Exploring the Interaction of Explanation Styles, Context, and Trust of AI Privacy Redaction in AI-mediated Interactions

Jul 16, 2026

Humans increasingly rely on AI-mediated communication to support everyday interactions, yet privacy-sensitive domains require the mediator to preserve confidentiality while sustaining user trust and adapting to user preferences. Prior research shows that explanations for redaction can improve trust versus offering no explanation. This work builds an AI mediation system that generates redacted dialogue and produces explanations in multiple styles. An online study with 249 participants examines how explanation style preferences vary with context and how trust aligns with users’ preferred explanations.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

Abstract

Introduction

Background: AI mediation and privacy redaction

Role of explanations and user trust

Preferences across context and redaction settings

Exploring the Interaction of Explanation Styles, Context, and Trust of AI Privacy Redaction in AI-mediated Interactions Roshni Kaushik Fujitsu Research of America Pittsburgh, Pennsylvania, USA[rkaushik@fujitsu.com](rkaushik@fujitsu.com) Maarten Sap Carnegie Mellon University Pittsburgh, Pennsylvania, USA[msap2@andrew.cmu.edu](msap2@andrew.cmu.edu) Koichi Onoue Fujitsu Research of America Pittsburgh, Pennsylvania, USA [konoue@fujitsu.com](konoue@fujitsu.com) arXiv :2607 .06687v 1 [ cs .HC] 7 Jul 2026 Abstract Humans are increasingly using AI-mediated communication to help facilitate human interactions; however, in privacy-sensitive domains, the AI mediator has the additional challenge of considering how to maintain user trust and adapt to user preferences while preserving privacy. Prior work has shown explanations of redaction improve trust compared to no explanations in privacy-sensitive scenarios. We explore this research area further by developing a system where an AI mediates conversations containing private information while providing explanations of different styles. The system uses an LLM for (1) information generation in a particular domain and with a specific amount of sensitive information to redact, (2) redaction of sensitive information,(3) explanation generation in a particular style, and (4) redaction of sensitive information in the generated explanation. We then conduct an online user study (􀀽 = 249) to understand the interaction between explanation styles, context (domain and redaction amount), and trust in these privacy-sensitive scenarios, where users choose which explanation style(s) they prefer for a particular context and rate their trust in the system with their preferred explanation(s) . Our results reveal that explanation preferences vary systematically with context, users’ trust in the system is linked to preferred explanation, and users display consistent individual differences in how they choose explanations and how their preferences depend on context. These findings highlight the importance and design implications of developing personalized, context-aware explanationsin privacy-sensitive scenarios and suggest that adaptive explanations are essential for designing privacy-aware and trustworthy AI mediators. CCS Concepts • Human-centered computing → HCI design and evaluation methods; User studies; • Security and privacy → Usability insecurity and privacy. Keywords Explanations; Trust; Privacy-preserving systems; Human–AI interaction; Transparency Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission [and/or a fee. Request permissions from permissions@acm.org](and/or a fee. Request permissions from permissions@acm.org). UIST ’26, Detroit, MI © 2026 Copyright held by the owner/author(s) . Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 [https://doi.org/XXXXXXX.XXXXXXX](https://doi.org/XXXXXXX.XXXXXXX) ACM Reference Format: Roshni Kaushik, Maarten Sap, and Koichi Onoue. 2026. Exploring the Interaction of Explanation Styles, Context, and Trust of AI Privacy Redaction in AI-mediated Interactions. In Proceedings of ACM (Association of Computing Machinery) Symposium on User Interface Software and Technology (UIST ’26) . ACM, New York, NY, USA, 18 pages. [https://doi.org/XXXXXXX.XXXXXXX](https://doi.org/XXXXXXX.XXXXXXX) 1 Introduction AI systems are increasingly used to mediate communication between people, particularly in settings where sensitive information must be selectively reve

FAQs

What problem does the study address in AI-mediated privacy redaction?

The study targets how an AI mediator can maintain user trust and adapt to preferences while preserving privacy when handling conversations that include private information.

How does the proposed system perform redaction and explanation?

It uses an LLM to generate domain information, redact sensitive content, generate an explanation in a chosen style, and then redact sensitive information again inside that generated explanation.

What do the results show about explanation preferences and trust?

Explanation preferences change systematically with context, users’ trust is linked to the explanation style(s) they prefer, and individuals show consistent differences in how they choose explanations and how those preferences depend on context.

Jul 16, 2026

Upload by River Wang