Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
Plausible Patients, Impossible Populations: Auditing Epidemiological Fidelity in Large Language Model Mental Health Simulations
TL;DR
Large language models like GPT-4o-mini and Gemini-3-Flash can generate plausible individual psychiatric cases but fail to accurately represent real population demographics in mental health simulations. This discrepancy highlights a significant gap between case-level validity and population-level fidelity, termed coherence-fidelity dissociation.
✦ Why It Matters
Evaluate AI-generated mental health data critically to ensure it aligns with real-world demographics.
Key Takeaways
- Assess AI-generated health data against real-world benchmarks.
- Recognize the limitations of individual case validity in population simulations.
- Investigate demographic disparities in AI outputs critically.
- Consider recalibration methods for AI-generated mental health data.
How It Works
The study involved prompting four language models to simulate psychiatric patients based on two different framing scenarios: clinician input and self-description. Each model's output was then scored against the PHQ-8, which measures mental health symptoms, allowing for a quantitative assessment of their fidelity to real-world data.
⚠ The Catch
Despite high individual case validity, the models produced population-level data that was significantly skewed, failing to accurately reflect real demographic distributions and mental health disparities.