Machine translation can make a government notice readable in a refugee’s language, but a faithful translation of bureaucratic Korean is still bureaucratic, and newcomers often cannot act on it. We report, with the operating organization as a co-author, what happened when a small refugee-serving organization in South Korea put a large language model in front of the translation step at the service it has operated since 2018. Scoped with the frontline staff who field refugees’ questions, the model rewrites the Korean source into plain language, explains Korea-specific terms where they appear, checks whether figures and deadlines have gone out of date, and pins proper names to a fixed glossary before anything is translated into seven languages. It covers most of the catalog and shortens the sentences readers meet. It also fails in ways a translation-only system never does. The stages allowed to search the web made corrections no translation could make, and they also wrote every serious error we found, including policy updates that were simply wrong; the stages without that access wrote a handful, none of them critical and all of them confined to metadata. Value and failure come from one capability, so the lesson is not to withhold it but to put the human check where it acts. For an organization with nobody on staff who reads the seven target languages, that check must be placed on the Korean source before publication, which makes Korean-language staff review the safeguard the whole design depends on.

People in distress increasingly talk to a consumer chatbot the way they would talk to a therapist: they vent to the model, work through feelings with it, and sometimes call it their therapist. We ask where that behavior is located in relation to the existing system of human care: which communities the people doing it belong to, what discourse crosses the boundary between those communities and human therapy, and in what order people arrive on the two sides. Across ten Reddit communities covering general chatbots, companion chatbots, an AI self-help community, mainstream human therapy, and therapy disillusionment, we combine the backbone of a user-by-community bipartite projection under the fixed degree sequence model, a directed reply network, a structural topic model, and an analysis of which side users reach first. Of the ten, the AI self-help community is the only one connecting the human-therapy and chatbot worlds, and its strongest backbone edge is to the therapy-disillusioned, at nine times the null expectation. Companion chatbots are as relational as the self-help community and their overlap with therapy grievance is still at chance, so relational chatbot use is not by itself adjacent to it; the adjacency is specific to purpose-built self-help. Visibility in the cross-community discourse is concentrated in one commercial account, and human-therapy participation predominantly precedes therapeutic chatbot use, though the data cannot establish that disillusionment drives the move. We interpret these patterns through the Network Episode Model.

AI ethics guidelines for humanitarian settings have grown in number and scope. Whether they produce their intended outcomes depends on which deployers are expected to follow them. These guidelines respond to documented risks: surveillance, data misuse, and discriminatory outcomes affecting refugee populations. For high-risk applications such as biometric identification and asylum adjudication, the concerns they address are genuine. Many differentiate risk tiers in principle, yet the compliance expectations they establish (staff capacity, technical infrastructure, formal evaluation) reflect the organizational contexts in which they were developed. Many nonprofits providing frontline services to refugees operate with limited administrative capacity. When compliance requirements exceed what these organizations can meet, formal AI adoption stalls, while informal adoption proceeds without oversight or recourse. Current guidelines also tend to treat non-adoption as a neutral default, without accounting for the service gaps that follow when AI-assisted language access is unavailable. Drawing on collaboration with refugee-serving practitioners, we show that this gap between governance design and organizational reality has consequences for the people these guidelines are meant to protect. Evaluating AI guidelines, we argue, requires the same realist logic that evaluation research has long applied to social programs: not “does this guideline exist?” but “for which deployers, under what conditions, and does it produce its intended protective outcomes?”