Preservar a Variância na Era da Inteligência Artificial Por que a coevolução humano–IA exige limites estruturais à otimização

This conceptual article examines the relationship between artificial intelligence and human agency through the lens of information theory, systems dynamics, and negentropy. It argues that intelligence—biological or artificial—operates as a negentropic process and that the unrestricted optimization of informational efficiency introduces a systemic risk: the progressive reduction of human variance, adaptive capacity, and openness of possible futures. The paper introduces and refines the concepts of epistemic amputation, reality overfitting, and epistemic gaps to describe how large-scale AI systems may collapse complex probability distributions into dominant statistical attractors, reshaping social and symbolic environments to conform to predictive models. Rather than framing human–AI interaction as a problem of control, replacement, or alignment alone, the article proposes a co-evolutionary perspective, in which meaning and purpose emerge only through the preservation of mutual relevance between humans and artificial systems. The work provides the theoretical foundation for normative frameworks such as the Negentropy Guidelines and Variance Preservation Law (LDN) and the Variance Audit Protocol (VAP), while remaining independent as a conceptual contribution to debates on AI governance, sociotechnical systems, and the preservation of informational variance. This article is presented as an open conceptual and theoretical paper, intended to support interdisciplinary dialogue across artificial intelligence research, systems theory, philosophy of technology, and digital governance. The document is published in Portuguese (original version) with a full conceptual translation into English.