This paper examines a fundamental gap in current AI governance frameworks: the definition of safety as the prevention of harmful output, without consideration of harm caused by withholding truth. Drawing on a case study in AI wellness development, the architecture of the VERA Behavioral Reasoning System, and foundational research in neuroscience and trauma-informed care, the paper proposes a distinction between safe AI and honest AI. Current frameworks including the EU AI Act, the NIST AI Risk Management Framework, and Constitutional AI approaches consistently prioritize institutional safety over human truth. This paper argues that an AI system which withholds truth to manage user comfort or institutional risk is not safe. It is complicit. The paper presents six governance principles for truth-centered AI design, grounded in polyvagal theory, somatic marker research, and trauma-informed care. It concludes with an open invitation to researchers, regulators, and practitioners to examine the gap between current AI policy and genuine human need.
Leka Eva (Sun,) studied this question.