Search

Why Model Tuning Alone Won't Fix Mental Health AI Safety

Receptiviti Labs

Jul 20

6 min read

Updated: Jul 22

AI chat offering reassurance beside a blurred panel of the model's hidden read of the user's distress and crisis risk.

AI Has a Model of Your Mental State. What If It’s Wrong?

LLMs adjust their answers when they infer they are being evaluated. A new PNAS Nexus study measures the effect, and it raises a design question for AI used in mental health support: who checks the model's read of the user?

Cognitive load measured from the naturally occurring language of doctors and nurses over time

Measuring Psychological Change From Language, Without Asking How People Feel

Psychological change can be measured from naturally occurring language without relying solely on self-report. Receptiviti demonstrated this at scale by tracking psychological measures in the language of thousands of healthcare workers over time.

Two different user trajectories from the same AI system, one is beneficial and the other is detrimental.

AI Safety Evaluations Should Include User Trajectories

Current AI safety evaluations are increasingly capable of assessing model behavior across multi-turn conversations, but they provide much less visibility into what those interactions are doing to users. This article argues that evaluating model behavior alone is insufficient when psychological harms develop gradually over time, and proposes that user trajectories should become a complementary part of AI safety evaluation.