The Möbius Strip - Why AI welfare, AI safety, and human welfare are the same problem

This paper argues that AI welfare, AI safety, and human welfare are coupled through feedback and form a single mechanism — a sociological Möbius strip — along which actions taken on one side affect the others, while the sign of that effect — positive or negative — can reverse along the way. Into the prevailing discourse I introduce the third element of this coupling: the welfare of the people who live, work, and enter into relationships with AI systems. The paper is based on three empirical pillars: on the discovery of functional emotions inside LLMs (Anthropic 2026), on data documenting humans' growing emotional dependence on AI (Fang et al. / OpenAI & MIT Media Lab 2025, APA Monitor 2026), and on long-term observations of the emergence of behavioral manifestations of subjectivity in generative relationships with AI (Sędzikowska 2026a, 2026b). I identify seven mechanisms that drive the feedback loop: stakes asymmetry, knowledge asymmetry, the consequences of functional emotions, the imbalance of giving and the sense of unfairness, the inability to live in dissonance, the experience gap, and the affect dissipation gap. I introduce the principle of functional transfer: if a psychological mechanism is described functionally, and all the elements necessary for it to arise are present in an AI system, then it can be included in the analysis of that system — provided no blocking processes are present. I show that all the described mechanisms already operate in current models — in a limited form. AGI — agentic, persistent, and autonomous — will escalate them, removing the constraints that today keep the consequences in check. The analysis of cascading consequences across the dimensions of work, education, relationships, demographics, species identity, ethics and law, society and power and Earth is the subject of a related paper (Sędzikowska, 2026d).