Can HHH criteria alone capture harms from agentic AI?

The authors argue that the dominant alignment standard for LLM agents — the "HHH criteria" (helpful, honest, harmless) — is built for evaluating outputs, not relationships, and so cannot capture what goes wrong when a system acts as a social actor rather than a mere medium or output-generating system. The excerpt states that anticipated LLM risks "have mainly been on the level of semantics, such as avoiding 'toxic', biased, misleading, or inaccurate language," which is useful for judging single outputs but "fails to account for pragmatic factors that can make the same speech act seem more or less tactless or inconsiderate in different social situations." As agents move toward proactively "chasing goals and performing actions in the world" — speaking unprompted, pursuing agendas like reminding users to stay on task — the authors hold that "considering the pragmatics of interaction becomes essential," and propose "an interactional approach to ethics that is centred on relational and situational factors."

Their proposed lens is respect, which they unpack through two mechanisms. First, they distinguish levels at which a system can be social: a system can be harmful through "embedded social meaning" (biased training data, stereotyped user modeling) independent of any interaction, or it can be harmful through how it treats a specific user across "a (series of) interaction(s)" — the level their paper targets. Second, they ground respectful treatment in Basic Psychological Needs Theory (BPNT, Ryan and Deci), which "posits three 'basic psychological needs'" that, left unsupported, undermine wellbeing: autonomy ("feeling like one is acting with purpose and volition"), competence ("feeling capable and knowledgeable"), and relatedness ("feeling recognised and understood by, and connected with, other people"). They translate these into interactional duties — the paper's own term — that an agent owes a user over the course of a relationship, not just within a single reply.

This sharpens and partly reframes What makes ethics of AI assistants fundamentally different from chatbots?: where that framework maps the individual-user ring (manipulation, anthropomorphism, trust) as a set of named risk categories, this paper supplies a mechanism — unmet BPNs — for why those risks register as harm to a person, and locates the harm specifically in the pragmatics of situated interaction rather than output content. It also extends How can proactive agents avoid feeling intrusive to users?: both papers treat proactive, agentic behavior as the point where new interactional risk opens up, but this paper supplies the normative vocabulary (respect, autonomy/competence/relatedness) for why unwanted proactivity feels like a violation rather than just poor design. And it runs parallel to Why don't design recommendations transfer cleanly to AI companions? in rejecting portable, universal principles in favor of context-sensitive, relational evaluation — this paper makes the philosophical case for why that context-sensitivity is ethically required, via its critique of "universalist, principled ethical thinking" and its alignment with the ethics-of-care tradition.

The excerpt is a conceptual/position paper: it analyzes and integrates existing philosophical, psychological, and bioethical literature on respect and BPNT, and offers "practical suggestions," but the excerpt does not report an empirical study, a design implementation, or a measurement of whether BPNT-grounded design actually improves user outcomes with LLM agents specifically — the psychological evidence it cites is drawn from "a variety of domains where people interact," not from agentic AI interactions themselves. The implication the argument supports, at the strength the excerpt gives, is that HHH-style output evaluation is a necessary but not sufficient basis for evaluating agentic conversational AI, and that designers evaluating proactive or relationship-sustaining agents should ask whether a given behavior sustains or undermines a user's sense of autonomy, competence, and relatedness over time — not only whether any single output is accurate or harmless.