This paper addresses detecting AI-generated code in out-of-distribution settings by fine-tuning CodeBERT on algorithmic code from C++, Python, and Java. While the model achieves near-perfect performance on training data (F1 = 0.9935), it degrades significantly on unseen languages and domains (F1 = 0.3532). The high recall (0.8789) but low precision (0.2210) indicates over-prediction of machine-generated code. Error analysis reveals three failure modes: domain mismatch, unfamiliar syntax patterns, and insufficient training. Multi-epoch training and domain-specific augmentation are needed to improve OOD generalization.