Research commentary · Research revisit · Disclosure, memory and privacy controls
After opening up to AI, what can you take back?
In a U.S. vignette study with 354 adults, deletion controls were rated favorably. A memory toggle showed different patterns for willingness, perceived protection and expected help. Those differences raise questions about what the controls actually do.
The author manuscript was submitted on 7 July 2026 and the proceedings paper appeared in August 2026. The study measured ratings of hypothetical controls.
Before speaking, you may want to know what will stay
Imagine opening a chatbot late at night to make sense of an argument. You might describe who was involved, what troubled you and details you would rather not read again. The next time you open the chat, you may want the system to remember the relationship while leaving out the details of that evening. Continuity and control over what stays are different questions.
A study by Kwesi and colleagues offers a way into those questions. The researchers compared descriptions of nine security and privacy controls and examined how people rated an imagined chatbot. Participants responded favorably to deletion controls, although they could understand the button and still doubt what the system would do. [1]
Deletion gives people a choice about what to keep
Deleting a single conversation served as the statistical reference. The study detected no significant difference from that reference in willingness to use for deleting the account and its data or for “anonymous chat,” defined in the study as choosing whether to save a conversation to account history. Training opt-out and local-only processing received lower willingness ratings. Participants rated descriptions at this stage, without operating the controls in a real product. [1]
Some respondents explained that they wanted to remove a particular conversation while keeping others. That distinction has practical consequences. Leaving a service entirely and clearing the record of a specific experience give people different choices. If clearing one conversation costs them access to everything else, they also lose material they may still find useful. An evaluation can therefore ask how precisely people can choose what to delete, beyond recording whether a delete option exists.
The scope of deletion depends on what the system does with the data. If a conversation disappears from view, can a later session still draw on related information? When does a request take effect, which stores does it cover, and are there retention exceptions? The study documented doubts about execution. Verifying how any actual implementation handles these questions requires a separate check. [1]
Willingness, reassurance and expected help can come apart
Cross-session memory illustrates the difference. Compared with deleting a single conversation, the memory toggle received lower willingness and perceived-protection ratings, with no significant difference detected in expected helpfulness. Some respondents valued the ability to carry context forward. Multifactor authentication showed a different split. Its perceived-protection rating did not differ significantly from the deletion reference, while willingness to use was lower. [1]
How much background people want to carry forward can change over time. Sometimes they want to work through a thought in one sitting. At other times, they want to pick up where they left off. A product could make three decisions visible: what the system can use now, what it can recall later and what the user can take back. An observer could then distinguish wanting a coherent reply from agreeing to longer retention.
Turn the setting’s name into something you can check
Settings grouped under one privacy menu may act at different stages. In the study scenarios, the memory toggle governed recall across sessions, training opt-out governed use of conversations in model training, and conversation deletion concerned the record in account history. Offering one control leaves other parts of the data journey to be explained. [1]
A follow-up evaluation could use fictional conversations with no real personal details. Ask people which pieces they want to bring into the next session, then observe whether they can find the setting, understand what is retained and carry out their choice. Separately, check what happens to test data in the interface, across sessions and within the stated scope of data processing. This would make comprehension, usability and execution separate outcomes, with room for people who choose to share less.
These are this site’s proposals for future evaluation. The original study used hypothetical scenarios and self-reports. It measured neither real disclosure behavior nor actual protection or health outcomes. Its U.S. English-speaking sample already used AI chatbots at least monthly for emotional or social support. The findings need further study with non-users and in other languages and cultures. The app review also reflects the screening criteria and period of August to September 2025. [1]
Original source and verification
- Kwesi, J., Cao, J., Cunningham, H., & Emami-Naeini, P. (2026). The Impact of Security and Privacy Controls on Users’ Emotional Engagement with Generative AI Chatbots. 35th USENIX Security Symposium, 1647–1666. Complete proceedings paper · Conference paper page · Author manuscript record.
The author manuscript was submitted on 7 July 2026, and the proceedings paper appeared in August 2026. On 6 November 2026 we checked the proceedings paper’s methods, Tables 1–5, limitations and appendices. The links lead to the same study. This article reproduces no figures, product screenshots or participant quotations.
This article contains AI-generated content. AI helped locate the original sources, check study details, and write and edit both languages. This site conducted no hands-on product testing or original interviews. The operating team develops Six Realms and provides Heart to Heart and Sleep, creating relevant commercial interests. This article does not evaluate the privacy implementation or support outcomes of those products.