TL;DR: Current geospatial AI (including satellites, sensors, and Earth models) is great at telling us what is happening in real-time, but still relies on humans to figure out why and what to do. The author argues the missing piece is causal reasoning — AI that doesn’t just perceive the world but understands cause and effect, can simulate intervention scenarios, and hands humans pre-reasoned decision paths rather than just better maps. They call this “Causal Planetary Intelligence.

---

Will Marshall‘s recent Substack essay on Planetary Intelligence (PI) is the clearest articulation I’ve encountered of where geospatial is genuinely heading. It’s a must read.

I want to extend that conversation, because I think there’s one critical piece still missing from the architecture.

The pieces of the puzzle

Will describes Planetary Intelligence (PI) as a real-time sensing and computing system that continuously observes the entire physical Earth and converts that observation into actionable decision support for humans. He breaks PI down into 4 components:

• Sensing: Continuous data streams – satellites, IoT, weather networks, cameras etc.
• Modelling: Large Earth Models (LEM’s) trained on that sensor data, analogous to how LLMs were trained on internet text.
• Edge Computing: Real-time data processing in space to reducing latency.
• Decision support: The output that flows back to human decision-makers inside their OODA loop – Observe, Orient, Decide, Act.

As Will also discusses in The Spillover Podcast, this gives humans more and faster information for decision making. In other words a faster delivered picture of what is happening right now – situational awareness – and the insight to help those who need to act, take decisive action ‘before it is too late.’ This makes ‘the collective smarter than its parts’.

Will’s argument is centred on bandwidth to reduce decision latency. PI and LEM’s are critical foundations, but as I will argue, are the first stage. The critical bottleneck still in place is the decision making bottleneck, that is the second stage and what I will address here.

What Geography Always Promised

I’m a passionate geographer. This is a discipline centred on causal explanation, helping us understand better the physical world. We explore questions like: why is Big Cottonwood Canyon – here in Utah – V shaped while Little Cottonwood Canyon, 2 miles further south, U shaped? We observe, gather evidence, build and test our hypothesis. Geography’s final output has always been understanding, not action.

Geospatial – GIS and remote sensing – has long been associated with geography. Software and toolkits (platforms) which allow us to process, store, analyse and visualize location based data. The commercialization of GIS shifted that final output, actionable insight delivered to decision-makers by analysts became the end product.

AI in the form of machine learning (ML) and generative AI (LLM’s) have added new tools to the toolbox. But the current geospatial industry business model is centred on an analyst in the middle. What I call the current paradigm or Horizon One. Will indirectly touches on this in his framing, suggesting the human moves from processor to strategic authority. The question he leaves open is: what does the machine need to become for that elevation of the human to actually happen?

This is the real paradigm shift or Horizon Two.

From Correlation to Causal Understanding

When we were children, we learned by exploring and observing. That means using our senses to understand the world around us and learning from that interaction – letting go causes a ball to fall to the ground.

We were building our own world model.

As Will points out LLM’s are pattern matching against text, predicting relationships between words rather than understanding the world those words describe. They are the librarian who has read every book in the world, can expertly draw from what he/she has learned, but has never set foot in the physical world. Unlike a child, LLM’s have no understanding of causality, they can describe a ball falling in exquisite detail but cannot reason about why, or what would happen differently on the moon.

Geospatial has been predominantly focused on correlational not causation. There are exceptions including physics-based simulations, process models, but these have remained narrow, domain-specific, and disconnected from the broader world of decision-making.

LLMs compound this problem rather than solve it. They are the most sophisticated correlational engines ever built, pattern matching across the entire written record of human civilization, but they have no causal structure underneath. As Will discusses, they can describe a flood in extraordinary detail but cannot reason about what caused it or what happens next if you intervene.

LEM’s extend LLM’s. They are trained on physical sensor data, making them able to perceive the world, but crucially they do not reason causally.

Enter world models.

World Models

The next advance in AI will be the release of world models say Yann LeCun – sometimes described as the godfather of AI – and many others. A world model doesn’t just represent the state of the world, it represents how the world works. It encodes not just what happens, but why — and what would happen differently if you intervened

LEMs provide the perceptual foundation world models need. In other words the eyes. The world model is the reasoning architecture that transforms what those eyes see into genuine causal understanding. Without Will’s planetary sensing as input, a world model reasons about abstractions. Without the world model on top, LEMs perceive without understanding.

This is the architectural leap my diagram above captures. The LEM sits at The Leap – the boundary between Horizon One and Horizon Two – precisely because it is the perceptual substrate that makes world models possible. But crossing into Horizon Two requires the additional causal reasoning architecture on top.

Together they cross the threshold from correlation to causation. From seeing to knowing.

So we add one more piece to Will’s essay – the critical causal element of PI.

Causal Planetary Intelligence

Will ends his essay reaching for something he doesn’t quite name. Giving AI planetary sensing is, he argues, an act of embodiment, the same developmental condition that allows a child to build a causal model of the world.

He is right. But he stops exactly where the most important question begins.

What happens when you close that loop at planetary scale? That is what I call Causal Planetary Intelligence or Causal PI.

Causal PI is the completion of Will’s vision, not a departure from it. Where PI delivers situational awareness – a continuously updated picture of what is happening – Causal PI delivers simulated foresight. The ability to reason about why things are happening, model consequences of interventions before they are made, and surface decision paths most likely to produce the outcomes that matter.

The architecture has three layers.

• The first is Will’s; that is planetary sensing at continuous global scale. The LEM layer. Perception grounded in physical reality.
• The second is the world model layer. These are the domain-specific causal engines that represent not just physical states but the mechanisms connecting them. A wildfire causal engine encodes relationships between fuel load, wind, humidity and topography, not as statistical patterns but as causal structures that can be simulated forward and interrogated. These are not general models. They are deep, domain-specific representations of how particular physical systems actually work.
• The third is the human layer. This is where Causal PI diverges sharply from the autonomous AI narrative. The human doesn’t disappear, we are indeed elevated. At Stage 4 the human is the decision validator. The causal engine runs the scenarios. The human decides with a clarity no dashboard could previously provide. At Stage 5 something more profound happens. The system becomes a second brain in the room, continuously simulating thousands of decision paths, surfacing options, reasoning over consequences, while the human operates as strategic authority. Not replaced. Not merely assisted. Genuinely augmented. This is what I term Augmented Planetary Intelligence.

Returning to Will’s fire chief. In his framing she receives a comprehensive situational picture; where the fire is, how fast it is moving, which communities are at risk. Extraordinary compared to what came before. But it is still fundamentally a better map. She still needs to make the causal leaps herself.

With Causal PI she doesn’t receive a map. She receives simulated decision paths. The causal engine has already run the wind-shift scenario, modelled the corridor failure, ranked evacuation sequences by predicted outcomes. She is not processing information. She is validating decisions the machine has already reasoned through, and making the final call with a quality of understanding no previous system could give her.

This is the elevation Will gestures at. The human moves from processor to strategic authority not because the machine has taken over, but because the machine has done the causal heavy lifting

Geography always promised causal explanation. Geospatial delivered correlation. LLMs compounded it. LEMs broke through to perception. World models introduce causal reasoning. Causal Planetary Intelligence closes the loop — sensing, understanding, simulating, deciding, at the scale of the entire Earth.

That is Horizon Two.

Will is building the sensing architecture. The causal reasoning layer I’ve described here is the natural next problem. I’d welcome the chance to think through it together. I suspect our perspectives are more complementary than either of us have fully mapped.

Matt Sheehan

Matt Sheehan is a geographer and AI strategist focused on one question: why does AI get better at seeing the world without getting better at understanding it? He is currently researching world models and causal reasoning engines as the next frontier of geospatial intelligence.