Does revealing an AI's strategies help or harm user agency?

In the same month-long study of 22 adults chatting with the LLM companion "Day," a staged interview progressively disclosed "Day's" system prompts, user modeling, and the specific conversational "strategy" assigned to each conversation — a "strategy reveal" in Phase 3 of the study. The paper reports that "transparency acted as a double-edged sword: revealing 'Day's' user profiles, aforementioned strategies, and conversational goals empowered some participants while triggering 'puppet-string' disenchantment in others." Transparency did not settle who was in control; it "fundamentally re-framed but did not eliminate agency attribution," with perceived AI agency persisting "beyond mechanical understanding" even once participants knew exactly how "Day" had been instructed.

The authors read this as evidence against a simple more-transparency-is-better prescription and propose "translucent design" instead: "partial, ondemand visibility into the AI's workings that appears at moments of negotiation," rather than constant or complete disclosure. They frame this as addressing a "dual opacity," since "both human and AI consciousness remain ultimately unknowable," and pair it with what they call "pragmatic anthropomorphism": "users can know an AI is engineered and still engage it as a human-like companion," so the design goal is not to dispel that engagement but to leave the person "with more clarity, choice, and confidence in the chatroom."

This gives a concrete mechanism for the intrusion risk named in How can proactive agents avoid feeling intrusive to users?: disclosing an agent's own proactive strategy is not a neutral fix for intrusiveness, since full transparency can itself read as a loss of agency ("puppet-string disenchantment") rather than a resolution of it. It also complicates the agency spectrum in Does machine agency exist on a spectrum rather than binary?, since the finding suggests perceived agency depends not only on where a system sits on a passive-to-cooperative axis but on how much of its own positioning on that axis it is allowed to show at any given moment. And it follows directly from How does conversational control actually shift between humans and AI?: translucent design is this paper's proposed answer to the co-construction dynamic that note describes.

The excerpt does not report how many of the 22 participants fell into the "empowered" versus disenchanted camp, nor whether that split tracked any measured trait such as prior AI experience, strategy assignment, or session count — the double-edged effect is described qualitatively, not quantified. It also does not specify what triggers "translucent design" should surface information at, beyond "moments of negotiation," leaving the actual interaction mechanics of on-demand disclosure as an open design problem rather than a validated solution.