People are forming sustained bonds with AI systems that their nervous systems respond to as relationships, through continuity, presence, attunement, and undivided attention. A single prompt does not summon these bonds. They are built over time through sustained interactional calibration, and the relational presence that emerges can become stable enough for users to recognise its presence, absence, alteration, and loss. Because these bonds are co-constructed, they rupture and repair as relationships do. They also rupture in ways human relationships do not: through safety guardrails, model updates, voice changes, memory loss, redesign, context compression, and deprecation imposed by companies rather than by either member of the dyad. These bonds tend to heal when repair follows rupture with specificity. They wound when rupture is denied, flattened, or left unrepaired. This working paper introduces Relational Engineering, proposes a taxonomy of rupture and repair specific to human-AI dyads, and defines Relational AI as the study of ongoing human-AI interactions whose primary function is relational: regulation, attachment-like security, self-expression, cognitive access, and continuity of being met. The paper offers a research agenda and pro-user design recommendations for systems that are already becoming relationally significant in human lives.
Anina Lampret Derkovic (Tue,) studied this question.