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

Abstract: 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.

• The paper reveals that self-efficacy declines (by 0.115 Likert points per turn) while trust increases (by 0.111 points per turn) during LLM-assisted writing sessions.
• It employs within-person change metrics and mixed-effects regression on data from 302 participants to analyze the behavioral correlates of editing and prompt transitions.
• Findings indicate that decreased self-efficacy is linked to greater LLM content adoption and diminished authorial agency, prompting design recommendations for preserving user ownership.

Authorship Drift in LLM-Assisted Writing: Dynamics of Self-Efficacy and Trust

Study Rationale and Methodology

The paper "Authorship Drift: How Self-Efficacy and Trust Evolve During LLM-Assisted Writing" (2602.05819) provides a rigorous empirical investigation into the temporal dynamics of self-efficacy and trust during human–LLM collaborative writing. The central hypothesis posits that these psychological constructs are not static but evolve throughout multi-turn LLM-assisted tasks, ultimately influencing both process-level behaviors and authorship outcomes.

To control for individual differences in scale interpretation, the authors use within-person change metrics and minimal detectable change thresholds (MDC) to categorize self-efficacy and trust trajectories into stable, increase, decrease, recovery, and reversion patterns. The experimental protocol involves 302 English-native, LLM-experienced participants sourced from Prolific, each tasked with a timed argumentative essay in a controlled, instrumented web platform.

Per interaction turn, participants must provide self-efficacy and trust ratings, operationalized as momentary beliefs in ability to complete the task unaided and in the LLM's reliability, respectively. Rich interaction logs are captured, including prompts, LLM responses, and essay drafts. Annotation of prompts enables segmentation into intention categories: drafting, editing, ideating, information searching, and reviewing.

Figure 1: The study's writing interface with left-panel LLM chat and right-panel essay editor, supporting segmented multi-turn interaction and intervention capture.

Trajectory Patterns of Self-Efficacy and Trust

The data reveal that 56.6% of participants maintain stable self-efficacy, 27.5% experience a reliable decrease, and 10.9% demonstrate recovery after an initial decline. Contrastingly, trust is predominately stable (76.5%), with 15.6% exhibiting a significant increase, notably among participants starting with low baseline trust.

Longitudinal modeling using linear mixed-effects regression identifies that, on average, self-efficacy decreases by 0.115 Likert points per turn while trust increases by 0.111 points per turn. Critically, higher trust buffers against self-efficacy decline ($\beta_\text{{turn × trust}}=0.038$, $p<0.001$), but the converse interaction is negligible; trust increases are more rapid for participants with higher self-efficacy but not predicted by self-efficacy alone.

Figure 2: Stable self-efficacy trajectories highlight a large median cohort preserving confidence throughout the LLM-assisted writing session.

Behavioral Correlates: Prompting Strategies and Transition Patterns

Prompt analysis exposes a strong behavioral linkage: participants exhibiting decreasing self-efficacy are statistically more likely to request direct LLM editing (Kruskal–Wallis $H=13.57$, $p<0.001$) and engage in drafting–editing or editing–editing prompt transitions. In contrast, those with recovering self-efficacy shift towards review-oriented prompts, requesting feedback over direct revision.

Figure 3: Editing prompt usage is disproportionately high in participants with decreasing self-efficacy, underscoring a shift towards LLM-driven revision agency.

Figure 4: Drafting-to-editing transitions are a key marker of self-efficacy decline, reflecting relinquishment of authorial control in iterative revision cycles.

Prompt intention also interacts with trust trajectories: participants with increasing trust demonstrate elevated frequencies of information-searching prompts (Mann–Whitney $p<0.05$), reflecting a more exploratory, less outcome-focused engagement with the LLM.

Figure 5: Information searching prompts are more prevalent in trust-increase trajectories, linking epistemic seeking behaviors to trust calibration.

Turn-level analyses further substantiate these findings: editing and drafting correspond with lower self-efficacy ratings (Fig. 7, in appendix), while editing and reviewing associate with higher trust (Fig. 8, in appendix).

Authorship Outcomes: Actual and Perceived

Actual authorship is objectively quantified via lexical overlap (ROUGE-3 recall) and semantic similarity (embedding cosine similarity) between the final essay and LLM outputs. Perceived authorship is captured through post hoc agency and ownership ratings. Notably, participants with decreasing self-efficacy evidence statistically greater LLM content adoption in both lexical and semantic measures and report diminished agency and ownership ($p<0.01$ for overlap, $p<0.001$ for agency).

Figure 6: Lexical overlap is highest among participants with decreasing self-efficacy, denoting heightened LLM content adoption and reduced textual originality.

This group also manifests the most pronounced authorship drift: both objective and subjective measures demonstrate a clear hierarchy (stable > recovery > decrease), directly tying self-efficacy trajectories to authorship integrity. No statistically meaningful differences in authorship outcomes are observable across trust trajectories.

Theoretical Implications and Design Recommendations

The decoupling of self-efficacy and trust trajectories exemplifies how LLMs, when perceived as reliable, simultaneously induce agency drift via social comparison and editorial delegation. Trust in the LLM calibrates quickly via repeated exposure to fluent outputs, but self-efficacy is systematically eroded, especially when users engage in direct editing delegation, thus relinquishing decision authority. This phenomenon aligns with educational and HCI literature emphasizing that agentic feedback—not corrective automation—supports robust self-efficacy and authentic authorship.

The findings demand a reconceptualization of self-efficacy as a dynamic state, not a static trait. Designs for future writing assistants should employ unobtrusive behavioral inference (e.g., prompt classification, keystroke logging) to identify vulnerable moments of self-efficacy erosion and trigger interventions supporting agency, such as promoting review/feedback over direct edits or visualizing collaboration history to enhance metacognitive reflection.

The results strongly imply that LLMs' increasing capabilities in human-AI writing contexts are nontrivially linked to authorship drift and agency loss, even when users self-report high initial competence. In educational or high-stakes creative scenarios, this exposes risks to skill preservation, ownership perception, and critical engagement, calling for longitudinal studies and system-level affordances that sustain user authorship and self-efficacy.

Conclusion

This paper offers a detailed, evidence-based account of the divergence between self-efficacy and trust in LLM-assisted writing. It exposes a robust relationship between declining self-efficacy, increased editorial delegation to LLMs, and measurable loss of authorship both perceptually and objectively. The theoretical and practical implications point toward the necessity of process-aware, trajectory-sensitive interventions in design and evaluation of human–AI collaborative systems to counteract subtle forms of agency erosion and preserve long-term user engagement and ownership.