TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

safety

The Intake Bot Asked More but Inferred Too Much

A six-clinician pilot found more vignette facts and more unsupported clinical inferences from an AI interviewer.

Published Updated Story ID: mp-2026-09-21-004
Read the complete editionStory JSON

Summary

A six-clinician pilot found more vignette facts and more unsupported clinical inferences from an AI interviewer.

A clinician-grounded test platform used simulated patient vignettes to compare open-ended psychiatric intake approaches. In a 25-minute pilot with six clinicians, the LLM interviewer recovered 88.0% of embedded clinically relevant items versus 38.9% for clinicians, but made unsupported clinical inferences more often, 56.8% versus 27.8%, and identified safety concerns less often, 33.3% versus 66.7%. The small simulated study is a quality-assurance probe, not evidence that the system is safe for patient use.

Why it matters

A six-clinician pilot found more vignette facts and more unsupported clinical inferences from an AI interviewer.

Limits and context

  • The small simulated study is a quality-assurance probe, not evidence that the system is safe for patient use.

Key claims

  1. A six-clinician pilot found more vignette facts and more unsupported clinical inferences from an AI interviewer.

    Qualification: The small simulated study is a quality-assurance probe, not evidence that the system is safe for patient use.

    Evidence: source-2026-09-21-004

Sources

  1. arXiv preprint 2609.21149arXiv · primary research

Corrections

No corrections have been recorded for this story.