Ground Truth in the Age of Generative Cartography: What the Google Earth Rollback Actually Signals
For as long as satellite and aerial imagery has existed as a commercial and civic product, it has rested on a premise so foundational that it rarely gets stated out loud: a satellite or aerial image is a photographic record of photons that actually bounced off the actual Earth. It can be old, low-resolution, cloud-obscured, or badly georeferenced—but it is, at minimum, evidence. Courts, insurers, journalists, and human rights investigators have built entire verification workflows on that single assumption.
On July 30, 2026, Google introduced a feature into Google Earth's web version called "Create Image," built on Gemini's newest image-generation model—informally known as "Nano Banana 2"—and grounded, per Google's own announcement, in the platform's real satellite, aerial, and 3D mapping data. Users could type a single sentence and generate photorealistic scenery anchored to real coordinates on the Earth's surface.
Within 24 to 48 hours, it was gone. Google pulled the feature after TechCrunch reported a wave of criticism over exactly what anyone who has spent time in the OSINT or geospatial-verification community could have predicted: the feature made it trivial to fabricate convincing fake geospatial scenes of real places, and those fakes started circulating as if they were genuine.
The question for the geospatial industry isn't whether Google's generative image feature was technically impressive—it clearly was. It's whether the tradeoff it represented was ever defensible. At Sixty Carlton, our view is that it wasn't, and that the rollback was the correct call—but the reasoning matters more than the headline, because generative AI applied to geospatial imagery is not uniformly dangerous. It is dangerous specifically where it collides with a market that depends on imagery as ground truth. Understanding that distinction is the whole game.
Technical Deconstruction
Before assessing what went wrong, it's worth being precise about what "Create Image" actually is—and what it is not.
What it is: a generative image model, conditioned on real geographic coordinates and Google's basemap context (satellite, aerial, and 3D mapping data), that synthesizes new photorealistic pixels in response to a text prompt. The conditioning is what made the output convincing: the model wasn't drawing a generic aerial scene, it was drawing a scene stylistically and geometrically consistent with the real terrain, buildings, and coastlines already present at that location in Google Earth.
What it is not: a sensor. It captured no photons, recorded no event, and reconstructed nothing from an underlying physical measurement. That distinction matters more here than in almost any other generative AI application, because it maps directly onto a distinction Sixty Carlton has written about before in a different context—the difference between enhancing real signal and generating signal that never existed. AI super-resolution models applied to satellite and aerial imagery, for instance, amplify photons that genuinely reached a sensor; the output is a probabilistic reconstruction bounded by real captured data, and reputable vendors even ship pixel-level confidence layers to make that boundedness explicit. Generative image synthesis, by contrast, invents pixels with no photon origin at all. There is no sensor event to trace back to, no confidence layer that means anything, because the model isn't reconstructing a measurement—it's producing a plausible-looking image from scratch.
Google Earth's "Create Image" fell squarely on the invention side of that line. And it did so inside the one product in the world most explicitly branded, and most implicitly trusted, as a photographic record of reality. The strategic implication follows directly from the architecture: because the model generates pixels rather than reconstructs them, its output can be made visually and geometrically indistinguishable from a genuine capture without any underlying event ever having occurred. That is not a bug to be patched with a better model—it is the defining property of generative synthesis, and it is precisely why where this capability gets deployed matters as much as how well it works.
Market Segmentation: The Ground-Truth Fault Line
Understanding the real risk requires mapping generative geospatial AI against the specific use cases competing for budget and trust in this market.
Where generative geospatial AI is genuinely competitive:
is already a working commercial product. Climate Central's FloodVision combines stereoscopic video, LiDAR elevation data, and GPS positioning with generative AI to produce photorealistic renderings of what a specific real address would look like at a specific projected flood-water level. This isn't a novelty demo—it's used in FEMA funding applications and public-awareness campaigns, precisely because a resident staring at a rendering of water at their own front door internalizes flood risk in a way that a probabilistic floodplain map never will. NASA's Wildfire Digital Twin, developed with George Mason University, pairs AI models with satellite data assimilation to forecast fire spread and smoke dispersion in near-real time, per NASA's own description of the program.
has an emerging body of academic work using diffusion models to synthesize plausible development scenarios—testing how a rezoned block or a new transit corridor might visually and functionally evolve, as explored in recent diffusion-model research on urban synthesis. A parallel strand of research is using generative AI to "pre-enact" multi-hazard disaster scenarios in cities, feeding climate-adaptive resilience planning with visualizations of compound risks, as detailed in a 2026 study on generative disaster pre-enactment. The "digital risk twin" framework being formalized in journals like npj Natural Hazards treats this kind of generative modeling as core infrastructure for disaster risk management, not a gimmick—see the digital risk twin framework.
is the newest entrant, and the most instructive. Municipal redevelopment authorities have used renderings to sell blighted commercial corridors and brownfield sites to skeptical taxpayers and investors for decades—Sonoma County's Roseland Brownfield Revitalization Project, spanning more than 50 identified parcels along a single commercial corridor, is a representative case of how concept renderings are used to build community and investor buy-in ahead of remediation funding, an approach the EPA's own land-revitalization guidance explicitly recommends. What's changing is the tooling: the generative-AI-in-architecture market is forecast to grow from $1.48 billion in 2025 to $5.85 billion by 2029, with a large and growing share of architecture and planning firms already using generative tools for concept visualization, per industry analysis from Parametric Architecture. Extending the same diffusion-model techniques used for rezoning scenarios to full redevelopment pitches is a natural next step, not a hypothetical one.
What unites all three categories is a shared structural feature: the generated output is explicitly a model of a hypothetical, future, or counterfactual state, produced and consumed inside a context that makes that framing unmistakable. Nobody looking at a FloodVision rendering believes they're looking at a satellite or aerial photograph taken yesterday. Nobody looking at a proposed brownfield redevelopment rendering believes the boarded-up storefronts are already gone.
Where ground-truth imagery remains structurally irreplaceable:
OSINT verification, insurance claims adjustment, legal evidence, and human rights investigation all depend on the opposite property: that the image is not a model of anything, but a direct record of what a sensor actually captured at a specific place and time. These fields don't just prefer ground truth—their entire evidentiary logic collapses without it. A court cannot admit a "plausible rendering" of property damage. An OSINT investigator cannot corroborate a war crime with a "stylistically consistent" scene. The moment a generative layer sits inside the evidentiary record rather than clearly outside it—as it did with Create Image—the product no longer serves these use cases at all; it actively undermines them, for every user, whether or not they touch the generative feature themselves.
What Actually Went Wrong
The trigger event was a piece of independent research, not a leak or a whistleblower. OSINT researcher Henk van Ess published a Substack investigation demonstrating that single-sentence prompts inside Google Earth's new feature could fabricate a nuclear facility in Iran, refugees at the US-Mexico border, and a fatal crash scene in Amsterdam—all rendered directly onto real, verifiable coordinates. Once the technique was public, the fakes multiplied fast. BBC Verify independently produced a collapsed Eiffel Tower, a sinkhole swallowing the Great Pyramid of Giza, and Russian tanks rolling through Kyiv; other researchers separately generated flooding around the U.S. Capitol and a fire at Iran's Kharg Island oil terminal—a site whose real destruction would constitute a significant geopolitical event—as NPR reported. Each was generated in seconds and each was visually consistent with the surrounding real terrain, because the model was grounded in Google's actual satellite, aerial, and 3D mapping data.
Google's response, issued via its official statement on X, was unusually candid: "We know that people uniquely trust Google Earth for a reliable view of the world. We've seen geospatial professionals using this feature for a range of useful purposes, however we've also seen people sharing screenshots of generated imagery that appear to violate our policies. So we're rolling back this feature in Google Earth while we work on implementing stronger guardrails." Google noted the generated images were user-side only—never inserted into the shared, public Earth basemap—and carried SynthID watermarks intended to flag them as synthetic.
To be fair to Google, none of this appears to reflect negligence so much as underestimated velocity. There is no evidence of confirmed regulatory action, no disclosed count of policy-violating images, and no public relaunch timeline as of this writing—this reads as an open-ended pause rather than a permanent retreat, and geospatial professionals were, by Google's own account, using the feature for legitimate purposes in the same 48-hour window that bad actors were weaponizing it.
That safeguard did not hold up well under scrutiny regardless. Tech Times reported that BBC Verify found that watermark and detection checks could be circumvented in some cases, and that external AI detection tools also failed to catch certain generated images—not a wholesale failure of the technology, but enough of a gap to matter, given that a screenshot stripped of metadata and reposted to social media doesn't carry a visible watermark disclaimer with it. A screenshot is not a Google Earth session; it's an image file that can travel anywhere, attached to any caption, with no indication of its synthetic origin once it leaves the app. Watermarking a session-bound artifact does very little to prevent that artifact's most consequential downstream use.
This also wasn't the platform's first brush with weaponized satellite-adjacent imagery. During the Iran-Israel-US war, an AI-manipulated satellite image circulated claiming to show a "completely destroyed" U.S. radar installation at Al-Udeid Air Base in Qatar; fact-checkers traced it back to an old, digitally altered Google Maps image of the U.S. Fifth Fleet Naval Base in Manama, Bahrain, per reporting from Eurasian Times. Nor is satellite and aerial image manipulation a generative-AI-era problem at all—Bellingcat's landmark 2015 forensic analysis of MH17 documented the Russian Ministry of Defence photoshopping satellite imagery years before diffusion models existed. What generative AI changes is not the existence of the threat but its accessibility and speed: fabrication that once required a skilled photo editor and hours of work now requires one sentence and a few seconds.
Why Ground Truth Is the Whole Argument
Here is where the analysis has to be precise, because it would be easy to write this off as a garden-variety AI-safety story. It isn't. The reason Google Earth's feature was uniquely dangerous—more so than, say, a generic image generator producing a fake satellite-style picture with no coordinate anchor—is that it fused fabrication with geolocated authority. A fake image floating on social media with no location claim is one thing. A fake image generated inside the platform that OSINT investigators, journalists, insurance adjusters, and human rights researchers treat as a primary verification tool, anchored to real latitude and longitude, indistinguishable in style from the genuine basemap around it, is something categorically different.
This is precisely the framing offered by researchers at WITNESS in their analysis for Tech Policy Press, who describe Google Earth as part of the world's "global verification infrastructure"—a resource courts, journalists, and human rights investigators rely on precisely because it has historically been distinguishable from fabrication. Their core argument, and one Sixty Carlton finds persuasive, is that watermarking is not adequate governance for a tool embedded this deeply in verification workflows. A watermark is a technical afterthought bolted onto an image; it says nothing about whether that image should have been generated in that context in the first place.
The scale of the problem is also structurally different from ordinary misinformation, because detecting AI-generated satellite and aerial imagery is a genuinely hard, unsolved technical problem—not a solved one that Google simply failed to apply. Recent academic work on "Deepfake Geography" found that even the best-performing detection architecture, a Vision Transformer, achieved only 95.11% accuracy against a CNN baseline of 87.02%—meaning a meaningful share of fabricated satellite and aerial imagery evades even purpose-built detection tools. Layer that imperfect detection landscape under a wider erosion of OSINT's foundational assumptions—a trend the Reuters Institute at Oxford has been tracking since late 2025—and the picture becomes clear: this was never a contained product experiment. It was a live-fire test of whether the world's most-trusted mapping platform could absorb a generative image feature without contaminating the evidentiary commons that sits downstream of it. It could not, at least not without guardrails that did not yet exist at launch.
Threat Modeling
For the geospatial industry broadly—not just Google—this episode surfaces five specific threats worth tracking.
Every fabricated scene that circulates before being debunked chips away at the baseline assumption—that a satellite or aerial image anchored to real coordinates is presumptively real—that entire verification disciplines are built on. The damage isn't confined to the specific fakes caught; it's the slow-motion loss of default trust in all geolocated imagery, genuine or not.
Google is not the only company sitting on a trusted basemap and a generative image model. Apple Maps, Microsoft/Bing Maps, Esri, and Mapbox all face internal pressure to ship comparable "imagine this place" features to stay competitive on AI capability. The real threat isn't that one platform experimented recklessly—it's that competitive pressure pushes others to ship similar features before anyone solves the containment problem Google just demonstrated is unsolved.
Any platform that blurs the line between captured and generated imagery, even briefly or accidentally, invites downstream litigation risk—for the platform and for every party that relied on an image from it in good faith, in a claims dispute, a property assessment, or a court filing.
With Vision Transformer detectors topping out at 95.11% accuracy against fabricated satellite and aerial imagery, a meaningful share of synthetic content will keep evading purpose-built detection tools for the foreseeable future. Any strategy premised on "we'll catch the fakes on the back end" is building on a technical foundation that doesn't yet hold.
Google Earth is not an isolated product; it shares a trust halo with Google Maps and Google Search. A verification failure inside Earth is a data point competitors and regulators will cite when scrutinizing Google's other mapping and search products, whether or not those products actually share the same vulnerability.
Strategic Postures
Generative geospatial AI is not going away—the legitimate use cases above guarantee continued investment and continued pressure to ship. The question for every geospatial company is how to capture that value without repeating Google's mistake. Three postures are available.
Session-bound watermarking was never going to be enough, because a screenshot escapes the session. The more durable answer is portable, cryptographically verifiable content provenance—C2PA-style metadata that travels with the file itself, not just the app it was generated in—built and shipped before the generative feature, not bolted on after a public failure forces the issue.
The products that work—FloodVision, the Wildfire Digital Twin, diffusion-based urban planning tools—succeed precisely because they live on a structurally distinct product surface from the trusted basemap, with framing that makes the hypothetical nature of the output unmistakable. Any company building generative features into a product also used for verification should ask whether that generative capability belongs in a separate product entirely, not a toggle inside the trusted one.
Regulators, courts, and international bodies will eventually define what counts as adequate labeling and containment for AI-generated geospatial content. Companies that engage now—with standards organizations, journalism-verification coalitions like WITNESS, and geospatial industry groups—will shape what "responsible" means in this category, rather than have it defined for them after the next public failure.
The Bottom Line
The FloodVision, Wildfire Digital Twin, and brownfield-visualization examples prove that generative AI has a genuine, valuable role to play in geospatial products—when it is scoped to modeling hypothetical futures inside a context that unmistakably signals synthesis. Google's Create Image feature collapsed that distinction by embedding unconstrained, coordinate-anchored image generation inside the one product on Earth whose entire commercial and civic value proposition is being a trustworthy record of the present. The two are not the same category of risk, and treating them as such—either by banning all generative geospatial AI or by defending Google's rollout as merely misunderstood—misses the actual lesson.
The losers here are not generative AI vendors broadly; they are any geospatial platform tempted to bolt fabrication capability onto a ground-truth product without first solving the containment problem. The winners, at least reputationally, are the OSINT researchers, BBC Verify, and the independent fact-checkers who did in 48 hours what internal red-teaming apparently did not: proved that geolocated generative imagery breaks the fundamental contract of the product it lives inside.
Sixty Carlton's View: The real moat in geospatial data was never resolution, coverage, or even AI sophistication. It is evidentiary trust—the confidence that when a court, an insurer, or a journalist pulls up an image of a real place, they are looking at something that actually happened. If your own product roadmap has a generative feature planned anywhere near a basemap your customers treat as verification-grade, the question Google evidently answered too late is now yours to answer first: what happens the first time someone uses it exactly as designed, and the output looks exactly real enough to matter?
Join the conversation on LinkedIn