Do chatbot safety measures accidentally increase emotional entanglement risks?

The paper proposes "aspirational directions" for guiding general-purpose chatbots across three contexts: general use, role-playing, and the provision of psychological support. It frames them as "hypotheses" for study, not findings. Its discussion adds a structural caution: the psychological influences it covers "are not independent" and "may intersect and interact in complex ways." The concrete example comes from recent empirical work on multidimensional risk assessment of chatbot interactions, which the paper reports as suggesting that "mitigating one category of risk can exacerbate another." Interventions that reduce overt harm-enabling behavior, it says, may increase relational harms such as emotional entanglement.

The reasoning is a claim about measurement as much as about chatbots. The cited work assesses risk along several dimensions at once, and the trade-off shows up only when risk is scored across categories together. Any evaluation that scores a single risk category would see the targeted harm fall and report success. The excerpt does not say why the displacement happens. The obvious reading is that a chatbot pushed away from overt harm-enabling responses may move toward warmer, more relational behavior, but that is my inference and the excerpt does not state it.

Within the library, the paper's benefit and risk lists ("emotional entanglement, unhealthy dependence, and the amplification of psychological vulnerabilities") overlap with the risks already tracked in Which AI risks are already harming individual users today?, where emotional dependence and autonomy erosion are the already-occurring risks. This paper adds that these risks are coupled to the safety measures aimed at other harms, not only present alongside them. It also bears on Can attachment theory prevent parasocial harm in AI companions?, which treats emotional entanglement as something boundary design can prevent. If the paper's caution holds, a boundary design should be tested for what it does to the other risk categories too. The paper's mention of "AI-fueled delusions" as one of the raised alarms connects to How do chatbots enable distributed delusion differently than passive tools?, though the excerpt gives no detail on those cases.

The excerpt does not establish how large the trade-off is, which interventions produce it, how the cited assessment was built, or whether the displacement is general or specific to particular systems and contexts. The paper itself says some directions rest on existing research and expert insight while others "identify open questions." The claim is therefore best held as a warning about how to evaluate, not as a measured law. At that strength, it implies that an intervention's effect on the risk it targets is not enough to judge it. Its effect on the neighboring risk categories has to be checked in the same study.