Generalizability is the ability of a model to keep working when conditions change. In AI systems, that means the model still performs well across different patients, organizations, languages, devices, or documentation styles instead of only matching the data it was trained on. A model with poor generalizability may look accurate in testing but fail in real use because the real-world input is messier, noisier, or distributed differently.
In cyber security, generalizability matters for phishing filters, malware detectors, fraud models, and anomaly detection tools. Attackers often try to evade detection by changing wording, file structure, traffic patterns, or infrastructure, so a brittle model can be bypassed easily. Defenders improve generalizability by training on diverse data, testing against shifted conditions, monitoring live performance, and retraining when behavior changes. In healthcare AI, the same idea affects whether a system works across patient groups and charting practices, which is why it is tied to safety and bias.