Google's Nano Banana 2 debacle: How an AI image generator broke satellite imagery trust in 24 hours

Examples of fabricated satellite imagery

Researchers and journalists rapidly demonstrated the danger by generating convincing fake satellite scenes, including:

• An Iranian nuclear facility appearing damaged or altered
• Refugees at the U.S.–Mexico border (created by Dutch OSINT researcher Henk van Ess)
• A plane crash in Amsterdam
• A bomb crater in Gaza
• Iran's Kharg Island on fire
• A collapsed Eiffel Tower
• Damage at the Great Pyramid of Giza

Other fakes included a flooded U.S. Capitol, a blast crater in Los Angeles, and a protest scene .

What researchers warned

Henk van Ess, a Dutch open-source intelligence researcher, told The Guardian that the tool "refused nothing" — it generated refugees at the Mexican border, a nuclear plant in Iran, and a Gaza hospital scene without any content filtering . He warned that the feature fundamentally undermined the trustworthiness of satellite imagery, which for years has been a critical tool for open-source investigators and fact-checkers .

Jake Godin of Bellingcat told NPR that the tool could "fundamentally destroy the credibility of satellite imagery as evidence." He noted that satellite images have historically been among the most trusted pieces of OSINT evidence, and this tool made it trivially easy to create convincing fakes that would require significant effort to debunk .

Why SynthID watermarks failed

Google had embedded SynthID digital watermarks — both visible labels and invisible metadata — into every Nano Banana 2 image. The company initially told users to check suspicious images using Gemini or Google Lens . But researchers found several critical failures:

• SynthID watermarks can be cropped or stripped via simple screenshotting or resaving, removing the invisible metadata
• The watermark only covers AI-generated portions, not the underlying satellite imagery, making it easy to claim an image was "real satellite data"
• Some fabricated images passed SynthID detection checks entirely, meaning the watermark system did not reliably flag them as AI-generated

The TechTimes described the rollout as "the fastest significant AI product reversal driven by a specific safety failure" — the guardrail Google named as its primary protection simply did not stop the fakes from circulating .

Google's stated reason for rolling back

Google said it pulled the feature after seeing "people sharing screenshots of generated imagery that appear to violate our policies" . The company stated it would "add stronger guardrails" before considering re-releasing the tool . In a statement on X, Google wrote: "We take misinformation seriously — every image created with Nano Banana in Google Earth includes the SynthID digital watermark" . But the damage was done within hours.

Why this matters

The incident highlights a growing challenge for tech companies racing to integrate generative AI into products that users trust for factual information. Satellite imagery has long been considered a reliable source of evidence for journalists, human rights investigators, and open-source researchers. By making it easy to create convincing fakes that inherit the credibility of Google Earth's brand, the Nano Banana 2 feature briefly turned a trusted mapping tool into a potential engine for misinformation.

Google has not announced a timeline for re-releasing the feature.