# Data Leakage

> Source: https://sukruyusufkaya.com/en/glossary/veri-sizintisi
> Updated: 2026-05-13T21:08:50.781Z
> Type: glossary
> Category: yapay-zeka-temelleri
**TLDR:** A situation where model performance appears misleadingly strong because it has learned information during training that would not be available in real use.

<p>Data leakage is one of the most overlooked yet dangerous mistakes in AI projects. Put simply, the model learns information during training that it would not actually have access to during real-world use. As a result, measured performance becomes artificially inflated and the system appears much stronger than it really is. Leakage may come from poor train-test separation, giving the model variables too close to the target, or mishandling time-based data. A project with leakage can look brilliant technically but fail quickly once deployed. That is why good evaluation is not only about generating metrics, but also about ensuring those metrics are honest.</p>