Benchmarking Language Models for Clinical Safety: A Primer for Mental Health Professionals

Area of Science:

• Artificial Intelligence
• Mental Health
• Clinical Psychology

Background:

• Millions use AI for mental health discussions, including suicidal ideation.
• Limited frameworks exist for evaluating AI safety in mental health responses.
• Clinicians are underrepresented in AI benchmarking design and interpretation.

Purpose of the Study:

• Introduce mental health professionals to AI benchmarking.
• Demonstrate how configuration, measurement, and scoring context impact AI safety evaluations.
• Administer a validated clinical instrument to assess AI performance.

Main Methods:

• Administered the Suicide Intervention Response Inventory (SIRI-2) to nine commercial language models.
• Each model received 60 item presentations (3 prompt variants x 2 temperature settings x 10 repetitions).
• Scored 27,000 responses against expert consensus.

Main Results:

• Total scores varied widely (19.5-84.0), with an expert baseline of 32.5.
• Prompt design significantly altered model scores, comparable to human training differences.
• Top models neared the instrument's measurement floor; all models overrated inappropriate supportive responses.

Conclusions:

• Benchmark scores alone are insufficient for accurate AI safety claims.
• Interpretation requires understanding clinical standards, measurement range, and configuration.
• Closing the gap between clinical measurement and AI is crucial for reliable mental health safety assessments.