Google has abruptly withdrawn a new AI-powered image generator from Google Earth less than 24 hours after its debut, after researchers and reporters warned it could be weaponized to create realistic, misleading satellite imagery.
The experimental feature, powered by Google’s Nano Banana model, appeared in the web version of Google Earth on July 30. Users could zoom to any point on the globe, hit a “create image” button, and then describe in text what they wanted to see. The system would then synthesize a satellite-style image that matched the prompt, overlaid on top of the familiar Earth interface.
By the following day, the tool had quietly vanished.
In a statement on X, Google acknowledged pulling the feature and stressed that people “uniquely trust Google Earth for a reliable view of the world.” The company said that while some geospatial professionals had quickly found constructive uses for the tool, others were sharing generated images that seemed to run afoul of its usage policies. As a result, Google said it was suspending the experiment and “re-evaluating” how such generative capabilities should be deployed in mapping products.
Although the tool was online for only a short period, it immediately triggered alarm among journalists, human rights researchers, and open-source intelligence (OSINT) analysts. Many of them depend on Google Earth and other commercial satellite platforms to verify conflict footage, document war crimes, and fact-check breaking news in near real time. A generative overlay that can fabricate convincing scenes risks undermining the credibility of that entire workflow.
The core fear was not simply that the tool could draw “pretty pictures,” but that it could be prompted to produce satellite-style images of events that never happened-or alter the apparent state of real locations. Imagery hinting at non-existent troop movements, destroyed infrastructure, mass graves, or fabricated environmental damage could be circulated as if it came from authentic satellite passes. Even if Google clearly labeled such creations, screenshots stripped of context could still spread across social platforms as supposed proof.
Researchers also noted that the user experience itself-allowing AI renderings to appear inside Google Earth-could blur a critical line. For years, the application has been perceived as a quasi-authoritative visual record of the planet: imperfect, but based on real-world sensor data. Embedding AI-generated scenes in the same interface could make it harder for casual users to distinguish what is observational imagery and what is synthetic.
This is especially sensitive for OSINT practitioners, who often start with satellite views from tools like Google Earth, then cross-check those images with videos, photos, and ground reports. They rely on the assumption that the base map is a fixed, factual reference point. Any erosion of that trust-whether through mistakes, manipulations, or the presence of generative overlays-risks cascading confusion in an already challenging verification landscape.
Google, for its part, has argued that generative AI can unlock powerful new use cases in geospatial analysis: quickly visualizing flood risk, simulating land-use changes, or drafting illustrative scenes for presentations and education. Early testers reportedly explored benign scenarios such as imagining a forest in different seasons or visualizing speculative city layouts. But those same capabilities make it trivial to create “evidence” of things that never occurred.
The broader context is an AI industry still grappling with deepfakes. Synthetic faces, voices, and videos have become increasingly lifelike and accessible. Until now, satellite imagery has been more resistant to casual fakery: it requires domain expertise, specialized tools, and a credible source. A ready-made generator built into a mainstream mapping product significantly lowers the barrier.
The episode also underscores a tension in how tech giants roll out experimental features. Many companies have adopted a “ship fast, iterate later” mindset, especially around AI. But when the product in question underpins journalism, human rights work, or crisis response, the cost of missteps is higher. OSINT analysts have repeatedly called for core reference tools-maps, satellite feeds, archival imagery-to be treated with more caution than consumer-facing chatbots or creative apps.
One of the central ethical questions is how to separate “illustrative” and “evidentiary” imagery. There is a strong case for generative tools that help people understand climate models, urban planning scenarios, or disaster preparedness by visualizing hypothetical futures. Yet those AI visualizations look increasingly similar to real aerial photographs. Without strict guardrails, prominent watermarks, and segregated interfaces, synthetic scenes can easily escape their intended context and be repurposed as alleged documentation.
Some experts argue that if generative imagery is ever allowed near products like Google Earth or other mapping services, it should live in a clearly distinct mode, with unmistakable visual treatment that differentiates it from real satellite data. Others contend that such tools should simply never coexist with real-world imagery in the same environment, given the stakes for misinformation and conflict reporting.
The speed of Google’s reversal suggests the company underestimated the reputational risk. It also highlights a growing difference in expectations between platform developers and specialist user communities. Engineers may see a creative sandbox; investigators see a potential contamination of one of their most important truth-finding tools.
More broadly, the controversy feeds into a larger debate about the future of trust online. As generative AI seeps into search results, mapping tools, office software, and image archives, users will increasingly struggle to know when they are looking at recorded reality and when they are seeing a machine’s best guess-or a user’s fictional prompt. The stakes are especially high for imagery that has long carried an aura of objectivity, such as satellite photography, CCTV footage, and newswire photographs.
Mitigating that risk will likely require a combination of technical and policy approaches. Technically, stronger provenance standards-such as cryptographic signatures attached at capture time-can help confirm when imagery is sensor-derived and unaltered. At the policy level, companies may need to adopt stricter pre-release risk assessments for AI features that intersect with critical information infrastructures like maps and factual archives.
There is also a role for user education. Journalists, researchers, and the general public will increasingly need basic literacy in how AI-generated images look, how they are labeled, and where they are most likely to appear. Yet leaning solely on media literacy is not enough; system design must minimize opportunities for confusion in the first place.
In the short term, Google’s decision to pull the Nano Banana image tool removes an immediate source of concern. But it does not resolve the underlying issue: geospatial deepfakes are now technically feasible and will continue to improve. Other companies or open-source projects may attempt similar tools, and malicious actors have every incentive to experiment with fabricating satellite-style imagery outside of major platforms’ oversight.
That leaves Google and its peers with a difficult task. They must find ways to harness generative AI’s value-making complex spatial data more understandable, more accessible, and more interactive-without corroding the fragile trust that underpins fact-checking, human rights documentation, and public understanding of global events. The brief life and rapid death of Google Earth’s AI image feature is likely to become an early case study in how hard that balance will be to strike.