Soil is not like concrete. You can specify concrete. You can test it in a controlled environment. Ground is what it is, and it varies in ways that even the most thorough investigation only partially reveals.
AI tools are entering geotechnical practice fast. Machine learning models now predict bearing capacity, flag liquefaction risk, interpolate between borehole data points, and automate parts of site characterisation that used to take weeks. A landmark study from the Laing O'Rourke Centre at Cambridge describes AI as a potential game-changer in geotechnics, with use cases spanning intelligent site investigation, predictive soil behaviour modelling, and design optimisation.
All of that is real. So is the risk when it goes wrong.
The Problem with Training Data
Every machine learning model learns from the data it was trained on. In geotechnical engineering, that means historical site investigation records, borehole logs, lab tests, and field measurements from completed projects. The datasets are often sparse, inconsistent, and geographically biased toward certain regions or soil types.
A paper in SIAM News makes this concrete. Standard tree-based ML models suffer up to 30% accuracy degradation when trained on datasets with fewer than 10,000 samples, which is a common reality in site-specific geotechnical work. Site investigations typically yield fewer than 1,000 data points per project. The model is being asked to generalise from conditions it may never have seen.
This matters enormously in the UK, where geological variability is extreme across relatively short distances.
UK-Specific Ground Risks AI Can Struggle With
Clay shrinkage and heave. According to the British Geological Survey, shrink-swell in clay soils costs the UK economy over £400 million a year, a figure predicted to rise above £600 million by 2050 as climate change intensifies. Root-induced clay shrinkage is responsible for 60% of all UK subsidence insurance claims. An AI model trained on national datasets can identify high-risk postcodes. It cannot tell you about the mature oak tree 12 metres from your proposed building footprint.
Chalk dissolution. Parts of Kent, East Anglia, and the South Downs sit on chalk formations with dissolution features including open voids, infilled cavities, and irregular depths to competent rock. These features are inherently site-specific. Dynamic probing, a common investigation method, can exaggerate apparent risks or miss real ones. AI models interpolating between investigation points are flying blind between boreholes.
Contaminated land. Urban brownfield sites in the UK frequently carry historical contamination from former industrial uses. The spatial distribution of contamination does not follow predictable patterns. A model trained on clean greenfield data will not flag what it has never learned to look for.
Ground Risk in the UK:
£400m annual cost of clay shrink-swell subsidence in the UK (BGS)
20-60% of linear and transport infrastructure projects delayed by unforeseen ground conditions (Gov.uk)
30% accuracy degradation in ML models with fewer than 10,000 training samples (SIAM)
What AI Cannot Replace
The TRC Companies assessment of AI in geotechnical practice makes the right point: rather than replacing engineering judgment, these tools help engineers reveal patterns more clearly, quantify uncertainty more rigorously, and make better-informed decisions. The operative phrase is rather than replacing engineering judgment.
A geotechnical engineer brings something no model currently replicates. They walk the site. They notice the pond that only forms in wet winters. They spot the Victorian ordnance map reference to a former brickworks. They recognise the clay exposure in a field drain that changes the risk profile entirely.
AI tools in geotechnics are at their best as pattern recognition tools over large datasets, uncertainty quantification aids, and productivity tools for routine report writing and data processing. They are not ready to make site-specific ground risk judgements independently.
"The model does not know about the Victorian brickworks. The engineer who walked the site does."
What Good Practice Looks Like
The UK government's Future of the Subsurface report notes that 20 to 60% of linear and transport infrastructure projects experience delays due to unforeseen ground conditions. That failure rate predates AI adoption. Introducing AI without robust human oversight will not fix it.
Good practice means using AI outputs as one input among many, not as a conclusion. It means documenting the limitations of any model used, the training data it was built on, and the local geological conditions it may not have encountered. It means a named, qualified engineer signing off the ground model, regardless of how much AI was used to build it.
The ground does not care about your model's accuracy metrics. It will behave how it behaves. The professionals responsible for understanding that bear the liability regardless of what the algorithm said.
AI Competence for Geotechnical and Civil Professionals
Understanding where AI adds value in ground investigations, and where professional judgement is irreplaceable, is core to responsible practice. Explore our built environment AI resources at responsiblewithai.com.
