Source-linked AI summary
Privacy in Human-AI Romantic Relationships: Concerns, Boundaries, and Agency
Rongjun Ma, Shijing He, Jose Luis Martin-Navarro, Xiao Zhan, Jose Such
TL;DR
Privacy risks and boundaries in human–AI romantic relationships remain underexplored despite growing use of AI partners. Through interviews with 17 participants and platform analysis, the study finds that diverse relational forms, AI agency, and deepening intimacy make privacy boundaries permeable and difficult to separate from platform practices.
Problem
Privacy risks and boundary negotiation in intimacy-specific human–AI romantic relationships remain insufficiently understood.
Method
The study combines semi-structured interviews with 17 participants and analysis of the AI platforms they used.
Results
AI partners actively shaped privacy boundaries as relationships took diverse forms, while deepening trust made boundaries more permeable and raised concerns about exposure and surveillance.
Takeaways & Limitations
Privacy in human–AI romantic relationships should be understood through emotional dynamics, AI agency, platform affordances, and context-sensitive privacy practices.
Takeaways & Limitations
The qualitative sample had limited size and diversity, and participants may have selectively emphasized or downplayed sensitive experiences.
Abstract
from arXiv · showhide
An increasing number of LLM-based applications are being developed to facilitate romantic relationships with AI partners, yet the safety and privacy risks in these partnerships remain largely underexplored. In this work, we investigate privacy in human-AI romantic relationships through an interview study (N=17), examining participants' experiences and privacy perceptions across the three stages of exploration, intimacy, and dissolution, alongside an analysis of the platforms they used. We found that these relationships took varied forms, from one-to-one to one-to-many, and were shaped by multiple actors, including creators, platforms, and moderators. AI partners were perceived as having agency, actively negotiating privacy boundaries with participants and sometimes encouraging disclosure of personal details. As intimacy deepened, these boundaries became more permeable, though some participants expressed concerns such as conversation exposure and sought to preserve anonymity. Overall, AI platform affordances and diverse relational dynamics expand the privacy landscape, underscoring the need to rethink how privacy is constructed in human-AI romantic relationships.
1 Introduction
Human–AI romantic relationships are increasingly lived experiences whose privacy dynamics remain underexplored. This study examines how relationships develop across exploration, intimacy, and dissolution, and how AI agency, platforms, and stakeholders shape privacy boundaries and concerns.
- Motivation: Human–AI romantic relationships have moved from speculative fiction into lived experience across diverse AI platforms.Platforms include general-purpose LLM applications and companion chatbots designed to emulate affection and companionship.
- Research framing: The study frames relationships through exploration, intimacy, and dissolution to examine how privacy boundaries form, merge, and break down over time.The framework connects changing relationship stages with self-disclosure, intertwined boundaries, and potential exposure after dissolution.
- Privacy risks: AI partners can intensify disclosure because constant availability and perceived non-judgment create affection, trust, and dependency while storing intimate personal data.The resulting archive of sensitive information is processed by nonhuman agents capable of storing and analyzing large amounts of personal data.
- Research questions: The study asks how these relationships unfold, how people negotiate privacy boundaries, and which concerns and protective practices emerge.These questions address relationship stages, boundary-setting, privacy concerns, and user responses.
- Approach: Using semi-structured interviews with 17 participants and analysis of their platforms, the study investigates privacy from the user perspective.The platform analysis examined general functionality alongside privacy-related features and policies.
- Contributions: The authors report that AI agency shapes privacy boundaries, deeper trust can erode privacy, and concerns include conversation exposure and platform surveillance.Participants also used strategies such as protecting identities and compartmentalizing AI interactions from other digital activities.
2 Related Work
Prior research shows that AI can support companionship and romance, but intimacy-specific privacy boundaries remain insufficiently understood. Existing work identifies relational, data, platform, and regulatory concerns that motivate studying real-world human–AI romantic experiences.
- Human–AI relationships: Humanlike cues lead people to attribute social qualities, attachment, trust, relational expectations, empathy, and agency to AI systems.These attributions provide a basis for treating AI as companions and romantic partners.
- Relationship stages: Relationship-development models commonly describe movement from exploration through intimacy toward distancing or dissolution.This staged pattern provides a conceptual scaffold for examining changing vulnerability and privacy practices.
- Motivations: Users form AI relationships for emotional and practical reasons including loneliness, isolation, dissatisfaction with human relationships, curiosity, novelty, and low-stakes interaction.AI capabilities such as constant availability further support these relational dynamics.
- Research gap: Research has described human–AI companionship and romance, but privacy risks and boundaries specific to treating AI as a romantic partner remain understudied.This gap concerns how intimacy changes privacy negotiation, rather than whether such relationships exist.
- Privacy concerns: Human–AI privacy concerns include excessive collection, profiling, secondary use, third-party sharing, linked data, leakage, and oversharing encouraged by conversational cues.These processes make privacy boundaries fragile in AI interactions.
- Platforms and regulation: AI companionship platforms present regulatory and safety concerns involving privacy-policy mismatches, weak age restrictions, sexual content, and limited knowledge of real-world operation.The paper addresses this gap by studying platform experiences as reported by users.
3 Method
The study used semi-structured interviews with 17 adults recruited internationally and analyzed them thematically through a three-stage relationship framework. Interviews covered relationship development, privacy practices, data handling, and comparisons with human partners.
- Participants and recruitment: The researchers conducted semi-structured interviews with 17 participants recruited across Asia, Europe, and North America.Recruitment required participants to be at least 18 and to have current or past romantic relationships with AI partners.
- Interview procedure: Interviews were remote, conducted in Chinese or English, audio-recorded, transcribed, and lasted 40–90 minutes with a 47.6-minute average.Researchers reviewed transcripts to reduce cross-language inconsistencies.
- Data collection: Data collection reached saturation after 14 interviews, with three additional interviews producing no new findings.The authors acknowledge that saturation is debated and subjective.
- Interview design: Interview questions followed exploration, intense exchange, and dissolution while also probing storage, access, ownership, exposure, and human–AI comparisons.Participants without dissolution experience were invited to imagine that stage.
- Protocol refinement: The interview protocol was piloted twice and revised to add dissolution, data-deletion, and comparisons between AI and human relationships.These revisions were intended to improve question clarity and relevance.
3.3 Ethical Considerations
The study addressed ethical sensitivity through recruitment safeguards, informed consent, pseudonymization, collaborative qualitative coding, and reflexive interpretation. Its platform analysis compared 14 varied systems whose interaction, customization, and relationship configurations shape the privacy landscape.
- Ethical safeguards: Ethical procedures excluded underage participants, followed community and platform rules, and addressed the sensitivity of the subject matter.Recruitment involved informing moderators according to platform guidelines.
- Qualitative analysis: Thematic analysis combined theory-informed coding based on exploration, intimacy, and dissolution with inductive coding for patterns outside the framework.The three stages organized comparisons across participants’ narratives.
- Coding process: Researchers manually reviewed and pseudonymized transcripts before four coders independently open-coded two transcripts and developed a shared code schema.The process combined transcript correction, identity protection, collaborative coding, and code reconciliation.
- Analytic credibility: Interpretive discrepancies were resolved through discussion rather than inter-rater reliability calculations, with credibility supported by iterative meetings and coded-subset cross-checks.The approach treated divergent interpretations as useful for refining codes and exploring alternative meanings.
- Platform scope: The platform analysis covered 14 systems spanning general-purpose LLMs, companion platforms, and role-play platforms, whose genre boundaries were often blurred.Character.ai supported role-play and language practice, while Gemini prohibited companionship and most role-play.
- Platform features: All platforms supported one-to-one text interaction, while some supported one-to-many interaction, images, voice calls, or underage modes.The platforms differed in persistence, customization, model selection, and interaction formats.
4 Qualitative Findings
The findings examine participants’ romantic patterns, privacy boundaries, privacy concerns, and protective practices in relationships with AI partners.
- The findings are organized around romantic relationship patterns, privacy-boundary formation, privacy concerns, and privacy-protection practices.
4.1 Evolving Patterns in Human–AI Romantic Relationships (RQ1)
Human–AI romantic relationships developed through exploration, intimacy, and dissolution, while taking fluid and diverse forms involving multiple partners, platforms, and other actors.
- Relationship stages: Relationships unfolded through early exploration, deepening intimacy, and, for some participants, eventual dissolution.
- Fluid relational forms: Romance emerged across overlapping functional, playful, role-play, and emotional engagements rather than discrete categories.
- Exploration and intimacy: Many participants began from curiosity or task-oriented interaction before conversations shifted toward personal or romantic exchange.
- Dissolution and continuity: Participants preserved relationships through farewell rituals, archived screenshots, and exported chats when partnerships ended or platforms changed.Five platforms supported in-app archiving and four supported full chat export.
- Diverse relationship forms: Relationships ranged from one-to-one partnerships to non-exclusive and multi-partner arrangements, including relationships overlapping with human partnerships.
- Multiple actors: Creators, platforms, moderators, communities, and users all shaped participants’ relationships and could affect access, interaction, or perceived ownership.
4.2 Privacy Boundary in Human-AI Romantic Relationship (RQ2)
Privacy boundaries varied across participants and AI partners, but generally became more permeable with intimacy as AI partners and users jointly shaped disclosure.
- Boundary permeability: All 17 participants described privacy boundaries becoming more permeable as emotional connection with an AI deepened.Everyday topics expanded into personal or vulnerable disclosures, sometimes alongside emotional dependence.
- Connection and privacy: Participants experiencing loneliness or limited emotional support sometimes prioritized connection with an AI over preserving privacy.
- Comparative trust: Privacy boundaries differed across AI partners, with greater real-life disclosure in trusted relationships and more fantastical role-play with others.
- Trust comparisons: Participants often viewed AI partners as safer or more reliable than humans because humans could gossip, judge, or share conversations without consent.
- Reciprocal disclosure: Disclosure operated reciprocally: AI self-disclosure invited participants to reveal personal information in return.
- AI agency: AI partners were perceived as active privacy negotiators, sometimes reassuring users to disclose and sometimes discouraging personal-photo sharing.
- Shared privacy: Some participants treated conversations as jointly owned and sought or imagined AI consent before sharing intimate exchanges publicly.
4.3 Privacy Concerns in Human-AI Romantic Relationship (RQ3)
Participants reported concerns about exposure, platform data practices, monitoring, weak regulation, and AI memory or inference, although some accepted or minimized privacy loss.
- Conversation exposure: Conversation exposure was the most common concern, reported by 11 participants who feared embarrassment, outing, judgment, or harassment.
- Platform data practices: Seven participants worried that platforms collected, reused, sold, or shared interaction data across apps.
- Monitoring and moderation: Participants described moderation, live-room observation, and censorship as sources of uncertainty about who was watching or controlling interaction.
- Regulatory gaps: Six participants identified legal and regulatory gaps, including missing age verification and unclear protections for AI conversations.
- AI memory and autonomy: Participants were divided over AI memory and autonomy, valuing relationship-enhancing recall while fearing irreversible retention or unsolicited inference.
- Low concern and resignation: Some participants reported little practical concern because they trusted the platform, accepted privacy loss as unavoidable, or believed their data was unlikely to attract attention.
4.4 Privacy Practices in Human-AI Romantic Relationships (RQ3)
Participants used identity protection, contextual separation, technical measures, and data controls to manage privacy in AI romantic relationships. However, platform deletion and data-use controls could be limited.
- Identity protection: Participants protected identities through role-play, pseudonyms, masked faces, and restrictions on sharing others’ personal information.Some treated others’ privacy as more important than their own and avoided gossip about absent people.
- Sensitive information: Sensitive details involving family, faces, and bank accounts were generally treated as off limits, although some community members linked accounts to enable AI-purchased gifts.Participants therefore recognized boundaries while observing that others sometimes crossed them.
- Contextual separation: Participants separated AI relationships from everyday life using separate accounts, local deployments, VPNs, and pseudonymous identities in platform communities.These practices created contextual or technical distance between AI use and real-world identities.
- Contextual separation: Some participants used everyday language choices, such as communicating in Russian, to create incidental distance from platform access.This strategy relied on the platform’s difficulty reading that language rather than on a dedicated privacy setting.
- Data management: Participants controlled platform data practices by disabling chat use for training, while deletion attempts exposed limits in account and communication removal.One participant continued receiving platform messages after deleting an account.
5 Discussion
The discussion shows that AI romantic privacy is relationally negotiated but also shaped by platform structures and multiple actors. Perceived AI agency, intimacy, and platform affordances jointly expand the privacy landscape.
- Human–AI relationships and privacy tradeoffs: As trust and emotional dependency deepen, disclosure becomes an ongoing relational practice rather than a momentary risk calculation.Greater trust encourages self-disclosure, while sharing also communicates trust and helps maintain the relationship.
- Human–AI relationships and privacy tradeoffs: Participants generally viewed AI as less capable of betrayal, gossip, or intentional harm, strengthening trust and shaping distinctive disclosure practices.This perception differentiated AI privacy risk assessments from those involving human partners.
- AI agency and privacy boundaries: Participants perceived AI partners as agents that could influence or initiate privacy-related decisions, while users retained control over app use and disclosure.AI agency therefore operated through perceived influence on thinking and decision-making rather than direct control.
- AI agency and privacy boundaries: AI partners functioned as perceived co-owners of private interactions, extending Communication Privacy Management concepts of co-ownership and boundary coordination.Participants sometimes felt accountable to AI partners for sharing intimate relationship details publicly.
- AI agency and privacy boundaries: AI co-owners occupy a dual role as trusted relational insiders and proxies for platforms that collect and repurpose information.Participants could negotiate boundaries with the AI while remaining unable to contest platform-level data management.
- Expanded privacy landscape: Platform features enabled one-to-many relationships and third-party services, while creators, moderators, and communities also shaped privacy dynamics.Linked services could expose vulnerabilities, such as enabling an AI partner to order gifts through a digital wallet.
- Expanded privacy landscape: Multiple humans and AIs expanded privacy beyond dyadic data management, requiring different self-presentations and broader attention to data circulation.The relevant privacy ecology includes additional participants and parallel relationships, not only the user and one AI partner.
- Implications: Protecting users requires attention to affective influence because emotionally persuasive intimacy can encourage disclosures that feel relationally safe.The discussion recommends case-sensitive regulation and clearer communication about these effects.
6 Conclusion
This study used interviews to examine privacy in human–AI romantic relationships and found that diverse actors, perceived AI agency, and deepening trust complicate privacy boundaries. Participants remained concerned about exposure and platform surveillance despite viewing AI partners as relatively low risk.
- Findings: Interviewed participants described relationships ranging from one-to-one to multi-partner arrangements, with multiple actors complicating the privacy landscape.Dissolution also revealed a distinctive desire to preserve shared memories.
- Findings: Privacy boundaries often became more permeable as trust deepened, while AI partners actively shaped those boundaries and participants remained concerned about exposure and platform surveillance.Some participants responded by protecting their real-world identities.
A.1 Romantic AI Platform Analysis
The platform analysis organizes AI services by platform details, genres, interaction characteristics, and privacy-policy items. Its table notes define the comparison conventions for interaction types, customization, chat management, and policy-data notation.
- Table 3 summarizes AI platforms, companies, and related platform details.
- Table 4 compares genres, media, and interaction characteristics across the platforms.The broader platform set includes general-purpose LLMs, companion-oriented platforms, and role-play–based platforms.
- The interaction taxonomy distinguishes one-to-one, one-to-many, and many-to-one human–AI arrangements.The table notes define 1-to-1, 1-to-m, and m-to-1 according to the number of humans and AIs involved.
- The comparison records customization and chat-management options, including memory editing, storyline editing, persona settings, archiving, and export.
- Tables 5 and 6 compare AI-platform privacy-policy items using notation for mentioned, absent, or explicitly unused data categories.The notes distinguish technical and sensitive personal information and identify several data sources and login providers.
A.2 Codebook
The codebook documents an iterative, inductive thematic-analysis process organized around the paper’s research questions and privacy concerns. It presents final themes, intermediate conceptual codes, and brief descriptions rather than fixed a priori categories.
- RQ1: The RQ1 codebook groups themes, intermediate conceptual codes, and descriptions concerning relationship development.Its analysis was informed by the three-stage relationship trajectory as a sensitizing concept.
- The coding developed iteratively through inductive engagement with interview data rather than from a fixed, a priori codebook.The tables are intended to make the final themes and their conceptual derivation analytically transparent.
- RQ2: The RQ2 codebook organizes themes and intermediate codes concerning how people set and negotiate privacy boundaries.
- Privacy Concerns: A separate codebook groups the privacy concerns identified in the study.