TrustMI Maps Controllable Trust Directions

TrustMI’s most compelling result is controllability, not proof that models possess a distinct internal “trust” faculty.

• –Frozen model weights make the method useful for probing safety-relevant behavior without retraining.
• –The study operationalizes trust as reliance on user claims, which may overlap with uncertainty, sycophancy, or refusal behavior.
• –Lowering trust could reduce prompt-injection risk, but excessive distrust may hurt helpfulness and increase false refusals.
• –Real-world tool traces and human evaluations will be needed to establish whether the direction generalizes beyond synthetic contrastive conversations.