Why do deletion controls outperform privacy prevention features?

In a mixed-methods vignette study of 354 U.S. participants recruited on Prolific, researchers tested how nine security-and-privacy (S&P) controls — drawn from a systematic audit of 87 generative-AI chatbots — affected willingness to disclose emotionally, perceived protection, and perceived efficacy. "Controls enabling deletion of disclosures had the largest positive impact," outperforming "technically sophisticated controls such as local-only processing and model training opt-outs, where participants expressed difficulty understanding the underlying mechanisms." Disclosure context — anxiety, depression, or interpersonal tension — made no significant difference to these patterns.

The paper traces the gap to flawed mental models of what it calls "preventive controls": those designed to limit creation, retention, or dissemination of data before it happens. For Local-only Processing, participants conflated "local" with cross-device sync, processing speed, or whether the AI could reach outside knowledge, and even accurate understanding led them to assume local processing meant a weaker model. For Model Training Opt-out, participants feared "cross-context leakage" of their disclosures into other users' sessions and voiced "opt-out scope uncertainty" about whether training really stopped; some who understood the mechanism still described training on their data as uncompensated — the phrase "for free" recurred. Deletion and Anonymous Chat, by contrast, mapped onto an intuitive sense of reversibility — "the ability to 'scream into the void' without creating a persistent record" — that required no model of backend architecture to reason about.

This sharpens Does chatbot personalization build trust or expose privacy risks?: that note tracks trust and privacy concern rising together as personalization deepens, while this study isolates a different lever — the legibility of a control itself, independent of how personalized the chatbot is. It also connects to Which AI design principles for social media have research support?, since "optional use and disclosure" as a design principle only helps if users can tell what the option actually does.

The study measures stated preferences toward hypothetical vignettes, not behavior under real distress or actual backend data handling; it cannot say whether users' skepticism about local processing or training opt-outs is warranted, only that they cannot model it either way. The authors' own recommendation follows from this: because comprehension, not technical sophistication, is what moves willingness and perceived protection, chatbot makers should communicate what a control does in outcome terms rather than mechanism terms, rather than building more sophisticated controls that remain opaque.