CONF Watawana_LREC2026_2026/IDIAP When Consistency Becomes Bias: Interviewer Effects in Semi-Structured Clinical Interviews Watawana, Hasindri Sankalpana Burdisso, Sergio Moreno-Galvan, Diego Aaron Sánchez-Vega, Fernando López-Monroy, A. Pastor Motlicek, Petr Villatoro-Tello, Esaú EXTERNAL https://publications.idiap.ch/attachments/papers/2026/Watawana_LREC2026_2026.pdf PUBLIC LREC 2026 Palma de Mallorca, Spain Proceedings of the Fifteenth Language Resources and Evaluation Conference 2026 ELRA Language Resource Association https://aclanthology.org/2026.lrec-1.185/ URL 10.63317/34hw23mzd8c7 doi Automatic depression detection from doctor–patient conversations has gained momentum thanks to the availability of public corpora and advances in language modeling. However, interpretability remains limited: strong performance is often reported without revealing what drives predictions. We analyze three datasets—ANDROIDS, DAIC-WOZ, and E-DAIC—and identify a systematic bias from interviewer prompts in semi-structured interviews. Models trained on interviewer turns exploit fixed prompts and positions to distinguish depressed from control subjects, often achieving high classification scores without using participant language. Restricting models to participant utterances distributes decision evidence more broadly and reflects genuine linguistic cues. While semi-structured protocols ensure consistency, including interviewer prompts inflates performance by leveraging script artifacts. Our results highlight a cross-dataset, architecture-agnostic bias and emphasize the need for analyses that localize decision evidence by time and speaker to ensure models learn from participants’ language. Our code is available at: \url{https://github.com/idiap/bias_in_daic-woz/tree/main/LREC_2026}.