Do deletion controls actually remove AI chatbot data as promised?

In the same 354-participant vignette study of security-and-privacy controls in generative-AI chatbots, deletion-based ("reversibility") controls dominated willingness to disclose, perceived protection, and perceived efficacy, yet "trust remains fragile, and participants often doubted S&P controls would function as promised." Participants preferred deletion over technically sophisticated alternatives like local-only processing, but still "expressed skepticism about whether such controls could fully remove their information or the inferences derived from their conversations" — the preference for deletion was not full confidence in it.

The paper names this the assurance gap, distinct from the comprehension gap it finds in preventive controls. With preventive controls, users don't understand the mechanism; with deletion, "users understand the promise but doubt the system will honor it." The distinction matters for design: explaining how a deletion control works does not resolve skepticism about whether the company will actually do what it says, because the gap is about credibility rather than legibility.

This complicates Does chatbot personalization build trust or expose privacy risks?, which treats rising trust and rising privacy concern as two outcomes of deepening personalization: here, privacy concern persists even for the control category users most prefer, so offering a preferred control redirects distrust rather than resolving it. It pairs with Why do deletion controls outperform privacy prevention features? from the same study: together, deletion wins on comprehension but not on assurance.

The excerpt does not establish whether this skepticism is accurate — the paper does not test whether deletion controls in the audited apps actually remove disclosures or the inferences drawn from them, only that participants doubt it regardless of the underlying truth. The authors' own recommendation is that credible assurance, not just interface simplicity, is required even for the most-preferred control category, a point they set against federal and state policymakers' recent divergence on regulating AI in sensitive, emotional contexts — a divergence the excerpt names but does not detail.