Overfitting remains a central challenge in modern data science, particularly as complex analytical tools become more accessible and widely applied in fields like chemometrics. This communication outlines a series of common pitfalls that lead to misleading and non-generalizable models-ranging from poor data quality and insufficient sample sizes to misuse of validation strategies and overly complex modeling choices. By illustrating a caricatured protocol for generating bad models, the paper emphasizes the importance of domain knowledge, appropriate experimental design, and rigorous validation. It advocates for "validity by design" as a proactive strategy to ensure robust, interpretable, and scientifically sound results.