From Interpretation to Restraint Biometric-Gated AI as a Deterministic Ethical Control Layer for Human–Machine Interaction

Prevailing approaches to AI ethics rely on declarative principles, psychological inference, or post-hoc accountability mechanisms. These paradigms systematically degrade under conditions of human cognitive and physiological instability—precisely where ethical guarantees are most critical. In high-risk and clinical contexts, human volatility amplifies the irreversibility of machine-mediated error, rendering intent inference and affective modeling not merely unreliable, but ethically hazardous.

This work introduces biometric-gated AI restraint, a deterministic control-layer architecture in which non-semantic physiological signals are used exclusively to curtail AI agency, without attributing mental states, intentions, or emotions. We formalize the Human System Volatility Index (HSVI) as a bounded, rule-based scalar derived from heart rate variability, sleep fragmentation, blood pressure variability, and glycemic instability. HSVI modulates AI initiative, complexity, persuasiveness, and reversibility via fixed thresholds and worst-case safety bounds.

We further extend the framework with a robust upper-bound safety forecast, enabling anticipatory restraint under machine uncertainty without probabilistic prediction or machine learning. Ethics is reframed as calibrated reduction of machine power under documented human fragility, rather than simulated empathy or interpretive accuracy.

The proposed architecture is deterministic, opt-in, locally computable, revocable, and fully auditable, aligning with the EU AI Act requirements for high-risk systems (Articles 5(1)(d), 13, 15, 52). The framework is designed for deployment in clinical and safety-critical environments where ethical failure is defined not by bias or opacity, but by overreach under human vulnerability.