Anil Kumar Korupoju, a senior surveyor at the Indian Register of Shipping, writes for Splash today, stressing the importance of data integrity. Artificial intelligence (AI) is increasingly influencing areas of vessel operation that were traditionally governed by human judgement and physical verification. Collision awareness tools, predictive maintenance systems, and structural monitoring platforms now contribute to decisions with clear safety implications. From a classification standpoint, that shift demands closer scrutiny. Much of the current discussion around AI focuses on model capability: predictive accuracy, anomaly detection, even the promise of autonomy. Far less attention is paid to the integrity of the data feeding these systems. In practical terms, the most serious weaknesses in maritime AI will not stem from flawed algorithms. They will stem from degraded, inconsistent, or poorly governed data. At sea, data is rarely pristine. Sensors drift over time. Calibrations shift. Automatic Identification System (AIS) signals can be incomplete in congested waters and vulnerable to manipulation in certain regions. Global Positioning System (GPS) quality fluctuates. Time alignment across systems is not always exact. These are not exceptional circumstances; they are routine operating conditions. An AI system has no inherent awareness that a sensor has moved out of calibration unless controls are designed to detect that change. It will continue to generate outputs that appear coherent and technically sound. The difficulty is that those outputs may gradually diverge from physical reality without triggering an obvious fault condition. Mechanical defects tend to reveal themselves clearly. Data degradation, by contrast, is often subtle and progressive. As AI moves closer to navigational decision support, that distinction becomes increasingly important. While high-performance GPUs have largely eliminated raw computational power as a bottleneck, fully autonomous manoeuvring in congested or ambiguous traffic is less constrained by computational power and more by complexity: dynamic and unpredictable patterns, rare scenarios and edge cases, conflicting inputs and the need for contextual judgement. Without structured revalidation and clearly defined operating limits, performance can drift outside the tested envelope without immediate visibility. From a classification perspective, reliance on a single data source is particularly concerning. AIS, for example, is valuable but not infallible. Cross-verification through independent sources has long underpinned safe navigation. AI systems that do not embed that discipline introduce predictable vulnerabilities. When digital tools influence navigational awareness or structural assessment, data governance becomes a safety matter, not a technical detail. That requires documented data lineage, clearly defined limits of use (the intended Operating Design Domain, or ODD), traceable version control, and rollback to previously validated configurations. It also requires continuous monitoring to detect performance drift and clear accountability for system updates. These are safeguards against gradual degradation, not responses to visible failure. For owners, the implications extend beyond technical reliability. If an AI system contributes to a navigational or maintenance decision, accountability does not sit with the algorithm. It sits with the operator. In the event of an incident, questions will focus on data in