Abstract: As the popularity of social AI grows, so has the number of documented harms associated with its usage. Drawing on Human-Computer Interaction (HCI) and Machine Learning (ML) literature, we frame the harms of AI companions as a $\textit{technological problem}$ and draw direct links between key technical design choices and risks for users. We argue that many of the observed harms are foreseeable and preventable consequences of these choices. In the spirit of $\textit{translational research}$, we offer concrete strategies to mitigate these harms through both regulatory and technical interventions, aiming to make our findings useful and actionable for policymakers and practitioners.