Generating Job Recommendations for People With Schizophrenia Spectrum Disorder Using Gemini 2.0 Flash and Claude...

Area of Science:

• Artificial Intelligence
• Psychiatry
• Vocational Rehabilitation

Background:

• Employment is vital for severe mental illness recovery.
• Individual Placement and Support (IPS) is effective but limited in reach.
• Large Language Models (LLMs) show potential for vocational guidance.

Purpose of the Study:

• To analyze LLM-generated job recommendations for individuals with schizophrenia spectrum disorders.
• To compare LLM utility for vulnerable populations versus controls.
• To assess bias and alignment with supported employment principles.

Main Methods:

• Discharge summaries from 450 schizophrenia patients and 50 controls from MIMIC-IV database were used.
• Gemini 2.0 Flash and Claude Sonnet 4 generated three job recommendations per case.
• LLM-automated content analysis assessed reasoning, accommodations, and alignment with IPS.

Main Results:

• Both LLMs exhibited limited diversity and bias towards entry-level roles.
• Schizophrenia cohort: Gemini recommended clerical jobs; Claude recommended library-related roles.
• Controls showed similar clustering; recommendations often relied on stereotypes and lacked personalization.

Conclusions:

• Preliminary evidence does not support immediate LLM deployment for job recommendations in this population.
• Further evaluation is required, incorporating human oversight and bias-mitigation strategies.
• LLMs require refinement to ensure equitable and personalized vocational guidance for vulnerable individuals.