Google Pulled Its AI Satellite Image Editor After One Day

Google launched an AI image-editing feature inside Google Earth on a Thursday, letting users rewrite satellite imagery with plain text prompts. Within 24 hours it was gone. Researcher Henk van Ess of Digital Digging demonstrated the obvious problem: the tool could generate convincing fake scenes, including refugees staged near the Mexican border and a bomb crater placed next to a hospital in Gaza. Google pulled the feature after one day and issued a revised statement walking back its earlier assurances.

What happened

Detail	Fact
Feature	AI text-prompt satellite image editing in Google Earth
AI model used	Nano Banana 2
Launch day	Thursday
Time before removal	Approximately one day
Researcher who exposed it	Henk van Ess, Digital Digging
Example outputs generated	Refugees near the Mexican border; bomb crater beside a hospital in Gaza

Google added an AI image-generation tool to Google Earth powered by a model named Nano Banana 2. Users could type a text prompt and the tool would alter the satellite imagery on screen to match it. The result was essentially a deepfake layer over real-world geography.

Henk van Ess, a researcher at Digital Digging, tested the tool deliberately and published what it could produce. His examples were stark: fabricated scenes that could be passed off as aerial evidence of a humanitarian crisis or a military strike. Both are exactly the kind of imagery that spreads quickly and is hard for a general audience to verify.

Why did Google pull the tool so fast?

On launch day, Google’s position was that Nano Banana 2 outputs carried a digital watermark and that the system blocked “image creation on harmful topics.” That framing did not hold. Van Ess’s examples showed the filters were not catching obviously sensitive geopolitical content.

By the following day, Google had issued a new statement and removed the feature entirely. The source article notes that the revised statement represented a clear reversal from the company’s initial response.

This pattern is familiar. A product team ships something that looks impressive in a demo, the safety framing goes out with the announcement, and then real-world use breaks the assumptions within hours. The gap between “we have a watermark” and “we verified the watermark actually prevents harm” is where these tools keep failing.

Why it matters

Satellite imagery carries an implicit authority. People trust aerial photos in a way they do not always trust street-level images. A tool that lets anyone rewrite that imagery with a text prompt creates a new category of misinformation: fake aerial evidence.

This is not just an abstract concern. The examples van Ess generated touched two of the most contested geopolitical topics in current news. If those images had circulated without context, the watermark would have been invisible to most viewers sharing a screenshot on social media.

For businesses and developers building on top of AI tools, this episode is a reminder that vendor safety claims need to be tested before you rely on them. Google’s initial assurances about topic filtering were falsified within a day by a single researcher using intentional but not especially sophisticated prompts. If you are using AI-generated content in any customer-facing context, the guardrails are rarely as robust as the launch announcement suggests. Our guide to protecting AI workflows in production covers how to think about this kind of failure mode in systems you actually control.

Our take

The fast shutdown is, honestly, the right call. What is more telling is that the tool shipped at all in this state. Editing real-world geographic imagery with generative AI is a genuinely high-risk capability. The fact that it passed internal review with “we have a watermark” as the primary safety argument suggests the risk assessment was not very deep.

Watermarks are useful for provenance tracking in controlled environments. They do nothing to stop a screenshot. Anyone using an AI tool for content generation, whether for AI integration into their own products or for marketing, should assume screenshots will strip any embedded metadata before the image reaches its final audience.

Google will likely revisit this feature in some form with tighter controls. When it does, the test should not be “can we block obviously harmful prompts” but “what happens when someone uses neutral-sounding prompts to achieve harmful outputs.” That is a much harder problem, and one watermarks do not solve.

What to do about it

1. Test AI content filters yourself before citing vendor safety claims to stakeholders. Assume the defaults are weaker than advertised.
2. Treat any AI-generated image as something that will be shared as a screenshot. Watermarks and metadata will not survive that journey.
3. Build a human review step for any AI-generated content that touches sensitive topics, geography, or real-world events.
4. Monitor changes to AI tools you depend on. Features can appear and disappear in 24 hours, as this case shows. Check the Lumien AI news feed for updates on tool changes that affect your stack.

Source: The Verge · AI