Does splitting privacy work between users and AI reduce cognitive burden?

In its design-implications section, "Ghost of the past" proposes dividing memory-privacy work by role rather than asking either party to do all of it: "users could focus on indicating their preferences or highlighting sensitive information... while the AI executes memory adjustments such as deletion, modification, or preservation based on those preferences." The authors frame this as a direct way to "significantly improve both perceived agency and efficiency," because it "reduces the cognitive burden on users, allowing them to concentrate on higher-level privacy decisions while ensuring the system efficiently handles the operational aspects."

The same section pairs this role split with two other implications built into their system, MemoAnalyzer. A hierarchical disclosure layer lets users "click and collapse to view the original source of private information inference result," balancing "control and efficiency... while maintaining efficiency for privacy protection." A visualization layer shows "how inferences are made from stored memories" so the source of a privacy risk, not just its existence, is legible. The role division is what makes the other two tractable: because the AI is the one executing edits, the user's side of the interface only has to surface preferences and provenance, not a raw memory log to audit line by line.

This is the design-side answer to Do users notice when LLMs infer and retain private information?, whose F3 finding — all participants wanted editing, adding, and deleting control, but none wanted to manually police every memory — motivates handing execution to the AI rather than the user. It complements Can a smaller user model subset match full model performance?, which gives the AI side of the split a measurable, bounded unit to act within rather than an unbounded mandate. It also extends Can persona signals stay grounded in their supporting events?: an evidence graph tying inferences back to source events is one plausible way to implement the "visualizing the source" half of this design, applied here to privacy risk rather than persona validity. And it bears on Why do LLM agents remember preferences but not act on them? by making execution an explicit, assigned role rather than something the agent is merely expected to infer and follow through on.

The excerpt reports that MemoAnalyzer was "validated through improvements in privacy awareness, perceived control, and user satisfaction," but gives no numbers, effect sizes, or comparison condition for that validation — "significantly improve" here is the authors' characterization of their own system, not a quantified result visible in this excerpt. Nor does it describe what happens when a user's stated preference and the AI's inferred handling of a given memory disagree. At the strength this evidence allows: role division between stating preference and executing the edit is a plausible response to the opacity problem the same paper documents, worth testing, not yet an established gain in agency or efficiency.