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

Large language models (LLMs) are increasingly used as collaborative partners in writing. However, this raises a critical challenge of authorship, as users and models jointly shape text across interaction turns. Understanding authorship in this context requires examining users’ evolving internal states during collaboration, particularly self-efficacy and trust. Yet, the dynamics of these states and their associations with users’ prompting strategies and authorship outcomes remain underexplored. We examined these dynamics through a study of 302 participants in LLM-assisted writing, capturing interaction logs and turn-by-turn self-efficacy and trust ratings. Our analysis showed that collaboration generally decreased users’ self-efficacy while increasing trust. Participants who lost self-efficacy were more likely to ask the LLM to edit their work directly, whereas those who recovered self-efficacy requested more review and feedback. Furthermore, participants with stable self-efficacy showed higher actual and perceived authorship of the final text. Based on these findings, we propose design implications for understanding and supporting authorship in human-LLM collaboration.

💡 Research Summary

Summary
The paper “Authorship Drift: How Self‑Efficacy and Trust Evolve During LLM‑Assisted Writing” investigates how users’ internal psychological states—self‑efficacy (belief in one’s ability to complete a writing task independently) and trust (belief that the LLM will reliably support the task)—change over the course of a multi‑turn interaction with a large language model, and how these dynamics relate to prompting behavior and perceived/actual authorship.

Study Design
A total of 302 native‑English participants (after quality control) were recruited via Prolific. Each participant wrote an argumentative essay using a GPT‑4‑based LLM in a web‑based chat interface. After every interaction turn, participants rated their current self‑efficacy and trust on a 5‑point Likert scale. The system logged all prompts, LLM responses, and the evolving draft. Essays were scored using the College Board rubric (automatically generated scores were manually verified). The top 10 % of essays received a performance bonus to encourage genuine effort.

Research Questions

1. What trajectory patterns emerge for self‑efficacy and trust, and how do they co‑evolve?
2. How are prompting strategies associated with those trajectories?
3. How do actual (objective contribution) and perceived (subjective ownership) authorship relate to the trajectories?

Key Findings

1. Dynamic Trajectories – Across participants, self‑efficacy tended to decline over turns while trust rose. Growth‑curve modeling identified three main self‑efficacy trajectories: (a) steady decline, (b) initial dip followed by recovery, and (c) stable high levels. Trust showed a relatively uniform upward slope across all groups.

2. Prompting Strategies – Participants whose self‑efficacy continuously dropped were more likely to issue “draft‑to‑edit” or “edit‑to‑edit” prompts, essentially asking the LLM to rewrite or polish their text directly. Those who experienced an early dip but later recovered tended to request “review” or “feedback” prompts, using the model as a consultant rather than a primary author.

3. Authorship Outcomes – Authorship was measured in two ways: (i) Actual authorship – the proportion of text attributable to the participant, computed via sentence‑level similarity between the final essay and the LLM’s contributions; (ii) Perceived authorship – participants’ self‑reported sense of ownership after the task. The “stable self‑efficacy” group scored highest on both dimensions, indicating that maintaining confidence supports both objective contribution and subjective ownership. Conversely, the “declining self‑efficacy” group showed lower actual contribution and felt less ownership.

Design Implications

• Detect Vulnerable Moments: Interfaces should monitor self‑efficacy signals (e.g., rapid rating drops) and, when detected, surface review/feedback tools rather than direct edit commands, encouraging users to stay cognitively engaged.
• Transparency & Calibration: As trust naturally rises, systems should continue to provide provenance information (e.g., highlight AI‑generated sentences) to prevent over‑reliance and preserve user agency.
• Support Self‑Efficacy Recovery: Offer reflective feedback loops (e.g., side‑by‑side comparison of user draft vs. AI suggestion) that let users see how their input shapes the output, thereby rebuilding confidence.

Limitations & Future Work

The study is confined to English argumentative essays and a single GPT‑4‑style model, limiting generalizability to other genres, languages, or model families. Self‑efficacy and trust were each captured by a single Likert item per turn, which may overlook nuanced facets such as anxiety or perceived control. Future research should explore multi‑modal tasks, cross‑cultural samples, and richer psychometric instruments, as well as adaptive UI prototypes that dynamically adjust assistance based on real‑time psychological state monitoring.

Conclusion
By tracking turn‑by‑turn self‑efficacy and trust, the authors reveal that authorship “drift” is not merely a function of how much text the model generates, but is tightly linked to users’ evolving confidence and reliance. Designing LLM‑assisted writing tools that recognize and respond to these psychological dynamics can help preserve human authorship while still leveraging the productivity gains of advanced language models.