Large language models offer new opportunities for processing historical documents, yet their application raises questions of reliability. We present the first comprehensive and explainability-driven framework for evaluating model design bias in multilingual historical news article extraction, using newspaper coverage of the 1908 Messina earthquake as our test case across German, English, and French sources. Through systematic analysis of six state-of-the-art models, we uncover three critical bias patterns that, in addition to data quality, compromise extraction quality: contextual integration bias, overconfidence bias, and preference bias. Our evaluation reveals that these biases stem from alignment procedures rather than training data limitations, findings that establish methodological foundations for responsible AI deployment in digital humanities.