Hidden bias in depression detection: When interviewer scripts fool AI models

Advances in language models have made it possible to automatically detect depression from doctor-patient conversations, but a new analysis reveals a troubling gap between reported performance and actual interpretability. Researchers examining three clinical interview datasets—ANDROIDS, DAIC-WOZ, and E-DAIC—have uncovered a systematic bias that lets models achieve high classification accuracy while sidestepping genuine patient language.

The problem with consistency#

Semi-structured clinical interviews are designed to maintain consistency across sessions by using standardized prompts and questioning sequences. However, this very consistency creates an unintended loophole. Models trained on the full conversation—including interviewer turns—learn to exploit these fixed prompts and their positioning in the dialogue. By recognizing which scripted questions appear at which points, the models can distinguish depressed from non-depressed participants at rates that look impressive on paper, often without actually analyzing what the patient says.

What changes when you remove the script#

When researchers restricted models to only participant utterances, two key things happened. First, the models no longer had access to the reliable cues embedded in interviewer language. Second, the decision-making process became more distributed across the conversation, reflecting a broader range of linguistic signals from the patient rather than narrow dependence on interview structure.

This finding is significant because it suggests that strong reported scores may not indicate genuine diagnostic reasoning—they may instead reflect the model's ability to memorize and exploit the artifacts of how semi-structured protocols work.

Why this matters for real deployment#

The research highlights a bias that persists across different datasets and neural architectures, indicating a deep structural problem rather than an isolated flaw. For depression detection systems to be reliable in clinical settings, they need to learn from actual patient language, not from the scaffolding of the interview itself.

The authors emphasize the importance of analysis methods that can localize where models place their decision evidence—tracking which speaker contributes what, and when. Without such scrutiny, models can achieve impressive benchmark results while remaining brittle and non-interpretable, potentially failing when deployed in real-world scenarios with different interview styles or structures.