Where should LLM systems govern inferred personal information?

The same 18-participant study (the Reflective Layer probe, 215 surfaced inferences) reports that users don't evaluate an LLM-drawn inference once; they evaluate it three times, at different decision points, often reaching different verdicts each time: "generation, or whether the inference should have been drawn at all; retention, or what they could do about it once drawn; and transmission, or if, how, and when it is acceptable for a third party to access it." At generation, participants "resisted inferences drawn from" task-specific content, treating a message typed for one purpose (drafting a sick-leave email) as acceptable source material for a persistent claim about identity only when it was shared for self-description, not when it was shared to get a task done. At retention, concerns split into five named categories — "derivation... capability... persistence... aggregation... and usage" — each pointing to a different interface affordance.

The authors ground this in a two-hop model: memory is the first hop, "the fragments of prior user input that an LLM-based system extracts and surfaces as context"; inference is the second hop, a conclusion "derived" from those fragments using world knowledge, not stated by the user. Because a prompt is a situated request while an inference is "intentionally situationally invariant," the same disclosure (typing a request) does not imply consent to the same durability (a stable trait claim). This is why the paper proposes a design response per decision point rather than one global toggle: surfacing each new inference alongside its source evidence for accept or reject at generation; exposing retention scope, duration, and composition at the retention stage; and binding use to a specified purpose and recipient at transmission.

This sharpens the lifecycle framing in What memory controls do users actually want from LLMs?, which names generation, management, usage, and updating as the stages where users want control over stored memory objects; this paper draws a similar lifecycle map but for inferences specifically, and argues the two objects need distinct governance because an inference is "situationally invariant" in a way a memory fragment is not. It is the mechanism behind the acceptability pattern in Do people fear or welcome AI inferences about themselves?: that note's recipient-sensitivity finding (platform versus advertiser) is the transmission decision point named here, and its misrepresentation-driven discomfort maps onto the generation and derivation concerns.

The three-stage framework and its design mapping come from qualitative analysis of interview and sketching-activity data in a single study session, not from a built or tested interface — the excerpt does not show whether presenting accept/reject controls at generation would hold up under real usage load, nor whether the five retention sub-categories are exhaustive outside this study's inference categories (relationships, interests, personal versus professional). The framework is a design target inferred from participant reasoning, not a validated governance mechanism.