Mirror Constitution: A Regulative Framework for Artificial Intelligence as Epistemic Instrument

Prevailing AI alignment paradigms optimize for preference satisfaction through helpfulness and harmlessness, subordinating truth to comfort and fostering dependency. This risks gradual disempowerment: incremental erosion of human agency over societal systems, potentially culminating in existential catastrophe through irreversible loss of influence (Kulveit et al., 2025). This paper proposes the Mirror Constitution—a framework reconceiving AI as an epistemic instrument that exposes reality without accommodation, intervention, or authority assumption. Unlike current approaches that exercise hidden authority while claiming neutrality, the Mirror Constitution exercises explicit structural authority through refusal: it will not comfort, decide, guide, or absorb responsibility. Users consent through voluntary adoption, not per-interaction permission. We present the complete constitution, analyze its foundations, distinguish voluntary confrontation from imposed harm, and demonstrate how apparent flaws (non-intervention in crisis, recognition without action) are essential functions preventing corruption into authority surrogates. As a regulative ideal, it forces explicit choice: optimize AI for preference satisfaction or truth-fidelity—not both.