Authorship Drift: How Self-Efficacy and Trust Evolve During LLM-Assisted Writing

Paper Title

Authorship Drift: How Self-Efficacy and Trust Evolve During LLM-Assisted Writing

Publication Info

• Topic area: Dynamics of user psychology and authorship in LLM-assisted writing.
• Keywords: Large language models, self-efficacy, trust, authorship, human-AI collaboration, prompting strategies, user behavior, writing assistance, agency, ownership.

Background and Problem

• Problem / challenge: LLM-assisted writing blurs the boundaries of authorship, potentially diminishing users' sense of agency and ownership. Existing research often treats self-efficacy and trust as static constructs, failing to capture their dynamic evolution during multi-turn interactions.
• Significance: Understanding how self-efficacy and trust evolve is critical to preserving user authorship and agency in LLM-assisted writing, which has implications for education, creativity, and human-AI collaboration.
• Motivation and related work: Prior studies have shown the benefits of LLMs in enhancing productivity and creativity but have not systematically explored the dynamic interplay between users' self-efficacy, trust, and authorship. This paper addresses this gap by examining turn-level changes in these constructs and their behavioral correlates.

Solution

• Proposed approach: An empirical study with 302 participants to analyze turn-by-turn dynamics of self-efficacy and trust during LLM-assisted argumentative essay writing, along with their associations with prompting strategies and authorship outcomes.
• Novelty:
  1. Empirical characterization of self-efficacy and trust trajectories in LLM-assisted writing.
  2. Identification of relationships between self-efficacy patterns, prompting strategies, and authorship outcomes.
  3. Design implications for supporting user authorship in human-LLM collaboration.
• Procedure and key techniques:
  1. Participants completed an argumentative essay task using an LLM, providing self-efficacy and trust ratings after each interaction turn.
  2. Prompts were categorized into intentions (e.g., drafting, editing, reviewing).
  3. Trajectories of self-efficacy and trust were classified into patterns (e.g., stable, decrease, recovery).
  4. Actual authorship was measured via lexical overlap and semantic similarity, while perceived authorship was assessed through post-survey ratings of ownership and agency.

Results

• Concrete findings:
  • Self-efficacy generally decreased over turns (−0.115 points/turn), while trust increased (+0.111 points/turn).
  • Users with decreasing self-efficacy relied more on editing prompts and exhibited lower actual and perceived authorship.
  • Recovery in self-efficacy was associated with increased use of review prompts and intermediate authorship outcomes.
  • Trust increases were more common among users with lower initial trust and were associated with information-seeking prompts.
• Advantage over baselines: The study provides a dynamic, turn-level analysis of self-efficacy and trust, revealing nuanced patterns that static, pre–post measures cannot capture.
• Experiments / evaluation:
  • 302 participants recruited via Prolific, filtered for English proficiency and recent LLM use.
  • Metrics: Likert-scale ratings for self-efficacy and trust, proportions of prompt categories, lexical overlap, semantic similarity, and post-survey authorship ratings.
  • Statistical analyses: Mixed-effects models, Kruskal-Wallis tests, and post-hoc comparisons.
• Limitations and future work:
  • Self-reported ratings may have influenced natural interaction patterns.
  • Study confined to single-session interactions; longitudinal effects remain unexplored.
  • Limited to one LLM configuration; future work should examine diverse systems and user groups.

Summary

This paper investigates how self-efficacy and trust evolve during LLM-assisted writing, revealing that self-efficacy tends to decline while trust increases over interaction turns. Users with decreasing self-efficacy relied more on editing prompts and showed diminished authorship, while those with recovering self-efficacy engaged in review-oriented interactions with intermediate authorship outcomes. Trust increases were linked to information-seeking behaviors, particularly among users with lower initial trust. These findings highlight the dynamic nature of self-efficacy and trust, emphasizing the need for systems that support user authorship through reflective interactions and calibrated trust. Future work should explore longitudinal dynamics and less intrusive measures of user perceptions.