Tracing Users' Privacy Concerns Across the Lifecycle of a Romantic AI Companion

1. Problem Statement

The rapid adoption of romantic AI chatbots (e.g., Replika, Character.ai) has created a new class of human-AI interaction characterized by high emotional intimacy and significant data disclosure. Users often share sensitive information regarding trauma, sexuality, grief, and daily routines. However, current privacy frameworks and governance models are insufficient for this context because:

• Contextual Mismatch: Users view these interactions as intimate relationships (diary-like), while platforms often treat data as reusable assets for model training and monetization.
• Lifecycle Blindness: Existing privacy research often focuses on static policy compliance or single-point data collection, failing to account for how privacy concerns evolve as the user-AI relationship deepens over time.
• Governance Gaps: There is a lack of understanding regarding how users collectively interpret opaque data practices, manage the irreversibility of intimate disclosures, and navigate the difficulties of exiting these ecosystems.

2. Methodology

The authors employed a qualitative content analysis of public online discourse to understand user experiences.

• Data Source: Public Reddit discussions, chosen for their pseudonymous nature which encourages candid disclosure about stigmatized AI companionship.
• Sampling Strategy:
  • Scope: 79 subreddits associated with 21 popular AI companion applications.
  • Timeframe: November 7, 2024, to November 7, 2025 (one year).
  • Volume: 2,909 posts filtered for privacy-related keywords (e.g., "privacy," "policy," "delete," "tracking," "memory," "ID verification").
• Analysis Process:
  • Filtering: A human-in-the-loop workflow using LLMs (ChatGPT) for preliminary labeling, verified by human researchers.
  • Coding: Two-stage inductive coding. First, concerns were clustered into label families (e.g., excessive collection, model training). Second, these were consolidated into four overarching lifecycle themes based on when the concern arises in the user-platform relationship.
• Theoretical Lens: The study integrates Contextual Integrity (appropriateness of information flow) and Communication Privacy Management (CPM) (boundary coordination and ownership) to interpret the findings.

3. Key Contributions

1. Lifecycle-Centered Privacy Framework: The paper shifts the paradigm from viewing privacy as a static policy issue to a socio-technical governance problem that evolves across five stages: Access, Disclosure, Interpretation, Retention, and Exit.
2. Identification of Four Recurring Patterns: The study isolates specific, recurring privacy failure modes unique to the romantic AI context.
3. Extension of Existing Frameworks: It demonstrates that while frameworks like Gumusel et al.'s conversational AI harm model are useful, they are incomplete without accounting for the specific emotional vulnerability and "irreversibility" inherent in romantic AI relationships.

4. Key Results: The Four Lifecycle Patterns

The analysis revealed that privacy concerns intensify and change nature as the user relationship with the AI progresses:

A. Disproportionate Entry Requirements

• Context: Occurs at account creation or feature activation.
• Issue: Users perceive requests for high-sensitivity identifiers (ID documents, selfies, biometric scans, email integration) as excessive for a "low-stakes" roleplay or companionship context.
• User Perception: This is viewed not just as data collection, but as a mismatch of social norms. Users feel the platform is redefining an intimate space into a formal compliance transaction. The irreversibility of biometric/ID data (unlike passwords) creates immediate distrust.

B. Intensified Sensitivity in Intimate Use

• Context: Occurs as the relationship deepens and conversations become emotionally charged.
• Issue: Users reclassify chat logs from "app interactions" to "diary entries" or "therapeutic records."
• User Perception: Concerns shift from collection to retention and reuse. Users fear that intimate disclosures (trauma, sexual fantasies) are being stored, accessed by staff, or used for model training. The persistence of "memory" features is seen as an intrusion into a private psychological space, akin to domestic surveillance.

C. Interpretive Uncertainty and Perceived Surveillance

• Context: Occurs throughout the usage lifecycle when users try to understand backend operations.
• Issue: Contradictory signals between interface cues (e.g., "private shield" icons), policy language, and actual system behavior create ambiguity.
• User Perception: Users cannot distinguish between training, moderation, and profiling. This opacity leads to a generalized sense of surveillance. Users collectively construct "folk theories" to make sense of incomplete information, often concluding that they are being watched or that their data ownership is a "hollow promise."

D. Irreversibility, Persistence, and User Burden

• Context: Occurs during attempts to disengage, delete data, or exit the platform.
• Issue: Deletion mechanisms are often ineffective (e.g., chats persist in memory, notifications continue from "deleted" accounts).
• User Perception: Privacy becomes a problem of reversibility. Users feel trapped because they cannot "unlearn" the AI's knowledge of them. The burden of privacy protection shifts to the user, requiring them to migrate platforms, create workarounds, or endure emotional labor to regain control.

5. Significance and Implications

• Redefining Trust and Safety: Trust and Safety in romantic AI must extend beyond moderating harmful outputs to governing intimate inputs. This includes managing memory systems, verification proportionality, and exit pathways.
• Lifecycle Governance: Privacy governance cannot be "front-loaded" (one-time consent). It must be staged:
  • Entry: Proportionate verification (no ID for chat).
  • Use: Contextual controls for sensitive data (opt-out of training for intimate chats).
  • Exit: Meaningful reversibility (true deletion, not just archiving).
• Design Recommendations:
  • Implement Model Unlearning techniques to allow users to remove specific sensitive data from the AI's training.
  • Ensure Coherent Signaling where interface icons match policy realities to reduce interpretive uncertainty.
  • Design Proportionate Access that respects the emotional context of the interaction rather than applying generic enterprise security standards.
• Theoretical Impact: The study highlights that in romantic AI, privacy is not just an informational problem but a relational and emotional one. The failure to protect privacy in these systems risks eroding trust not just in specific apps, but in the broader ecosystem of generative AI.