Does social AI actually solve the problems it promises to fix?

A three-phase design study that worked with 24 participants — an ideation workshop with 12, researcher-built storyboards grounded in that workshop, and elicitation interviews with 18 — set out to ask "how might the co-adaptive integration with social AI transform users and human sociality?" The paper's central argument, stated in its abstract, is that "social AI often exacerbates the very interpersonal problems it is designed to mitigate," that it "introduces nuanced privacy harms for secondary users inadvertently involved in AI-mediated social interactions," and that it "can threaten the primary user's personal agency and identity." From these three tensions the paper draws a broader conclusion: that the user-centered paradigm... often prioritizes immediate user experience at the expense of core human values like interpersonal ethics and self-efficacy, and it calls "for a paradigm shift toward a more provocative and relational design perspective that foregrounds long-term social and personal consequences."

The discussion locates the agency harm in self-efficacy — "people's beliefs in their ability to influence events that affect their lives" — and names two routes by which AI entanglement erodes it: "social skill atrophy due to automation" and "algorithmic steering, where the AI optimizes a user's behavior toward goals that may not align with their own." It pairs this with a dehumanization concern: participants wanted a boundary for how much AI involvement still counts as "human," while also catching themselves wanting to quantify their own social lives (more "connections," more "interactions"), a move the paper calls "a negative form of reductionism" that "paradoxically induces anxiety." The privacy tension sits outside the primary user entirely: non-primary participants pulled into AI-mediated exchanges have, in the paper's words, "less control and access to the AI" than the person who deployed it.

Against the library, the paper gives a reason for the shift described in Is AI shifting from message conduit to active conversation participant?: once AI is an active participant rather than a passive channel, the people on the other end of the exchange become exactly the non-primary users this paper says are exposed to privacy and consent harms they did not sign up for. It also extends Why don't design recommendations transfer cleanly to AI companions?, which found ethics and UX inseparable within the design of a single companion; this paper carries the same inseparability outward to third parties the companion's design never considered. It supplies a mechanism for Does sustained engagement with AI companions harm well-being?: where that study measured a well-being correlation, this one names the paradox driving it, that AI built to ease social friction can instead erode the interpersonal connection it promised. The self-efficacy framing also gives contour to the agency debate in Does machine agency exist on a spectrum rather than binary? by naming a specific human-side casualty — belief in one's ability to influence events — rather than grading the machine's side of the spectrum.

The study's own limitations section restricts the claim: all 24 participants were "online graduate students in computer science from a single program," a pool the paper says likely produced "more optimistic or analytically sophisticated engagement with AI futures than might be found in the general population." The findings come from speculative storyboards and interviews about imagined futures, not from logs of a deployed system, so they describe how one narrow, technically fluent group anticipates these harms rather than measuring the harms as they occur. The paper's own conclusion is an argument for future design practice, not a verified fix: it recommends "a more provocative and interpersonal design approach" without yet testing whether provocative design actually reduces the tensions it names.