

The governance of an AI service cannot be a phase that follows launch. It must begin before launch, before any damage to trust.

The lesson for any organisation with a trust asset

Google will be fine. It is so large, so entrenched, that it can withstand even a crisis like that of Nano Banana 2 on Google Earth, withdrawn within twenty-four hours.

Few other companies, however, could withstand the same phenomenon.

We are well aware of the pressure today to integrate AI capabilities into every existing product and service (indeed, we discussed it in our article on AI Washing). However, one must not give in to the temptation to do so regardless of the product category. Or at the expense of service quality.

The governance question to ask before any AI integration is not only “what can this tool do?”, but: “Which property of the product makes what we already do credible?” That question must then be paired with another: where must human oversight remain before a new capability reaches the public?

If the integration preserves value, then it is right to proceed. If it erodes it, or could erode it, it should without question be avoided.

Google can withstand impacts like this. Smaller organisations, however, may not.