Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat
Sign Up
Cited by 8Innovative Human-Technology InteractionHuman-Automation Interaction and SafetyUsability and User Interface Design

The DOI matches its Crossref record; 100 of the 100 references it cites were found in a registry or an index; none of the 74 sources its claims cite is checked for support yet.

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

Open access
Park, Yeon Su, Arvi, Nadia Azzahra Putri, Kim, Seoyoung, Kim, Juho·2026·Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems
DOI 10.1145/3772318.3790276
Open this paper in CoChat — free account

Abstract

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.

Overview

CoChat Summary

TL;DR

This study tracked turn-by-turn self-efficacy and trust in 302 participants completing LLM-assisted argumentative essays, finding that self-efficacy generally decreased while trust increased. Stable self-efficacy was associated with higher actual and perceived authorship, whereas decreasing self-efficacy was linked to more editing prompts and lower authorship.

Key findings

  • Across 302 participants and 1,410 prompt–response pairs, self-efficacy decreased by 0.115 points per turn (p < 0.001) and trust increased by 0.111 points per turn (p < 0.001); higher trust buffered self-efficacy decline (turn × trust β = 0.038, p < 0.001).
  • Self-efficacy trajectories were 56.6% stable, 27.5% decrease, 10.9% recovery, 4.0% increase, and 1.0% reversion; trust trajectories were 76.5% stable, 15.6% increase, 5.3% recovery, 1.7% reversion, and 1.0% decrease.
  • The decrease self-efficacy group used a higher proportion of editing prompts than the stable group (p_adj < 0.001) and showed more drafting–editing and editing–editing transitions (p_adj < 0.001 and p_adj < 0.01); the recovery group used more reviewing prompts than the decrease group (p_adj < 0.05).
  • The decrease trajectory had higher lexical overlap and semantic similarity with LLM-original content than stable (p_adj < 0.01 for both) and lower perceived ownership and agency; stable > recovery > decrease for ownership, and stable had higher agency than recovery and decrease. Trust trajectory patterns showed no significant differences across authorship measures (p > 0.05).

Methods

  • Online experiment with 302 final participants (313 recruited, 11 excluded) from Prolific: native English speakers, approval rating ≥95%, and LLM use within the last month; participants wrote a minimum 300-word argumentative essay within 30 minutes using an embedded GPT-4.1 chat and text editor.
  • Data included pre-survey baseline self-efficacy/trust, turn-by-turn 7-point self-efficacy and trust ratings before each prompt, post-survey ownership and agency, and logs of prompts, LLM responses, essay states, and timestamps.
  • Analysis used a 2-point Minimal Detectable Change threshold to classify trajectories; linear mixed-effects models with random intercepts and slopes; qualitative prompt intention coding (κ = 0.867 and 0.789); ROUGE-3 lexical overlap and gemini-embedding-001 semantic similarity for actual authorship; and Kruskal-Wallis, Mann-Whitney U, and Dunn tests.

Limitations

  • Stated: turn-by-turn self-reported ratings may have influenced natural interaction; 76.7% reported no effect, but 19.2% found the ratings distracting or made them more mindful, and logging was limited to in-platform events, so external tool use could not be determined.
  • Stated: the study identifies empirical patterns but not qualitative reasons behind them; it was confined to a single session, did not account for individual or system factors such as task proficiency or LLM tone, and used a single LLM.
  • Evident from the text: the appendix turn-level analysis aggregates across interactions without normalizing for varying numbers of turns per participant, so highly active users may contribute proportionally more data points; the authors call these associations exploratory rather than definitive.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3772318.3790276

Cites

100 of 100 references verified0 citation errors · 0 not checkable

Statistics

Figures

Image 1
CC-BY · From the PDF
Figure 2 Normalized overlays of self-efficacy and trust trajectories for the five major trajectory patterns. The horizontal axis shows the time-normalized turn-by-turn progression from initial pre-survey rating (left) to final rating (right). The vertical axis shows the relative change in Likert-scale points after zero-centering all trajectories at each participant's pre-survey rating. Thin translucent lines represent individual participants, the solid line shows the group median, and the shaded band indicates the interquartile range.
Figure 2 Normalized overlays of self-efficacy and trust trajectories for the five major trajectory patterns. The horizontal axis shows the time-normalized turn-by-turn progression from initial pre-survey rating (left) to final rating (right). The vertical axis shows the relative change in Likert-scale points after zero-centering all trajectories at each participant's pre-survey rating. Thin translucent lines represent individual participants, the solid line shows the group median, and the shaded band indicates the interquartile range.CC-BY · From the PDF
Image 3
CC-BY · From the PDF
Image 4
CC-BY · From the PDF
Figure 4 Proportions of information searching prompts across different trust trajectory patterns. (* 𝑝 < 0 . 05 )
Figure 4 Proportions of information searching prompts across different trust trajectory patterns. (* 𝑝 < 0 . 05 )CC-BY · From the PDF
Image 6
CC-BY · From the PDF
Image 7
CC-BY · From the PDF
Figure 6 Actual (a-b) and perceived (c-d) authorship across different self-efficacy trajectory patterns. (* 𝑝 < 0 . 05 , ** 𝑝 < 0 . 01 , *** 𝑝 < 0 . 001 )
Figure 6 Actual (a-b) and perceived (c-d) authorship across different self-efficacy trajectory patterns. (* 𝑝 < 0 . 05 , ** 𝑝 < 0 . 01 , *** 𝑝 < 0 . 001 )CC-BY · From the PDF
Figure 7 Turn-level self-efficacy ratings across prompt categories (* 𝑝 < 0 . 05 )
Figure 7 Turn-level self-efficacy ratings across prompt categories (* 𝑝 < 0 . 05 )CC-BY · From the PDF
Figure 8 Turn-level trust ratings across prompt categories (** 𝑝 < 0 . 01 )
Figure 8 Turn-level trust ratings across prompt categories (** 𝑝 < 0 . 01 )CC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 1 Not valid · 37 Not checked · 117 Own results and interpretations

73 more claims

Create a free account to see all 155 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 42
  • The quote only partly states this claim. 1
  • Probably not a substantial claim. 44

Create a free account to see these claims.

Create a free account to see these assumptions.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentz = -4.9reported p < .001 · recomputed p < .001

    “z = -4.90,p < .001”

  • Consistentz = 2.63reported p < .01 · recomputed p = .009

    “z = 2.63, p < .01”

  • Consistentz = 4.1reported p < .001 · recomputed p < .001

    “z = 4.10, p < .001”

  • Consistentz = 10.13reported p < .001 · recomputed p < .001

    “z = 10.13,p < .001”

  • Consistentz = 3.67reported p < .001 · recomputed p < .001

    “z = 3.67, p < .001”

5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Citations

Cites

Show all 96 references

Research works printed but not linked: 5

  1. Albert Bandura. 1997. Self-efficacy: The exercise of control. Macmillan.
  2. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems , H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 1877–1901.
  3. Cecilia Ka Yuk Chan. 2023. Is AI Changing the Rules of Academic Misconduct? An In-depth Look at Students’ Perceptions of ’AI-giarism’. arxiv:https://arXiv.org/abs/2306.03358 [cs.CY] https://arxiv.org/abs/2306.03358 · arXiv
  4. Sture Holm. 1979. A Simple Sequentially Rejective Multiple Test Procedure. Scandinavian Journal of Statistics 6 2 (1979) 65–70. http://www.jstor.org/stable/4615733
  5. Brian Lubars and Chenhao Tan. 2019. Ask not what AI can do, but what AI should do: Towards a framework of task delegability. In Advances in Neural Information Processing Systems , H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett (Eds.), Vol. 32. Curran Associates, Inc.

Cited by

Related

See all

Metadata

Authors
Park, Y. S., Arvi, N. A. P., Kim, S., & Kim, J.
Affiliations
Korea Advanced Institute of Science and Technology (KR)
ORCID
Venue
Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2026-04-13
Pages
1-18
DOI
10.1145/3772318.3790276
OpenAlex
W7154187432
Cited by
8
References
96
Citations by year
2026: 8
FWCI
29.26
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Research Foundation of Korea (RS-2025-00557726)
  • Institute of Information & Communications Technology Planning & Evaluation of Korea (RS-2024-00443251)
Topics
  • Physical Sciences › Computer Science › Human-Computer Interaction › Innovative Human-Technology Interaction
  • Social Sciences › Human Factors and Ergonomics › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
Keywords
authorship, large language models, human-LLM collaboration, self-efficacy, trust, prompting strategies, text editing

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-14, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-14 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.