Scholars actively debate what artificial intelligence is for humans — a cultural tool, a means of digital externalization, a digital significant Other, or a future alternative, replacement for the human. Far less frequently discussed is who the human is for technologies and their designers. The paper shows that AI technologies — understood not as a single engineering solution but as a system of heterogeneous approaches — not only rest on fundamentally different conceptions of the human, but also transmit these images through their architectures, thereby facilitating corresponding practices and shaping culture, new worlds and a new forms of being. The architectures of most types of AI both, implemented in practice and only discussed in the literature (generative, human-centered, trustworthy, explainable, interpretable, embodied, situated, enactive, social, as well as designs developed within sociotechnical and actor-network frameworks) rely on reduced, incomplete images of the human being. Among the principal risks associated with broad deployment of such systems is the risk of dehumanization — the elimination from human practices of those features that make them genuinely human: contextual understanding, meaning-making, moral and ethical responsibility, deep interpersonal bonds, coaction, variability, autonomy and self-determination. In these architectures the human too often appears only as a performer of a social role (user, client, worker, patient, student, etc.). Overcoming the reduction of the human in AI architectures is possible through redesign of those architectures. Corresponding technologies must move beyond simplified anthropologies and be built on complex images of the human being. Such technologies can not only facilitate human development and the realization of human potential, but also contribute to the formation of a truly human-centered culture, a culture of dignity (A.G. Asmolov). Practical tenets intended to underlie the design of non-reductionist architectures are formulated.