Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication

Paper Title

"Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication"

Publication Info

• Topic area: AI-assisted communication in romantic relationships, focusing on authorship and authenticity.
• Keywords: AI assistance, romantic communication, authorship, authenticity, trust, disclosure, apologies, boundary requests, human-AI collaboration, interpersonal communication.

Background and Problem

• Problem / challenge: The increasing use of AI tools in intimate communication raises concerns about perceived authorship, effort, and authenticity, especially in morally sensitive contexts like apologies and boundary-setting.
• Significance: Understanding how AI assistance and disclosure affect relational outcomes is critical for designing tools that preserve trust and authenticity in close relationships.
• Motivation and related work: Prior research highlights tensions between AI-mediated communication and perceptions of sincerity, but lacks scenario-specific evidence on how AI assistance and disclosure interact to shape trust and relational outcomes. This study addresses gaps in causal evidence, mechanisms, and sender-receiver miscalibration.

Solution

• Proposed approach: A scenario-aware attributional framework that examines how AI assistance (light vs. heavy) and disclosure (co-sign vs. none) influence perceived ownership, authenticity, and relational outcomes in romantic communication.
• Novelty:
  1. Scenario-specific causal estimates of AI assistance and disclosure on ownership, authenticity, trust, and outcomes.
  2. A mechanism linking help and disclosure to relational outcomes via ownership and authenticity.
  3. Evidence of sender-receiver miscalibration in AI-assisted communication.
  4. Design guidance for preserving voice, restoring ownership, and aligning disclosure with context.
• Procedure and key techniques:
  • Study 1: Observed sender behavior (N = 152) in composing AI-assisted messages for apologies and boundary requests, generating a curated message corpus.
  • Study 2: Conducted a 2 × 2 randomized experiment (N = 704) to test receiver judgments of messages under varying help and disclosure conditions.
  • Analyzed relationships between authoring telemetry (e.g., edit distance, idiosyncrasy cues) and receiver perceptions.

Results

• Concrete findings:
  • Heavier AI assistance (full drafts) reduced perceived ownership and authenticity, while light assistance (tone rewrites) preserved these qualities.
  • Disclosure effects were scenario-dependent: co-signing light help in apologies increased authenticity and forgiveness, but co-signing heavy help reduced both.
  • In boundary requests, co-signing had neutral to mildly negative effects, regardless of help level.
  • Personalization cues (e.g., shared memories, nicknames) enhanced ownership and authenticity, while greater drafting distance had the opposite effect.
• Advantage over baselines:
  • Light assistance with co-signing improved relational outcomes in apologies compared to heavy assistance or no disclosure.
  • Competence gains from AI assistance did not compensate for authenticity losses.
• Experiments / evaluation:
  • Study 1: Mixed design with sender-side telemetry and self-reports, focusing on authoring behavior and anticipated partner reactions.
  • Study 2: Mixed-effects models with receiver-side judgments of curated messages, testing the mechanism from help/disclosure to outcomes.
  • Metrics: Ownership, authenticity, trust facets (integrity/benevolence vs. competence/clarity), forgiveness, compliance.
• Limitations and future work:
  • Independent raters in Study 2 lacked dyad-specific context, limiting generalizability to real couples.
  • Focused on two scenarios (apologies and boundary requests); future work should explore other relational acts, cultural contexts, and synchronous modalities.
  • Disclosure manipulation tested only one framing; future studies should vary framing, placement, and specificity.

Summary

This research demonstrates that AI assistance and disclosure reshape the social meaning of romantic messages by altering attributions of authorship and effort. Heavier assistance reduces perceived ownership and authenticity, while light assistance preserves these qualities. Disclosure effects depend on the scenario: co-signing light help in apologies enhances authenticity, but co-signing heavy help harms it. Personalization cues and restrained system drafting are key to maintaining authenticity. The findings inform design principles for AI tools in intimate communication, emphasizing the need to preserve voice, restore ownership, and tailor disclosure to context. Future work should extend these insights to broader relational scenarios and real-world dyadic interactions.