PaperlayerView exact revision on arXiv

Source-linked AI summary

The HaLLMark Effect: Supporting Provenance and Transparent Use of Large Language Models in Writing with Interactive Visualization

Md Naimul Hoque, Tasfia Mashiat, Bhavya Ghai, Cecilia Shelton, Fanny Chevalier, Kari Kraus, Niklas Elmqvist

arXiv:2311.13057v4cs.HC

TL;DR

Writers need ways to retain agency and ownership while using LLMs and meeting AI-writing policies. HaLLMark captures and visualizes writer–LLM provenance; creative writers reported control and ownership, though the study focused on creative writers.

  • Problem

    Writers need ways to track AI contributions to protect agency and ownership and to report LLM use under policies.

    Source: §1 · p. 2

  • Method

    HaLLMark is a writing probe that stores and visualizes writer–LLM interactions using text highlighting and a visual timeline.

    Source: §1 · p. 2, §7 · p. 13

  • Results

    Writers valued explicit representations of AI influence, and HaLLMark helped them maintain control and ownership while supporting transparent policy compliance.

    Source: §6 · p. 11

  • Takeaways & Limitations

    Interactive provenance offers writers a way to reflect on AI contributions and communicate AI’s influence to publishers without manually preparing disclosures.

    Source: §1 · p. 2, §6 · p. 11

    The study included only creative writers, whose experiences may differ from those of academic writers or journalists.

    Source: §6.3 · p. 12

Abstract

from arXiv · show

The use of Large Language Models (LLMs) for writing has sparked controversy both among readers and writers. On one hand, writers are concerned that LLMs will deprive them of agency and ownership, and readers are concerned about spending their time on text generated by soulless machines. On the other hand, AI-assistance can improve writing as long as writers can conform to publisher policies, and as long as readers can be assured that a text has been verified by a human. We argue that a system that captures the provenance of interaction with an LLM can help writers retain their agency, conform to policies, and communicate their use of AI to publishers and readers transparently. Thus we propose HaLLMark, a tool for visualizing the writer's interaction with the LLM. We evaluated HaLLMark with 13 creative writers, and found that it helped them retain a sense of control and ownership of the text.

1 INTRODUCTION

Writers face tension between using LLMs and retaining agency, ownership, and accountability under AI-writing policies. HaLLMark captures and visualizes writer–LLM interaction; creative writers reported greater control and ownership.

1 INTRODUCTION

  • Writers worry that attributing AI contributions can threaten agency and ownership, while policy compliance requires tracking their use of LLMs.The introduction notes that writers have few mechanisms for tracking accountability under these policies.Source: §1 · p. 2
  • HaLLMark stores and visualizes writer–LLM interaction over time, highlighting AI contributions to support reflection on AI use.The authors argue that interactive provenance exploration can improve writers’ agency, control, and ownership of the final artifact.Source: §1 · p. 2
  • Creative writers using HaLLMark reported greater control and ownership, and confidence that it could support policy compliance and transparent communication.The evaluation involved writers using the tool to write a short story during remote sessions.Source: §1 · p. 2

2 RELATED WORK

The related work spans writing support, concerns about LLM use, text visualization, and provenance tracking in collaborative work. It situates HaLLMark as addressing the specific challenge of interactive provenance and transparency in AI-assisted writing.

2.1 Writing Support Tools

  • HaLLMark shares generative writing support with tools such as CoAuthor and Wordcraft, but focuses on provenance, policy conformity, ownership, and transparency.Prior writing tools also support organization, revision, and creative tasks.Source: §2.1 · pp. 2–3

2.2 Concerns around LLMs as Writing Assistants

  • Writers’ concerns include agency, ownership, ethics, plagiarism, and stereotyped text, while policies ask them to track and report LLM interactions.The paper identifies uncertainty about how writers can operationalize those policies and report AI use transparently.Source: §2.2 · p. 3

2.3 Visualization for Text and Writing

  • Text visualization supports comprehension of text, but its application in writing tools is limited, and cited LLM visualizations do not target writing provenance or transparency.Examples of writing-oriented visualization include tools for social bias, character traits, and poetic sound.Source: §2.3 · p. 3

2.4 Visualizing and Tracking Contributions in Collaborative Work

  • Prior systems visualize collaborative contributions and human–AI provenance, but AI-assisted writing adds challenges around policy requirements and writers’ interactive needs.Related work includes visualizations for collaborative writing and systems for examining human–AI contributions.Source: §2.4 · pp. 3–4

3 FORMATIVE ANALYSIS OF AI-ASSISTED WRITING POLICIES

A review of publishing policies found shared expectations to disclose AI contributions and keep human responsibility, alongside differences in domain-specific guidance and ambiguity about disclosure. The authors distilled these policies into three information dimensions for AI-assisted writing.

3 FORMATIVE ANALYSIS OF AI-ASSISTED WRITING POLICIES

  • The authors reviewed policies from government, creative, educational, and academic organizations, noting that their list was not exhaustive and policies may evolve.The sources included the U.S. Copyright Office, Authors Guild, MLA, ACL, and ACM.Source: §3 · p. 4
  • The policy analysis used thematic coding by two authors, with a codebook and a second coding round; inter-rater reliability was 0.91 by Jaccard similarity.The coders resolved disagreements together, and the full team participated in regular discussions.Source: §3 · p. 4

3.1 Patterns and Differences

  • Across policy domains, authors are expected to disclose AI contributions and remain responsible for verifying and modifying AI-generated content.The policies also reject AI authorship.Source: §3.1 · p. 4
  • Creative policies emphasize copyright and permission for training on books, academic policies give explicit use and reporting guidance, and educational policies stress plagiarism and integrity.Academic policies also emphasize fact-checking and referencing more than creative policies.Source: §3.1 · p. 4
  • Disclosure guidance is ambiguous: policies vary on what to report and where, from brief notes or AI-provided text to method-section reporting.ACL discourages direct use of AI-generated text, while Nature does not specify what authors should report.Source: §3.1 · p. 4

3.2 Information Typology

  • The typology distinguishes prompts seeking new content from requests to edit existing text; policies generally exempt routine editorial tasks but expect disclosure of generative prompts.ACM asks authors to provide a list of generative prompts used for new content.Source: §3.2 · p. 4
  • Policies ubiquitously call for disclosing AI-generated text, and some also require acknowledging AI inspiration or ideas even when generated text is not used directly.Some publishers may reject work that is mostly AI-generated.Source: §3.2 · p. 4
  • Some publishers set AI-use thresholds, such as the Authors Guild’s 5% limit, but do not specify how to measure or report that percentage.The threshold is uncommon in the reviewed policies.Source: §3.2 · p. 4

4 THE HALLMARK SYSTEM

HaLLMark is a technology probe combining a text editor, an LLM, and an interaction history, designed around policy-informed provenance and writer judgment. Its interface links prompts and text through cards, visual summaries, a timeline, and manual annotations.

4 THE HALLMARK SYSTEM

  • HaLLMark combines a text editor, an LLM, and an interaction history to explore how provenance tracking can support writers and inform future systems.The probe was informed by a formative analysis of AI-writing policies and feedback from two domain experts.Source: §4 · pp. 4–5

4.1 Design Rationale

  • The design aims to capture prompt category, use of AI responses, and summary statistics, then connect interaction history to the artifact and let writers edit provenance.Because AI influence can be implicit, the system leaves some provenance judgments to writers rather than enforcing policy compliance.Source: §4.1 · p. 5

4.2 Visual Interface

  • The interface divides the screen among a text editor, GPT-4 interaction, and provenance visualization, with controls to hide modules and redistribute space.Source: §4.2 · p. 5

4.2.1 Prompting LLMs.

  • Writers can prompt GPT-4 with free-form or suggested tasks and optionally select text from the editor as prompt context.The selected text is automatically added to the context box.Source: §4.2.1 · p. 5, Figure 2 · p. 6

4.2.2 Prompt Card.

  • Each prompt and response appears in a card categorized as generating new content or editing existing text; the soft-prompt classifier achieved 96% accuracy on 150 labeled prompts.Cards expose prompt context and provide copy and redo controls.Source: §4.2.2 · pp. 5–6, Figure 3 · p. 6

4.2.3 Visualizing AI vs. human provenance.

  • The provenance view summarizes prompt counts and AI contribution, while a timeline places prompts alongside writing activity and links them to text.Copied AI text is marked AI-written; writers can manually label text AI-influenced, a category the system does not automatically detect.Source: §4.2.3 · pp. 6–7, Figure 4 · p. 7

4.2.4 Linking Visualization and Artifact.

  • Writers can select text to label or remove AI-written or AI-influenced annotations, link it to a prompt, and inspect that prompt from the annotated text.These actions support manual annotation when the system cannot automatically annotate text.Source: §4.2.4 · p. 7, Figure 5 · p. 7

5 EVALUATION

The evaluation used HaLLMark as a probe with 13 creative writers to investigate how provenance externalization relates to LLM interaction, ownership concerns, AI-use communication, and policy conformity. The authors chose creative writing as a case study, so findings remain specific to that use case.

5 EVALUATION

  • The study used 13 creative writers to examine provenance externalization in AI-assisted co-writing through four questions about interaction, ownership, communication of LLM use, and policy conformity.Source: §5 · pp. 7–8

5.1 Different Forms of Writing

  • Although HaLLMark may apply to other writing types, the evaluation results remain specific to the creative-writing use case.Source: §5.1 · p. 8
  • Creative writing was selected because writers face agency and ownership concerns, while hallucination is less problematic there than in nonfiction and academic writing.The authors cite prior reports of creative writers’ ownership and agency concerns when using LLMs.Source: §5.1 · p. 8

5.2 Study Conditions

The study compared a text-only baseline with HaLLMark in a counterbalanced, repeated-measures design. Participants were published creative writers with varied experience and prior exposure to LLMs.

5.2 Study Conditions

  • Participants completed a counterbalanced, repeated-measures comparison of a ChatGPT-like editor with prompt history and HaLLMark with its interactive provenance support.The baseline was chosen to represent current LLM writing interfaces and isolate the impact of provenance visualization.Source: §5.2 · p. 8

5.3 Participants

  • The 13 participants were recruited through university writing and English or literature departments, and all had published work in their portfolios.They ranged from 19 to 56 years old and had 5 to 26 years of creative-writing experience.Source: §5.3 · p. 8
  • All participants had prior LLM experience or awareness, but none described LLMs as critical to their writing; five used ChatGPT infrequently for editorial tasks.Two had explored narrative angles, and two had participated in the 2023 WGA/SAG-AFTRA strike.Source: §5.3 · p. 8

5.4 Tasks

Writers wrote short stories while using GPT-4 for a fixed 20-minute period, allowing the study to examine interaction with two conditions despite the lack of objective creative-writing goals.

5.4 Tasks

  • Writers used the study interfaces to write short stories and could prompt GPT-4 during a fixed 20-minute session.The task was chosen because creative writing is difficult to structure around objective goals.Source: §5.4 · p. 8

5.5 Measures

The study combined semi-structured interviews with 7-point ratings to capture writers’ experiences, including abstract concepts that are difficult to operationalize quantitatively.

5.5 Measures

  • Semi-structured interviews asked how each condition affected interaction with LLMs, agency, control, ownership, communication, and transparency.The authors chose qualitative methods because the tasks lacked objective goals and relevant concepts are difficult to measure quantitatively.Source: §5.5 · p. 8
  • Three 7-point ratings assessed comfort publishing under one’s name and perceived help with transparency and guideline conformity.The ratings covered ownership, communication and transparency, and conformity to the given AI-assisted writing guideline.Source: §5.5 · p. 8
  • Six 7-point agreement items assessed AI helpfulness, ease, experimentation, iteration, pride in the output, and perceived story uniqueness.The items were used to capture usability of LLM support.Source: §5.5 · pp. 8–9

5.6 Procedure

Participants reviewed an AI-writing policy, prepared story ideas, and completed two timed writing sessions using the study’s two interfaces, followed by surveys and an interview.

5.6 Procedure

  • Before the sessions, participants reviewed the U.S. Copyright Office policy and considered two story plots and settings without beginning to write.Each interface was introduced with a brief demo, and participants explored its features using a training story.Source: §5.6 · p. 9
  • Participants wrote a short story for 20 minutes in each condition and did not need to finish the story.The session was explicitly framed as a timed experience.Source: §5.6 · p. 9
  • After each session participants completed a survey; after using the other interface for the second story, they completed another survey and a semi-structured interview.The interview covered experience with both conditions and use of the LLM in the writing process.Source: §5.6 · p. 9

5.7 Analysis Plan

The authors analyzed interview transcripts thematically and used bootstrap confidence intervals and standardized effect sizes for quantitative measures and subjective ratings.

5.7 Analysis Plan

  • Two authors independently open-coded anonymized interviews, refined themes through discussion with the research team, and reported initial Jaccard agreement of 0.86.The coders met regularly to discuss and refine codes and themes.Source: §5.7 · p. 9
  • Quantitative measures and subjective ratings used 95% confidence intervals from 1,000 non-parametric bootstrap iterations and reported Cohen’s d.The authors chose estimation methods instead of traditional null-hypothesis-based statistical testing.Source: §5.7 · p. 9

5.8 Results

HaLLMark changed how writers interacted with AI and supported their reported control, ownership, transparency, and policy conformity. Participants also found it usable, though they suggested improvements and some raised concerns about sharing its information.

5.8.1 RQ1: Interaction with LLM.

  • 3.48% of final-story text was AI-written with HaLLMark, versus 13.66% with the baseline; writers also labeled 2.99% as AI-influenced.The baseline had no option to label AI-influenced text.Source: §5.8.1 · p. 9, Figure 6 · p. 10
  • Writers used similar numbers of prompts across conditions, while interviews indicated HaLLMark encouraged awareness and active evaluation of AI assistance from the start.Participants described monitoring AI contributions and modifying machine-generated text.Source: §5.8.1 · p. 9

5.8.2 RQ2: Agency and Ownership.

  • HaLLMark improved writers’ reported control and ownership; ownership ratings averaged 4.92 with HaLLMark and 3.39 with the baseline.The reported standardized effect size was 0.46 (Cohen’s d).Source: §5.8.2 · pp. 9–10

5.8.3 RQ3: Communication and Transparency.

  • Participants generally preferred HaLLMark for communicating AI contributions, though two had reservations; transparency ratings averaged 6.31 versus 3.14 for the baseline.One participant worried readers might criticize transparent AI use, while another preferred not to share summary statistics.Source: §5.8.3 · p. 10

5.8.4 RQ4: Conformity to AI-assisted Writing.

  • All participants preferred HaLLMark for evaluating policy conformity; ratings averaged 6.00 with HaLLMark and 2.83 with the baseline.The reported standardized effect size was 2.10 (Cohen’s d).Source: §5.8.4 · p. 10

5.8.5 Subjective Perception of LLM Support.

  • Participants’ subjective ratings of LLM-support usability dimensions showed no significant differences between conditions.The ratings covered six dimensions.Source: §5.8.5 · p. 10, Figure 8 · p. 11

5.8.6 Cognitive Load and Usability.

  • Participants found HaLLMark easy to use without reporting excessive cognitive load, but noted that always-on highlighting could distract during long-term use.Participants also requested options including toggling highlighting, adding notes, and controlling generation randomness.Source: §5.8.6 · p. 11

6 DISCUSSION, LIMITATIONS, AND FUTURE WORK

HaLLMark’s provenance display helped writers track AI contributions, maintain control and ownership, and communicate use of AI. The discussion considers writers’ preferences, practical uses, study boundaries, and ethical concerns around transparency.

6 DISCUSSION, LIMITATIONS, AND FUTURE WORK

  • Tracking AI contributions helped writers maintain a comfortable level of assistance, feel control over the process, and improve ownership of the final artifact.Participants also said HaLLMark could support transparency and policy conformity without manually preparing disclosures.Source: §6 · p. 11

6.1 Normalize Transparency and Accountability in AI-assisted Writing

  • Some writers feared public disclosure could invite criticism and diminish the perceived human contribution, despite generally welcoming provenance sharing.The authors argue for transparency and accountability alongside more charitable responses from readers, publishers, and other writers.Source: §6.1 · p. 11

6.2.1 Writers Want to Use LLMs for Content Generation, not Editing.

  • Writers mostly wanted GPT-4 to generate content or ideas, not whole stories or editorial changes, but the 20-minute sessions may not represent long-form use.The authors suggest the contrast with a survey reporting that 60% wanted editorial use may reflect tangible tool experience versus preconceived perceptions.Source: §6.2.1 · pp. 11–12

6.2.2 When to Write, When to Prompt, and When to Verify?

  • Writers’ preferred timing for writing, prompting, and provenance tracking varied, motivating tools that let them switch freely and a planned longitudinal study.The study observed participants turning the three modules on and off in different combinations.Source: §6.2.2 · p. 12

6.2.3 Adopting HaLLMark in Writing Tools.

  • HaLLMark is currently standalone, but the authors describe adapting it as a plugin or desktop app and sharing structured provenance across tools.They propose JSON storage and copy-and-paste actions as one route to platform-independent provenance.Source: §6.2.3 · p. 12

6.2.4 HaLLMark as a Reading Tool.

  • Writers could share an interactive document, static report, or structured provenance data so publishers, reviewers, or readers can inspect AI contributions.Publishers can decide how provenance is shared with a wider audience.Source: §6.2.4 · p. 12

6.3 Limitations

  • As a technology probe, HaLLMark is not production-ready, and the study covers creative writers rather than the broader population of writers.Long timelines may become difficult to interpret; academic and journalistic settings require further study because they rely on verifiable facts and evidence.Source: §6.3 · p. 12

6.4 Ethical Concerns of LLM based Co-Writing

  • The authors frame AI as a tool under human verification, while noting risks of publisher surveillance and that writers chiefly sought to improve their craft, not produce copy faster.They also support making tools freely available while prohibiting use to cross the WGA and SAG-AFTRA picket line.Source: §6.4 · p. 13

7 CONCLUSION

The conclusion presents HaLLMark as a technology probe for prompting and recording human–AI writing provenance. Writers valued visible AI influence and reported a smoother workflow than with the default ChatGPT interface, while the authors identify several open research needs.

7 CONCLUSION

  • HaLLMark combines LLM prompting with stored provenance, visualizing prompts and AI influence through text highlighting and a timeline.The technology probe was designed based on a review of generative AI guidelines from professional and research organizations.Source: §7 · p. 13
  • Writers valued explicit representations of AI influence, and HaLLMark’s prompting interface yielded a smoother, more integrated workflow than the default ChatGPT interface.The evaluation involved writers creating a short story or non-fiction article.Source: §7 · p. 13
  • Human–AI co-writing remains a nascent, controversial area, with future work needed on prompt scaffolding, provenance tracking, and non-repudiation of textual outputs.The authors describe HaLLMark as neither a final nor optimal approach to transparency or prompting.Source: §7 · p. 13

Related papers

Loading 2311.13057v4…