At 09:04 UTC on 23 September 1999, the Mars Climate Orbiter's carrier signal stopped. The probe had spent nine months travelling to Mars and was meant to enter orbit at 226 kilometres above the surface. It passed at 57. The Martian atmosphere did the rest. NASA's investigation board took seven weeks to find the cause, and it was not in space: a piece of ground software reported the small thruster impulses in pound-force seconds when the interface specification called for newton-seconds. A factor of 4.45. Nobody caught it in nine months of navigation. The orbital model was flawless. The input data was not.
It is worth retelling that story now, when every week brings a new model promising to read the landscape without anyone setting foot on it. The lesson of 1999 has not expired. It has changed scale.
The 80% myth and what the surveys actually say
You will have heard the figure: data scientists spend 80% of their time cleaning data. It is one of those statistics that circulate without an owner. Leigh Dodds of the Open Data Institute followed the citation trail in 2020 and arrived at a 2016 CrowdFlower survey that Forbes made famous: 60% of time went on cleaning and organising data, another 19% on collecting it. The 80% is the sum. Later surveys come in lower. Anaconda, with nearly 2,400 responses from more than a hundred countries, put data preparation at around 45% of the working day in 2020, with cleaning as the single heaviest task.
Whether it is 45 or 80 hardly matters. What matters is that, in a discipline sold on its models, half the job is still fixing what arrives. And the Forbes survey contained another figure almost nobody quotes: for 57% of respondents, cleaning data was also the least enjoyable part of the work. The task that most determines the outcome is the one nobody wants to do.
The idea is not new either. Garbage in, garbage out first appeared in print in an Indiana newspaper on 10 November 1957, in the words of a US Army computer specialist working on the earliest machines. Seventy years on it remains the most repeated and least applied phrase in computing.
What geospatial AI has genuinely changed
This is not scepticism about the technology. The leap is real. In July 2025 Google DeepMind released AlphaEarth Foundations, a model that fuses optical imagery, radar, LiDAR and climate simulations to describe every 10-by-10-metre cell on the planet with a 64-dimensional vector. The company reports roughly 24% lower average error than other systems across 15 evaluation datasets, and more than 50 organisations were testing it before launch. Esri, meanwhile, ships a library of over a hundred pretrained models for ArcGIS: building footprints, land cover, tree segmentation, vehicle detection. You can use them without labelling a thing.
Notice one detail in Esri's documentation. The models come ready to run, yet the company itself recommends retraining them for "your location, landscape, resolution or sensor type". That is an honest admission that a general model does not know your data. And AlphaEarth, for all its ambition, reduces the world to ten-metre cells. Useful for agricultural policy or deforestation tracking. Not enough to tell you whether an embankment has shifted three centimetres or whether a power conductor is breaching its clearance over a track.
What has changed, then, is the cost of interpretation. Classifying a point cloud or segmenting an orthophoto, which ten years ago meant weeks of manual work, is now a matter of hours if the data is good. The bottleneck has moved upstream, towards the source.
Where the error is manufactured
In geospatial work, bad data is rarely an absurd value that leaps off the screen. It is a quiet, systematic error, like the one that killed the Mars Climate Orbiter. A LiDAR sensor with a miscalibrated mounting angle produces a point cloud that is perfectly dense and perfectly displaced. A flight with insufficient overlap leaves gaps that the software interpolates with great elegance. A thermal camera without radiometric correction delivers striking images and false temperatures. A ground control point measured in a hurry shifts the whole georeferencing by a few centimetres nobody will notice until two surveys fail to match.
No AI model corrects any of that. On the contrary: it learns it. A classifier trained on misaligned clouds will learn that power lines are twice as thick and that slopes have steps. And it will do so with great confidence, because a model's confidence does not measure input quality; it measures consistency with what it has already seen. The better the model, the more convincing the error.
That is why NASA's board did not stop at the units file. It pointed to what surrounded it: an overloaded navigation team, no independent review, and signs that something was off which were written down as noise. The pattern is familiar. When an organisation trusts the model, it stops looking at the data.
Capturing well is the expensive part, and the part that counts
Here is the thesis, stated plainly: in geospatial engineering, the value and the risk still sit in the capture. In a well-planned flight, with the right sensor for the question being asked. In calibration, which is not a laboratory formality but a decision that conditions every subsequent measurement. In point-cloud classification, where a mass of points becomes ground, vegetation, building or conductor, and where a poorly defined criterion contaminates everything downstream.
Three phases, really. Capture: sensor, platform and flight plan. Structure: align, calibrate, classify, turn a measurement into data you can work with. Interpret: the model, the analysis, the decision. AI has made the third phase dramatically cheaper. It has not touched the first, and the second only works if the first was done properly. Anyone pouring all their attention into phase three is optimising 20% of the chain and inheriting the other 80% from someone they have never met.
This has a practical consequence for anyone commissioning work. The question is not which model the supplier uses. It is which sensor they flew, how they calibrated it, what overlap the mission had, how many control points they measured and by what criteria they classified. These are dull questions. They are the only ones that matter.
What we do about it
At Azimutal we have spent three decades measuring territory, and some time ago we made a distinctly unglamorous decision: to control all three phases. We fly our own sensors, airborne and terrestrial LiDAR, RGB, multispectral and thermal, and we calibrate them ourselves, because a borrowed sensor is data whose origin we cannot defend. We classify point clouds with our own software, because classification is where the criterion is set, and the criterion is not something you outsource. Only then do we interpret: orthophotos, mapping from 1:500 to 1:50,000, network models.
That does not make us immune to error. It makes us accountable for it, which is different. When an orthophoto of a wind farm or an inventory of power lines does not add up, we know which phase to look at and we have the trail to do it. More than 300 wind farms, more than 100 solar plants and more than 10,000 kilometres of power lines captured have taught us something simple: the models get better every year, and every year the source data is still what decides whether the result is any use.
AI will keep surprising us. But it will not bring back the Mars Climate Orbiter, and it will not straighten a badly flown survey. A model can only ever be as good as the data that someone, somewhere, decided to capture with care.
Sources
- Mars Climate Orbiter Mishap Investigation Board, Phase I Report — NASA, 1999
- Mars Climate Orbiter Failure Board Releases Report — NASA press release via SpaceNews, 1999
- Cleaning Big Data: Most Time-Consuming, Least Enjoyable Data Science Task, Survey Says — Forbes / CrowdFlower, 2016
- Data Prep Still Dominates Data Scientists' Time, Survey Finds — Datanami / Anaconda State of Data Science, 2020
- Do data scientists spend 80% of their time cleaning data? Turns out, no? — Leigh Dodds, 2020
- Is This the First Time Anyone Printed 'Garbage In, Garbage Out'? — Atlas Obscura, 2017
- Google DeepMind unveils an AI model that acts as a virtual satellite for mapping the entire planet — The Decoder, 2025
- Pretrained Deep Learning Models for ArcGIS — Esri, 2026



