Why the North Warning System No Longer Fits the Threat
The North Warning System (NWS) was constructed through the late 1980s and into the mid-1990s as the successor to the Distant Early Warning (DEW) Line. It consists of 15 long-range radar (LRR) stations equipped with AN/FPS-117 radars — 11 in Canada and 4 in Alaska — along with approximately 36 short-range radar (SRR) stations in Canada and additional SRR sites in Alaska. The system provides overlapping radar coverage from the Beaufort Sea through Baffin Island and into Labrador, feeding data through NORAD's continental air defense network and ultimately to Cheyenne Mountain.
The NWS was designed around a specific threat model: Soviet strategic bombers flying predictable subsonic flight paths toward North American targets. The AN/FPS-117 is effective at that task. It is not optimized for detecting hypersonic glide vehicles (HGVs), which fly non-ballistic trajectories at altitudes and speeds that generate different radar signatures than aircraft or ballistic missiles. Nor was it designed to detect low-observable cruise missiles — particularly sub-launched variants that approach from offshore Arctic patrol areas rather than over the radar's detection horizon.
The fundamental mismatch is geometric and spectral. The NWS stations were positioned for their DEW Line successor role, with spacing and look angles appropriate for intercepting aircraft. A hypersonic weapon that descends into the coverage envelope from high altitude, or a submarine-launched cruise missile that stays below the radar horizon until close approach, may appear as a very short track with limited time for human or automated assessment before it reaches its target.
What Canada's Commitment Covers
In June 2022, Canadian Defence Minister Anita Anand announced that Canada would invest C$38.6 billion over approximately 20 years in NORAD modernization. The investment spans four general areas: Arctic and continental surveillance, space-based surveillance, command and control infrastructure, and support infrastructure to sustain operations in the Arctic environment.
The surveillance component includes over-the-horizon radar (OTH-R), which uses skywave propagation to detect objects beyond line-of-sight range — relevant for detecting aircraft and some missile launches far earlier than ground-based line-of-sight radar allows. Space-based sensing integration is also part of the program, leveraging Space Force surveillance satellites alongside emerging commercial space situational awareness providers. Maritime domain awareness capabilities address the submarine-launched cruise missile risk from Arctic approaches.
This is a 20-year commitment across multiple acquisition phases, not a single program. Each phase represents a distinct program of record with its own requirements document, solicitation, and T&E plan. The integration challenge is not confined to any single radar program — it spans the system-of-systems architecture that must make sense of data from all these sources simultaneously.
The Multi-Sensor Fusion Requirement
The technical challenge at the center of NORAD modernization is not sensor performance in isolation — it is the data architecture that turns feeds from multiple heterogeneous sensor types into reliable, timely threat assessments.
Each sensor type in the modernized architecture produces data with different characteristics. A ground-based radar provides high-update-rate tracking data with strong geometric accuracy but is limited by line of sight. A space-based infrared sensor detects launch signatures from orbit but has different update rates and coverage geometry. An OTH-R system provides extended range but with lower resolution and more environmental variability. A maritime radar or acoustic sensor may detect a submarine departure or missile launch from waters that ground radar cannot cover at all.
Track correlation — the process of associating multiple sensor observations with a single physical object — is conceptually straightforward but operationally complex at scale. When the system is tracking hundreds of commercial aircraft, military aircraft on known flight plans, and satellite objects simultaneously, the algorithm must correctly associate a radar track from Canada with a space sensor track from a different orbit and a maritime contact from a different sensor entirely, in real time, across the full surveillance zone. Errors in track association — either incorrectly merging two separate objects or failing to connect observations of the same object — directly degrade the quality of the threat picture.
Classification is the second problem. Once a track is established, the system must assess what the object is: commercial aircraft, friendly military aircraft, neutral state aircraft, cruise missile, ballistic missile, or hypersonic glide vehicle. In the Arctic, commercial air traffic on polar routes and military traffic on exercise patterns coexist with the threat environment. Classification depends on combining radar signature (shape, radar cross-section, altitude, speed), trajectory analysis, and contextual data including known flight plans and patterns of life.
Where AI Has a Defensible Role
The scale of the multi-sensor fusion problem creates a legitimate case for AI-enabled processing. Human operators cannot manually correlate tracks from dozens of sensor feeds across a surveillance zone spanning thousands of kilometers. The question is not whether to use automated processing — the NWS already uses automated track management — but what AI-enabled capabilities are appropriate and what requirements they must meet.
AI has a specific, defensible role in three areas of this architecture.
First, track correlation at scale: machine learning models trained on historical sensor data can associate observations from heterogeneous sensors more accurately than rule-based correlation algorithms when the sensor geometry is complex. The specific requirement is that the model's association decisions are interpretable — operators and program managers need to understand why the system merged two tracks or kept them separate, not just accept an opaque output.
Second, anomaly flagging: a track that does not match any known flight plan, exhibits an unusual trajectory, or generates an ambiguous classification should be flagged for human attention rather than silently classified as benign or dropped. This is a case where AI adds value by identifying uncertainty rather than resolving it — the system surfaces the track with confidence indicators and supporting evidence for a human assessor.
Third, signature analysis support: classification models trained on known radar signatures for threat classes can generate probability distributions over candidate classifications, giving human operators a structured basis for assessment rather than a binary answer. This is most useful in ambiguous cases where no single sensor provides definitive classification.
The appropriate boundary: AI should not make autonomous threat designation decisions. Under DoD Directive 3000.09 (updated 2023), human judgment is required for lethal force decisions. In the context of early warning, the decision to initiate an alert chain that could lead to an engagement recommendation involves enough consequential ambiguity that human review must remain in the loop before escalation. The system design question is where exactly that human review checkpoint goes — and requirements documents for NORAD C2 modernization need to specify it explicitly rather than leaving it to software developers.
The Asymmetric Cost of Two Error Types
Any AI-enabled classification system in this context has two failure modes, and their costs are not symmetric.
A false positive — misclassifying a commercial aircraft or friendly military flight as a threat — initiates an alert chain that consumes command attention, can trigger defensive posturing visible to adversaries, and erodes confidence in the system over time if it happens repeatedly. This is costly but not catastrophic.
A false negative — failing to classify an actual threat, or classifying it with low enough confidence that it passes without human review — has a categorically different consequence. At the extreme, a missed warning on a nuclear-armed hypersonic weapon cannot be corrected after the fact.
Requirements documents for the classification component must specify acceptable false positive and false negative rates separately, for each threat class and against realistic edge cases including adversarial evasion tactics. A system optimized only on false positive rate — which is the natural result of training on available data dominated by benign tracks — will systematically underweight threat signatures that look like benign traffic. T&E plans must test both failure modes explicitly, including against synthetic adversarial tracks designed to defeat the classification model.
The Cybersecurity Architecture Problem
The NWS stations are geographically dispersed across one of the world's most logistically challenging environments. Replacing them with networked digital sensors expands the attack surface for a critical national security system. An adversary who can suppress a radar station's data feed — by jamming, cyber intrusion, or physical interference — without triggering a network-level alert creates a detection gap.
The data pipeline from sensor to C2 system must be designed with Zero Trust architecture principles: validate source identity and data integrity at each transfer point rather than trusting the network. This means sensor data should carry cryptographic signatures that the C2 system can verify, and anomalies in data quality or update rate should generate alerts rather than silent failures.
The bilateral Canada-US governance structure adds a classification and data-sharing requirement. Sensor data from Canadian-operated stations feeding a joint C2 system requires clear agreements on handling, access control, and what happens when the two nations have different assessments of a track's classification. These governance requirements need to be architecture requirements in the contract, not left to memoranda of understanding drafted after system delivery.
What Program Offices Need to Specify Before Soliciting
For vendors and integrators responding to NORAD modernization solicitations, the architectural clarity of the requirements document determines whether a proposal can be evaluated or is effectively competing on faith.
Requirements documents need to specify: sensor data interface standards (format, update rate, latency budget) for each sensor type the C2 system will receive; the track correlation algorithm's expected performance metrics and the test methodology for measuring them; the classification confidence threshold design — at what confidence level and for what threat class does the system require human review before escalation; and the cybersecurity architecture baseline, including cryptographic signing requirements for sensor data.
T&E plans need to address false positive and false negative performance separately, against both nominal threat profiles and adversarial edge cases. Integration testing must address the transition between new sensor feeds and the legacy C2 tail, including failure modes where new sensors provide data and legacy systems attempt to correlate it.
The 20-year modernization timeline means that AI classification models will need to be updated as threat characterizations evolve — new missile types, new delivery modes, new tactics. Requirements documents should specify the model update process and re-validation requirements. A system whose classification component can only be updated through a major program modification will fall behind the threat within a few years of delivery.
Sources and further reading
- Government of Canada, NORAD Modernization announcement, June 2022
- NORAD, About NORAD, organizational and mission overview
- U.S. Air Force Air Combat Command, ACC AMIC Det 1 and the North Warning System, DVIDS, August 2025
- DoD Directive 3000.09, Autonomous Weapons Systems, November 2023 update
- Government Accountability Office, Missile Defense: Actions Needed to Improve DOD Reporting on Homeland Defense Capabilities, GAO-22-105047, March 2022
—-
The data integration challenges in NORAD's sensor architecture — multi-source correlation, classification confidence, adversarial edge case handling, and verification of model outputs before operational decisions — are a specific instance of the broader AI verification problem in high-stakes defense systems. Spartan X's work on AI verification and multi-model consensus is oriented precisely toward this requirement: establishing warranted confidence in an AI output when the cost of a wrong answer is asymmetric and the operational consequence is immediate.
