During my time working at Google, we used to describe Google Earth and Maps as a mirror of the real world.

It wasn’t just a marketing slogan; it was a guiding principle. To us, the earth-observation and geospatial teams, the core value proposition of Google Earth was objective truth. It was a digital double of our planet built on photons bounced off satellites, stitched aerial photography, and meticulously modeled 3D geometry. If you zoomed into a street corner in Tokyo, a rural farm in Kansas, or Teddington High Street, you trusted that what you were looking at actually existed.

That absolute bedrock of trust was structurally shattered this week—briefly, thankfully, but dangerously—when Google rolled out (and subsequently had to reverse) its “Nano Banana 2” AI image generation feature directly inside Google Earth.

Yes, you read that right. A feature named after a fruit was unleashed to let users paint synthetic, hallucinated realities straight onto the planet’s trusted baseline layer.

The “Nano Banana” Blunder: When a Mirror Becomes a Magic Eye Poster

The pitch from the product team was dripping with standard Technical Product Manager buzzwords. They wanted to help users “visualise the past,” dream up real estate projects, or see what an empty lot might look like as a community garden.

Sounds harmless enough in a vacuum, right or on your desktop GIS system?

Except the internet immediately did what the internet always does when handed a frictionless tool for synthetic fabrication. Within hours of launch, researchers, OSINT investigators, and digital watchdogs began using Nano Banana 2 to drop fake refugee crises at borders, plant nuclear power plants in restricted zones, and simulate fatal crashes or geopolitical flashpoints on real streets.

Because the resulting screenshots maintained the authoritative interface of Google Earth, they could easily be weaponized as pristine disinformation. By July 31st yes that’s just yesterday !—barely 24 hours after shouting about how users could now render ancient Pompeii or futuristic sci-fi utopias—Google was forced to retreat, pulling the feature and admitting they needed to work on “stronger guardrails”.

How did a company that once deeply respected the sanctity of geospatial truth sign off on something so fundamentally reckless?

How did a company that once deeply respected the sanctity of geospatial truth sign off on something so fundamentally reckless?

The Perverse Mandate: “Add AI to Everything”

We are living through a corporate panic. In boardrooms across Silicon Valley, leadership is terrorized by the prospect of being left behind in the generative AI gold rush. The mandate from the top has devolved into a blunt directive: Add AI to everything, everywhere, immediately. I felt this at the end of my time at Google I can only imagine how much stronger this has become..

It doesn’t matter if the product calls for it. It doesn’t matter if the product’s entire historical equity is built on being an objective record of reality. If a product doesn’t have a generative text-to-image prompt box slapped onto it, senior leadership panics.

In their haste to check an AI box, the architects of this feature completely lost sight of what makes Google Earth valuable. Google Earth is not a canvas for creative writing; it is a reference library. You don’t want an AI “interpreting” or “reimagining” a map any more than you want an AI to creatively hallucinate your bank statement.

A Total Amnesia of “Geo” Values

OK this sounds like the rantings of someone who used to work for Google Geo of course, but I believe this blunder also points to a deeper cultural rot within the current configuration of the company: a severe lack of institutional memory.

When you cycle through staff, or when leaders who have no foundational background in Geo-products are handed the keys to legendary mapping infrastructure, you get a profound disconnect. New arrivals to the Geo team saw Google Earth as a static UI waiting to be “unlocked” by generative media. They treated it like a blank social media feed.

Veterans of the Geo organization understood an unwritten rule that took decades to build: Geospatial trust is hard to win and terrifyingly easy to lose. Google Earth is relied upon by human rights investigators, journalists, emergency responders, and courts of law as a baseline truth layer of our planet. Once you train the public to accept that the satellite view can be whimsically altered by typing a prompt about a “lakeside cabin” or a “futuristic metropolis”, you destroy the integrity of the platform. If anything on the screen can be AI-generated, then nothing on the screen can be trusted.

The Mirror Can Never Be Unbroken Entirely

Google’s quick rollback of Nano Banana 2 is a welcome relief, but the damage to the narrative is already done. The genie is out of the bottle. Bad actors have already captured screenshots, and the illusion of Google Earth as an trusted, objective mirror of the world has a permanent crack in it.

If Google wants to retain any remaining credibility for Google Earth, leadership needs to stop treating its core products as testbeds for whatever experimental model management wants to hype next.

Google remains a trustworthy company, however the technology deployed is not Google’s alone and there are other organisations who would use this capability for their own goals, this is a wake up call to the industry to be careful what you trust when it comes to Geospatial as much as Social Media these days.

keep the hallucinations out of the geography

Some things in this world should remain real. Leave the bananas in the fruit bowl, and keep the hallucinations out of the geography.

Who else got into geography to change the world rather than just digitise it?

If you read Rohan Silva’s recent essay in The Times, you’ll have seen him articulate a shift that is quietly rewiring the way we think about knowledge work. Silva captures a profound transition brought on by the rise of Artificial Intelligence: we are moving away from a world defined by mastering the how, and entering a world where value is entirely driven by the what.

Reading his piece, I couldn’t help but reflect on how perfectly this mirrors the evolution—and the immediate future—of our own geospatial industry.

For decades, to be a “Geographer” or a geospatial professional meant acting as a gatekeeper of the how. If a city planner, a climate scientist, or a logistics manager wanted to understand the spatial dynamics of their problem, they had to come to us. And our work was intensely mechanical. We spent our days wrangling shapefiles, fighting with map projections, writing complex SQL queries, and navigating the Byzantine user interfaces of desktop GIS software.

The barrier to entry for spatial analysis wasn’t the ability to think spatially; it was the ability to operate the machinery. We were defined by our technical execution.

AI is dismantling that barrier, and it is doing it faster than many in our industry realise, a point made forcefully by Peter Rabley at this year’s GeoIgnite.

With the advent of large language models and generative AI, the mechanics of spatial analysis are being abstracted away. We are rapidly approaching a point where a user can simply look at a map and ask, in plain English: “Show me the neighborhoods with the highest density of elderly residents that are also within a 100-year flood zone, and highlight the ones lacking accessible public transit.”

The AI will write the query, fetch the data, perform the intersection, and render the visualization. The machine handles the how.

For some in the geospatial community, this feels like an existential threat. If a machine can run a spatial buffer and render a heat map in three seconds based on a voice prompt, what is the role of the GIS analyst?

But as Silva’s essay implies, this isn’t the end of our profession—it is a massive elevation of it. By stripping away the friction of the how, AI frees geographers to focus entirely on the what.

When the mechanics of map-making are automated, the true value of a Geographer is revealed. Our expertise was never really about knowing which buttons to click in a software package; it was about understanding the world as a complex, interconnected spatial system.

In the AI era, our job shifts from being operators of software to being interrogators of data.

• What are the ethical implications of the spatial models we are building?
• What context is the AI missing about this local community?
• What are the systemic spatial inequalities—in health, climate vulnerability, or infrastructure—that we should be pointing these powerful new tools toward?

AI doesn’t know what questions are worth asking. It doesn’t possess geographical curiosity. It doesn’t understand the lived reality of a neighbourhood, the historical context of a border, or the nuances of human geography.

We are moving from a paradigm of spatial mechanics to one of spatial strategy. AI is finally allowing us to put down the digital spanners, look up from our screens, and focus our energy on what we are actually trying to solve. For anyone who got into geography to change the world rather than just digitise it, this is exactly where we want to be.

Next month’s blog post will be all about my other passion.. aviation photography, it’s the holiday season after all !

Geographers are Architects, Not Draughtsmen

For decades, the primary output of our profession was the map. We were the custodians of the lines, polygons, and points that represented the physical world, carefully plotting data for others to interpret. But as our industry evolves deeper into the era of Location Intelligence and GeoAI, the role of the geographer must fundamentally shift. We can no longer afford to operate simply as draughtsmen, rendering the environment onto a screen.

To remain relevant and support the next generation of decision-making, we must move up the value chain. We must become architects.

From Cartography to Geographic Architecture

Looking back at the history of WebGIS, the development of the internet had a profound impact on our field. It successfully democratized geographic information, taking maps out of the hands of specialists and putting them into the pockets of the public. However, in the rush to make spatial data universally accessible, the deeper underlying geographical principles were often left behind. We delivered the visualisation but frequently lost the analytical rigor.

Today, artificial intelligence presents a parallel, yet far more profound, opportunity. Just as the web democratized the viewing of maps, AI has the potential to democratize the techniques of quantitative geography. It can take complex spatial analysis out of the academic silo and make it an accessible, everyday tool for solving real-world problems.

Why AI may changes the Role..

But there is a catch. To achieve this, we have significant work to do to ensure that AI systems understand the unique nature of geographic information. Spatial data is not just tabular data with coordinates attached. If we simply feed locations into Large Language Models, we quickly see their limitations—such as their inherent struggles to genuinely grasp the topological and spatial realities of something as fundamental as a city’s street network.

AI must be explicitly taught to comprehend foundational geographic concepts. It needs to account for spatial autocorrelation—the reality that near things are more related than distant things—and it must be able to navigate the persistent statistical traps of the Modifiable Areal Unit Problem (MAUP).

Relying solely on opaque geospatial foundation models or black-box spatial embeddings will not be enough here. We cannot entrust critical spatial reasoning to systems that cannot explain their geographic logic. We need transparent approaches that respect the science of “where.”

As I have noted previously the massive momentum to develop “Word Models” through a brute force recreation of synthetic worlds based of sensor data may shortcut this need for fundamental understanding, but surely an approach that embeds geographical reasoning based on the principles of geographic knowledge offers a useful shortcut?

Decision Layers in Practice

This is exactly where the geographer as an architect becomes indispensable. Our focus must shift from designing general-purpose maps to engineering “Decision Layers.”

This is a vernacular that is rapidly gaining traction in the defence and intelligence communities, and it represents a fundamental evolution in how we view our output. In this framework, the map itself becomes the Decision Layer. It is no longer a passive; exploratory repository of topographic features and points of interest left to the user to decipher. Instead, it is a highly targeted, dynamic visualization synthesized to illustrate a very specific decision point.

the map becomes the Decision Layer

A true Decision Layer strips away the extraneous “noise” of traditional cartography. It presents only the exact spatial variables, constraints, and operational realities required for a leader to make a precise call. Whether it is a battlefield commander determining a line of advance, a logistics director re-routing a compromised supply chain, or an urban planner approving a new infrastructure project, the map exists solely to provide the distilled, actionable truth for that singular moment.

A draughtsman asks, “How should this data look?” An architect asks, “What spatial logic is required to drive this specific decision?”

We must be the ones who define what these Decision Layers are, ensuring the right context, scale, and relationships are modelled into the AI systems that generate them. It is no longer just about obtaining the appropriate content; it is about deeply understanding the decision-making process itself.

In an AI-powered world, the ultimate value of a geographer isn’t in drawing the map—it is in architecting the spatial reasoning that powers the decision. It is time as Geographers we step fully into that role.

The Problem with Local Trust

The core issue appears to be one of local network architecture. In its original implementation, if a user enabled “Share Location” in the Starlink debug settings, the dish would broadcast its coordinates to any device on the local area network (LAN) without requiring authentication.
In a world of “Zero Trust” architecture, this is a significant vulnerability. Any compromised IoT device on the same Wi-Fi network—from a smart fridge to a cheap security camera—could silently poll the dish for its exact physical coordinates. Furthermore, mobile apps that usually require OS-level permissions to access a phone’s GPS could bypass those restrictions entirely by simply “asking” the Starlink dish over the Wi-Fi.

Geopolitics and Kinetic Risk

Beyond the digital privacy concerns, there is a very real physical safety dimension. Starlink has become critical infrastructure in modern conflict zones.
In these environments, location data is a weapon. By leaving a local, unauthenticated API active, SpaceX was inadvertently creating a “homing beacon” for anyone who could gain even a foothold on a local network. If a terminal’s coordinates can be scraped via a simple script, that terminal (and the people using it) becomes a target for kinetic strikes. By moving this data behind an authenticated, cloud-based wall, SpaceX is essentially “hardening” the terminal against being used as a targeting coordinate source.

The Impact on the “Resilient PNT” Community

The removal is a blow to those looking for Resilient Positioning, Navigation, and Timing (PNT). Many maritime users relied on the Starlink location as a “sanity check” against their primary GPS. Because Starlink operates in Low Earth Orbit (LEO) with high-gain directional antennas and complex encryption, it is significantly harder to spoof than traditional MEO-based GNSS signals.
While the “pseudo-GPS” wasn’t a formal service, it was a proof-of-concept for how LEO constellations can serve as a backup to our aging GPS infrastructure.

Where do we go from here?

SpaceX isn’t removing location data entirely; they are shifting it to their Telemetry API. The catch? This is largely an enterprise-facing feature, likely requiring higher-tier subscriptions and proper authentication tokens.
This move signals the end of the “Wild West” era of Starlink data. As the platform matures from a disruptive startup service into a piece of global critical infrastructure, the “fun” features that allow for easy tinkering are being traded for the “boring” but necessary features of security and liability management.
For the geospatial professional, it’s a reminder that even the most innovative positioning sources are ultimately beholden to the security requirements of the platforms they run on.

But the tectonic plates of artificial intelligence are shifting. As a recent MIT Technology Review piece highlighted, the frontier of AI is rapidly advancing beyond Large Language Models (LLMs) that merely predict text, toward “World Models” that simulate physical reality. Driven by visionaries like Fei-Fei Li (World Labs), Yann LeCun at Meta, and heavily funded initiatives from Google DeepMind and OpenAI, world models seek to endow AI with spatial, physical, and causal “intuition” if not understanding!

For the geospatial industry, this represents far more than a software update. It is the paradigm shift we often talk about. It threatens the core business models of traditional data providers and GIS toolmakers alike, forcing us to confront a profound philosophical question: what happens when the market for our digital representations of the Earth transition from measured realities to synthetic, probabilistic simulations?

The Brittleness of Language vs. The Geometry of the Map

To understand the specific nature of this disruption, we must first dispel a common misconception about the capabilities of current AI. The MIT Technology Review highlighted a fascinating study where an LLM was asked to navigate a simulated New York City. The model could flawlessly recite turn-by-turn directions from one point in Manhattan to another based on its training data. However, the moment it was forced to take a detour, it failed catastrophically.

This failure was not a critique of modern mapping technology; it was a glaring exposure of the LLM’s spatial brittleness. The LLM was simply predicting the next logical word in a sequence. It had no “mental map” of New York, no understanding of intersecting grids, and no intuition for spatial workarounds.

Traditional geospatial routing—the kind powering Google Maps or a HERE navigation system—does not suffer from this specific brittleness. If a water main breaks on 5th Avenue and the road is closed, traditional routing algorithms instantly recalculate the optimal path. They do this brilliantly using established mathematical models (like Dijkstra’s algorithm) applied to a highly structured database of road networks.

However, this traditional system, while mathematically robust, is fundamentally rigid. It is a series of hard-coded spatial queries run against a static database. The routing algorithm doesn’t “know” what a water main break is, nor does it understand the physical physics of traffic flow; it merely knows that a specific line segment on a graph now has an infinite time-penalty, so it searches for the next mathematically shortest line segment.

This is the exact limitation that World Models are being built to solve.

From Models to Engines

A World Model does not just query a database of street nodes; it simulates the environment itself. It operates much closer to human spatial intuition. If a human encounters a blocked street, they don’t just calculate the next mathematically viable sequence of turns; they understand the physical constraints of the neighbourhood, the flow of pedestrians, the width of the city streets, and the likely downstream effects of the blockage.

World Models aim to give AI this causal and physical understanding of the environment. Google DeepMind and World Labs are already building models that generate interactive, 3D virtual environments from simple prompts. These aren’t just pretty 3D pictures; they are physics-aware models.

For the geospatial industry, the leap from a “database of coordinates” to a “causal simulation of reality” renders traditional methodologies incredibly vulnerable. If an AI can natively understand spatial relations, cause-and-effect, and physical geometry, the old way of managing spatial data begins to look like using an abacus in the age of the microchip.

The Commoditisation of “Ground Truth”

The most immediate and existential threat will be felt by the geospatial data companies—the maintainers of the map. Firms like HERE and traditional surveying organisations like the Ordnance Survey possess immense competitive moats because they own proprietary, high-precision, heavily curated datasets. Their entire product is verified “ground truth.”

World models threaten to commoditize this ground truth by generating it synthetically and continuously. The article notes that Niantic (the creators of Pokémon Go) is utilizing billions of crowdsourced smartphone images to build the pieces of a spatial world model to guide delivery robots. This bypasses the need for traditional, centralised mapping fleets.

More radically, world models possess the ability to interpolate and probabilistically generate spatial data. If a World Model has ingested enough video data of a city’s architecture, street widths, and typical traffic patterns, it doesn’t necessarily need a fresh LIDAR scan of a specific side-street to know what is there. It can probabilistically simulate the street, complete with physics-compliant surfaces, lighting, and spatial boundaries, on the fly.

If the tech giants can generate real-time, interactive 3D simulations of any environment using a mix of text, scattered crowdsourced video, and predictive spatial intelligence, the business model of selling static, highly expensive HD maps faces an inevitable collapse. The data companies will be forced into a painful pivot: their vast historical archives will be incredibly valuable as initial training data for these models, but once the models are robust, the ongoing value of traditional, manual map updates will plummet.

They must transition from selling records of the past to facilitating predictions of the present.

The Toolmakers’ Dilemma: ESRI and the Generative Leap

While data companies face commoditisation, the software toolmakers like ESRI face the threat of obsolescence through user-interface revolution. ESRI’s ArcGIS is the undisputed heavyweight of spatial analytics. It is an indispensable tool for urban planners, environmental scientists, and logisticians.

Yet, GIS is fundamentally analytical and representational. The workflow is manual and layered: a user imports data, applies spatial joins or buffers, runs a query, and outputs a 2D or 3D visualization.

World Models represent a leap from the analytical to the generative. Instead of an emergency planner using GIS software to overlay a flood-risk polygon onto a city map to calculate affected building footprints, a World Model allows the planner to simply prompt: “Simulate a Category 4 hurricane hitting the Miami coastline at high tide, and highlight structural failures in residential zones.” The model—understanding the physics of fluid dynamics, the structural integrity of different building materials based on historical data, and the 3D topography of the city—would generate a real-time, interactive simulation of the disaster. The user isn’t joining tables or managing shapefiles; they are conversing with a physics-engine that understands geography.

If AI systems can natively execute complex spatial workflows through natural language and return interactive simulations, the traditional GIS interface becomes a bottleneck. To survive, ESRI and its competitors cannot continue with their current approach of just bolting an LLM chatbot onto their existing software. They must fundamentally rebuild their platforms, transitioning from passive repositories of spatial layers into active, world-simulating engines.

The Philosophical Chasm: The Map That Dreams

Beyond the shifting corporate landscapes and disrupted business models lies a profound philosophical dilemma. The geospatial industry has always been anchored to a sacred concept: absolute fidelity to physical reality. The map is a contract of truth. If a map says a road exists, the road must exist.

But as we transition to World Models, we enter the territory famously described by French philosopher Jean Baudrillard in Simulacra and Simulation, referencing the analogy of Borges’ Map. Baudrillard theorised a state where the simulation of reality, as illustrated by the famous short story of Borges’ Map, becomes so pervasive and detailed that it precedes and eventually obscures the real world—where the map becomes the territory.

World Models are inherently probabilistic. When a system generates a 3D environment based on a mix of real data and predictive algorithms, it is not merely recalling reality; it is hallucinating a highly plausible reality based on statistical weights.

What happens when we begin to run our physical world based on the outputs of a synthetic simulation?

If an autonomous vehicle navigates a city street using a generative world model rather than a deterministic HD map, it is navigating a probabilistic representation of that street. If the model statistically determines that a dark patch of asphalt is likely a shadow rather than a pothole, or that a newly constructed glass facade reflects open sky, the synthetic world clashes violently with the real one.

The well-documented danger of LLMs is that they confidently hallucinate facts. The impending danger of World Models is that they confidently hallucinate reality.

For a traditional cartographer, an error is a mislabeled street—a verifiable departure from ground truth that can be manually corrected. But in a World Model, the concept of ground truth is fluid. If an urban planner uses a world model to redesign a traffic intersection, and the model subtly hallucinates the turning radius of a delivery truck because its internal physics engine approximated the data, the resulting real-world concrete will be poured based on a synthetic lie.

We are moving from an era of cartography to an era of spatial generation. The cartographer meticulously measures what is; the generative AI probabilistically dreams what might be.

Navigating the Uncharted

The developments discussed by MIT Technology Review—the reallocation of OpenAI’s resources toward world simulation, the birth of World Labs, the laser focus of the industry’s brightest minds—are the latest warning sirens for the geospatial sector.

The era of the static, queried map is drawing to a close.

Legacy companies have survived massive technological shifts before, evolving from paper charts to digital databases, and from desktop software to cloud infrastructure. But the rise of the World Model is entirely different. It is not a new medium for displaying spatial data; it is a fundamental replacement for how spatial intelligence is computed.

To survive the coming decade, the geospatial industry must accept that its future does not lie solely in capturing reality with higher fidelity. The future belongs to those who can build the most robust, physics-aware, and dynamically predictive simulations of reality. They must evolve from being the archivists of the Earth to becoming the architects of its digital twin.

Yet, as we eagerly hand over the spatial mechanics of our world to generative models, we must proceed with profound caution. In our rush to build AI that truly “understands” the physical world, we run the very real risk of creating systems that replace our shared, tangible reality with a plausible, synthetic dream.