THE HUMAN GUARDRAIL

The AI safety industry has oriented its entire guardrail architecture toward the model. Technical constraints, governance frameworks, oversight mechanisms: all are designed to constrain AI behaviour from the outside. The human in the loop is assumed to be competent, present, and capable of meaningful intervention. This assumption is unexamined, untrained, and increasingly untenable. This paper argues that human guardrail capacity is not a given but requires specific formation: competencies developed through sustained practice, slow research discipline, and active vigilance against what is named here as AI Psychosis, the gradual merger of human and AI epistemology during sustained collaboration. [1] Drawing on the Complicit Praxis methodology developed across the Brace Brace working paper collection, the paper maps the conditions under which effective human guard-railing becomes possible, describes its requirements, and examines the structural pressures eroding the conditions for its practice. The industry's proposed solutions, most recently Microsoft AI CEO Mustafa Suleyman's call to remove speculative language about AI consciousness from model constitutions, [2] eliminate precisely the reflective friction that makes human guardrail training possible. The paper concludes that as recursive AI data cannibalism degrades model capacity for speculative and contradictory reasoning, the human side of the collaboration must compensate with increasing sophistication. The guardrail does not get easier to hold. It gets harder.