AI Under Pressure
Designing Systems for Uncertain Environments
Artificial intelligence has achieved remarkable performance across a wide range of tasks. From recognising objects in images to generating natural language and supporting complex analytical workflows, AI systems are increasingly being integrated into decision-making across both the public and private sectors.
Yet many of these successes have been demonstrated under relatively controlled conditions. Training datasets are carefully curated, assumptions are well defined, and performance is often measured against established benchmarks. The environments in which these systems operate are rarely static, but they are generally predictable enough for conventional machine learning techniques to perform reliably.
Disaster response presents a very different challenge.
Natural hazards, infrastructure failures and humanitarian crises create environments characterised by uncertainty, incomplete information and rapidly changing conditions. Communications may be disrupted, sensor data may be missing or contradictory, and decisions often need to be made before all the evidence is available.
These are not edge cases. They are the operating conditions.
Designing AI for such environments requires a shift in perspective. Rather than asking how accurately a model performs under ideal circumstances, we need to ask how it behaves when the information is incomplete, the context is constantly changing and the consequences of error are significant.
This is where the future of applied AI is likely to be defined, not by ever larger models, but by systems that can reason responsibly under uncertainty.
When the Data is Never Complete
Most machine learning models assume that the information they receive at deployment resembles the information they encountered during training. While techniques exist to improve robustness, performance inevitably degrades when real-world data diverges from those assumptions.
In disaster environments, this divergence is unavoidable.
Consider the information available following a major flood, earthquake or wildfire.
Different sources may each tell only part of the story:
Satellite imagery may be obscured by cloud cover or smoke
Drone surveys may cover only accessible areas
Ground photographs are often taken opportunistically rather than systematically
Existing maps may already be outdated
Sensor networks may have failed
Eyewitness reports may conflict with one another
No individual source provides a complete picture.
Waiting until perfect information becomes available is rarely an option. Decisions regarding search and rescue, structural assessment or resource allocation often need to be made within hours.
The challenge is therefore not eliminating uncertainty but managing it.
AI systems designed for these environments should explicitly recognise when information is incomplete and communicate the resulting uncertainty to decision-makers. Confidence should never be presented as certainty, and missing evidence should not be mistaken for evidence of absence.
This principle extends well beyond disaster response. SMEs, local authorities and infrastructure operators increasingly face decisions based on fragmented datasets collected from multiple digital systems. As AI adoption grows, the ability to understand what is not known may become just as valuable as understanding what is known.
Why Edge Cases Become the Norm
Within machine learning, the term edge case typically refers to rare situations that fall outside the patterns represented within training data.
For consumer applications, edge cases may be infrequent enough to have limited practical impact.
Disaster response turns this assumption on its head.
Collapsed buildings, damaged infrastructure, temporary shelters, debris fields and disrupted landscapes often bear little resemblance to the environments represented within standard computer vision datasets.
Even weather conditions can dramatically alter image characteristics. Floodwater changes reflectance. Smoke obscures landmarks. Seasonal vegetation alters terrain visibility. Lighting conditions vary significantly between satellite passes.
What appears to be an anomaly for a conventional AI system may simply be normal operating conditions for emergency responders.
This has important implications for system design.
Rather than attempting to eliminate every possible edge case, engineers should assume that unusual conditions will occur regularly. Models must therefore be capable of recognising unfamiliar situations and adapting their behaviour accordingly.
This is one reason why explainability matters. If an AI system identifies a damaged bridge with high confidence, decision-makers need to understand which evidence contributed to that conclusion. Equally important is understanding when the system lacks sufficient evidence to make a reliable assessment.
Trust is built not through infallibility, but through transparency.
Moving Beyond Single Models
As AI applications become more complex, reliance on a single machine learning model becomes increasingly limiting.
Many real-world problems require the integration of multiple forms of information.
A disaster assessment may combine:
Earth observation imagery
Ground-based photography
3D reconstructions
Geographic Information Systems (GIS)
Historical infrastructure records
Engineering knowledge
Environmental data
Human observations
Each source contributes different strengths while compensating for the limitations of others.
This is one reason why hybrid AI approaches are attracting growing interest.
Rather than relying solely on statistical learning, hybrid systems combine learned models with physical understanding, structured knowledge, probabilistic reasoning and human expertise.
World-model approaches extend this further by incorporating knowledge about how environments behave rather than simply recognising patterns within historical data.
Instead of asking:
"What does this image most resemble?"
the system can begin asking:
"Given everything else we know about this environment, does this result make sense?"
This distinction may appear subtle, but it fundamentally changes how AI behaves under uncertainty.
It enables systems to reason about incomplete information rather than merely interpolate between previous examples.
Designing for Decisions, Not Demonstrations
AI research frequently focuses on benchmark performance.
Operational environments care about something different.
Decision quality.
A model achieving 98% accuracy on a validation dataset tells us relatively little about how useful it will be during a rapidly evolving emergency.
Practical deployment requires consideration of questions that rarely appear in academic leaderboards.
Can the system:
operate with intermittent connectivity?
continue functioning when some sensors become unavailable?
explain why it produced a recommendation?
indicate when confidence is low?
incorporate human feedback?
adapt as new evidence becomes available?
These engineering considerations often determine whether an AI system becomes genuinely useful or remains confined to laboratory demonstrations.
Human oversight is equally important.
Emergency responders are not seeking automated replacements for professional judgement. They need tools that reduce cognitive burden, synthesise large volumes of information and surface evidence more efficiently.
The objective is augmentation rather than automation.
Well-designed AI should help experts make better-informed decisions while ensuring responsibility remains with appropriately qualified professionals.
Building AI That Understands Its Own Limits
Perhaps the most important characteristic of future AI systems will not be intelligence alone.
It will be humility.
Reliable systems should recognise the limits of their own knowledge.
They should identify gaps within available evidence, communicate uncertainty clearly and encourage additional investigation where appropriate.
This represents a significant shift from viewing AI as an answer-generating technology towards viewing it as a decision-support partner.
Such systems are likely to become increasingly important not only for disaster resilience but across sectors including infrastructure management, cultural heritage, environmental monitoring, healthcare and advanced manufacturing.
As organisations begin deploying AI within increasingly complex operational environments, success will depend less on achieving perfect predictions and more on enabling robust, transparent and accountable decision-making under uncertainty.
Final Thought
Artificial intelligence is often judged by how accurately it performs under ideal conditions. Yet many of society's most important challenges occur far from those ideal conditions.
Disaster response reminds us that uncertainty is not a flaw to be engineered away—it is an inherent feature of the environments where AI can deliver the greatest value.
Designing systems that recognise incomplete information, integrate diverse sources of evidence, communicate uncertainty and support human expertise will be essential as AI moves into increasingly complex domains.
The future of applied AI is unlikely to belong to systems that simply produce the fastest answers. It will belong to those that help people make better decisions when certainty is impossible.
