Google pauses Earth AI image tool after experts warn of fake satellite scenes

Google paused a new generative-image feature in Google Earth less than 48 hours after launching it on July 30, 2026, saying it would “work on implementing stronger guardrails” after researchers and journalists demonstrated the tool could produce realistic but false satellite scenes that spread easily online. The company said it had seen “people sharing screenshots of generated imagery that appear to violate our policies.” Google paused the feature on X and in a statement reported by BBC.

Why the pause matters: the tool let users prompt synthetic imagery layered over real satellite, aerial or 3D basemaps — effectively creating plausible-looking geospatial fabrications with Google’s branding attached. That combination, experts told reporters, lowers the bar for disinformation to look authoritative and circulate widely.

Google pauses "Nano Banana 2" integration after tests

Google introduced the web-only image-generation workflow — described in some coverage as integrating a model called “Nano Banana 2” — into Google Earth on July 30, 2026, then pulled or paused it within about two days after BBC Verify and AFP published test findings and France24/AFP reported that researchers could fabricate explosions, bomb craters and other false events. Google told reporters images created by its tools contain SynthID digital watermarks and that it blocks “image creation on harmful topics,” but acknowledged people had shared screenshots that appeared to violate those policies.

Google framed the feature’s purpose as helping users “visualize history, create real estate plans and more,” but the rapid rollback signalled a mismatch between product ambition and safety controls. The company also suggested users could verify generated content through its Gemini chatbot or Lens, per the BBC report.

AFP tests produce vivid false scenes despite watermarks

Independent tests published by AFP generated a range of fabricated scenes, including a claimed explosion in Paris, a nuclear site in Iran, bomb craters in Russia and an Islamic State training ground in Syria, plus humanitarian and migration scenes such as floods in Bangladesh and a migrant caravan at the US–Mexico border.https://www.france24.com/en/live-news/20260731-google-launches-new-satellite-image-ai-tool-alarming-researchers These examples underline the tension between Google’s technical mitigations — invisible SynthID watermarks and filters — and the practical ease of producing shareable screenshots that evade detection.

Experts quoted by AFP and BBC were blunt. Brady Africk of the American Enterprise Institute said the update “makes it easier to generate convincing fake satellite imagery that can spread quickly online and mislead the public.” An unnamed disinformation researcher told AFP: “The primary, and perhaps only, practical use of this tool appears to be the creation and dissemination of mis- and disinformation.” https://www.france24.com/en/live-news/20260731-google-launches-new-satellite-image-ai-tool-alarming-researchers

Those critiques expose a broader problem for major platforms rolling generative AI into trusted products: credibility is portable. When Google-branded imagery appears to show real-world events, even a small amount of polish can fool readers and spread faster than platforms can verify and label content.

Google’s track record with image-safety safeguards gives it tools to respond quickly, but the episode echoes prior launch missteps across the industry where company assurances about filters and provenance lagged behind on-the-ground misuse. Competitors such as OpenAI and Adobe offer image generation with different safety postures — OpenAI focuses on general-purpose model guardrails, Adobe markets enterprise controls — but none combine large-scale geospatial basemaps with consumer-facing synthesis at planetary scale, which is what raised alarm bells here. (bbc.com)

Closing: Google says it will rebuild stronger guardrails before restoring the feature; the next concrete metric to watch is whether the company publishes a transparent safety audit or third‑party tests showing the SynthID watermark and content filters reliably prevent the kinds of fabricated scenes AFP and BBC replicated. If Google cannot demonstrate robust, externally verifiable defences, regulators and civil-society researchers are likely to press for stricter limits on synthetic overlays of real-world geospatial imagery. (bbc.com) (france24.com)