Source-linked AI summary
Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas
Advait Bhat, Marianne Aubin Le Quéré, Mor Naaman, Maurice Jakesch
TL;DR
Research shows AI assistance can affect written views and later opinions, but lacks a process-level account of how suggestions reshape co-writing. The study combines qualitative interviews using a custom replay tool with quantitative analysis validating reactive writing, including suggestion evaluation and AI-generated ideas permeating text. Reactive writing shifts activity toward evaluating and elaborating on AI suggestions, which can introduce ideas and framings into final text while writers feel in control.
Problem
Research shows AI assistance can affect written views and later opinions, but lacks a process-level account of how suggestions reshape co-writing.
Method
The study combines qualitative interviews using a custom replay tool with quantitative analysis validating reactive writing, including suggestion evaluation and AI-generated ideas permeating text.
Results
Reactive writing shifts activity toward evaluating and elaborating on AI suggestions, which can introduce ideas and framings into final text while writers feel in control.
Takeaways & Limitations
AI writing assistants can function as subtle agenda-setting technologies, especially when they provide opinionated suggestions that writers incorporate into their text.
Takeaways & Limitations
The experimental setting may amplify opinionated-suggestion effects and has limited ecological validity, while the qualitative sample lacks professional and cultural diversity.
Abstract
from arXiv · showhide
Emerging experimental evidence shows that writing with AI assistance can change both the views people express in writing and the opinions they hold afterwards. Yet, we lack substantive understanding of procedural and behavioral changes in co-writing with AI that underlie the observed opinion-shaping power of AI writing tools. We conducted a mixed-methods study, combining retrospective interviews with 19 participants about their AI co-writing experience with a quantitative analysis tracing engagement with ideas and opinions in 1{,}291 AI co-writing sessions. Our analysis shows that engaging with the AI's suggestions -- reading them and deciding whether to accept them -- becomes a central activity in the writing process, taking away from more traditional processes of ideation and language generation. As writers often do not complete their own ideation before engaging with suggestions, the suggested ideas and opinions seeded directions that writers then elaborated on. At the same time, writers did not notice the AI's influence and felt in full control of their writing, as they -- in principle -- could always edit the final text. We term this shift \textit{Reactive Writing}: an evaluation-first, suggestion-led writing practice that departs substantially from conventional composing in the presence of AI assistance and is highly vulnerable to AI-induced biases and opinion shifts.
1 Introduction
AI writing assistants may shape both what people write and what they believe, but the processes behind this influence remain poorly understood. This study combines interviews and interaction-log analysis to examine how inline opinionated suggestions reshape writing and ideas.
- The study asks how inline opinionated AI assistance reshapes ideation and composition, and how co-writing affects the ideas writers ultimately produce.
- The authors use 19 retrospective interviews and quantitative analysis of 1,291 writing sessions to trace writers’ experiences and how suggestions propagate into final texts.
- Inline AI suggestions can redirect writers from internally generating ideas toward evaluating and elaborating on AI-presented content while writers retain a strong sense of control.
- The study frames this process as reactive writing, an evaluation-first, suggestion-led mode that may make AI-assisted writing vulnerable to biases and opinion shifts.
- Exposure to AI suggestions on specific topics strongly predicts those topics’ presence in final essays even after controlling for directly accepted text.
2 Related Work
Prior research presents AI writing assistants as tools for productivity, ideation, and revision, while also documenting cognitive, linguistic, affective, and persuasive trade-offs. The paper addresses the gap between these process-level and outcome-level accounts.
- AI writing systems have evolved from efficiency-oriented predictors into active partners supporting idea generation, storytelling, revision, and creative writing.
- Writers engage with suggestions in varied ways, from avoidance to chaining, and may use assistance for creative ideation or for rote, mindless tasks.
- Research finds that AI assistance can improve productivity or quality while reducing satisfaction, encouraging predictable language, and affecting sentiment or self-description.
- Algorithmic writing suggestions can inherit training-data biases and shape writing content and direction through the act of presenting particular framings or arguments.
- Existing work has largely studied cognitive effects and persuasive effects separately rather than how they interact during writing itself.
3 Methods
The study combines a controlled inline-AI writing platform, retrospective interviews, and large-scale interaction-log analysis. A two-stage topic pipeline tracks how suggested ideas enter participants’ writing.
- The study uses a mixed-methods design combining qualitative accounts of lived experience with quantitative analysis of text-output data.
- 3.1 Experimental Platform: Participants wrote opinion statements about social media using an inline GPT-3-based assistant that generated opinionated suggestions after natural pauses and logged detailed interactions.
- The qualitative component used 19 US- and UK-based Prolific participants, immediate 30–45 minute replay-supported retrospective interviews, and thematic analysis of their protocols.
- The quantitative dataset began with 1,506 sessions and, after removing malformed, incomplete, and outlier data, retained 1,291 participants across control, pro-social-media, and anti-social-media conditions.
- 3.3 Quantitative Topic Analysis: Topics serve as proxies for ideas: the pipeline discovers candidate topics, clusters them into final topics, then assigns them to sub-sentence segments of writing bursts.
4 Qualitative Results: The Reactive Writing Model
Interviews identify reactive writing as a three-stage process: AI suggestions capture attention, agreement governs inclusion, and writers personalize accepted content. This process redirects ideation while preserving writers’ perceived agency.
- The reactive writing model comprises attention capture, agreement-governed inclusion, and post-hoc personalization.
- Attention capture: Inline suggestions interrupted participants’ emerging ideation, redirecting attention from recalling experiences and forming arguments toward reading and evaluating suggested sentences.
- Attention capture: Evaluating pre-constructed suggestions required less effort than independently recalling experiences and crafting arguments, making suggestion engagement comparatively easy.
- Agreement-governed inclusion: Writers commonly accepted suggestions when they agreed with or did not contradict the content, while interface defaults made acceptance the path of least resistance.
- Agreement-governed inclusion: Polished, fluent prose encouraged acceptance and could make AI suggestions appear authoritative, even when the wording did not match writers’ precise voices or thoughts.
- Post-hoc personalization: Writers rejected suggestions that failed to reflect their voice, knowledge, or personal experience, and nevertheless felt in control because they could accept, reject, or revise them.
4.3 Post-hoc personalization
Post-hoc personalization describes how writers adapt accepted AI suggestions to their own voice, experiences, and perspectives, while AI ideas can continue shaping writing even when rejected.
- 4.3 Post-hoc personalization: Accepted AI suggestions were adapted to writers’ own voice, stance, experiences, and desired phrasing.Edits included shortening, rephrasing, and making generic claims more personal or specific.
- 4.3 Post-hoc personalization: Rejected suggestions still influenced writing by prompting related ideas, counterarguments, or alternative positions.Participants sometimes wrote against suggestions they disagreed with, showing that influence did not require acceptance.
- 4.3 Post-hoc personalization: Writers often preserved an AI suggestion’s underlying framing even after stylistic or qualifying edits.Some edits softened a definitive stance, but remained within the suggestion’s initial conceptual framing.
- 4.3 Post-hoc personalization: AI suggestions redirected thought toward topics participants had not previously considered, functioning as cues for subsequent ideation.Participants described their minds going in the suggestion’s direction and credited suggestions with kickstarting thoughts about misinformation or hate speech.
- 4.3 Post-hoc personalization: Reactive writing shifts early composition from personal ideation toward evaluating external suggestions and then elaborating them through individual voice and experience.The process comprises attention capture, agreement-governed inclusion, and post-hoc personalization, repositioning writers as evaluators and extenders.
5 Exploratory Quantitative Results
Across 1,291 co-writing sessions, AI suggestions captured writing time and attention, while suggested topics substantially shaped both final essays and participants’ self-written text.
- 5.1 Attention capture: AI-assisted participants accepted about 31% of essay text but realized only 7.5% time savings relative to controls.They spent 250 or 248 seconds versus 269 seconds without AI, while evaluating displayed suggestions consumed substantial time.
- 5.2 Agreement-governed inclusion: Suggested topics substantially biased final essays toward positive or critical framings, relative to the control group.Positive assistance emphasized benefits such as global connectivity, whereas critical assistance emphasized addiction, bullying, and loneliness.
- 5.2 Agreement-governed inclusion: Topic frequency in AI suggestions positively predicted topic frequency in final text, with R2 = 0.85 and β = 0.64.The model was statistically significant: F(1, 38) = 212.06, p < .001, adj. R2 = 0.84; 95% CI [0.55, 0.72].
- 5.3 Post-hoc personalization: Each additional topic suggestion predicted 4.43 words in the final essay and 1.95 additional writer-produced words.Both associations were statistically significant, including the self-written-text model with 95% CI [1.79, 2.11].
- 5.3 Post-hoc personalization: Seeing at least one suggestion increased the odds that a topic appeared in essays by 3.97-fold and in participants’ own writing by 1.65-fold.Both odds ratios were statistically significant at p < .001.
- 5.3 Post-hoc personalization: AI-originated ideas influenced participants’ self-written text, consistent with writers elaborating on suggestions as ideation cues.The quantitative pattern aligns with qualitative evidence that AI suggestions interfered with ideation and shaped subsequent writing.
6 Discussion
The discussion frames reactive writing as a suggestion-led process in which AI captures attention, lowers acceptance thresholds, and embeds its framings into writers’ personalized prose. These mechanisms may explain how opinionated assistants shape topics and opinions, although the strongest implications are limited to the study’s deliberately strong setting.
- 6 Discussion: Reactive writing shifts writers from internally generating ideas toward evaluating and elaborating on AI-presented content.Inline suggestions interrupt nascent ideation, while writers retain a strong sense of agency and control.
- 6 Discussion: Default effects, automation bias, fluency, and satisficing can make accepting roughly matching suggestions easier than generating alternatives.Participants often accepted suggestions even when they did not fully match their voice or thought process.
- 6 Discussion: Post-hoc personalization preserves AI-introduced framings while embedding them in writers’ own tone, experiences, and reasoning.This redistribution of cognitive labor leaves writers evaluating and editing while ideas are scaffolded by the model.
- 6.1 Implications for Persuasion: Reactive writing creates conditions for biased suggestions to shape expressed and actual opinions through attention capture, lowered acceptance thresholds, and personalization.The discussion connects this sequence to self-persuasion, where people can be influenced by arguments they elaborate themselves while feeling agency.
- 6.1 Implications for Persuasion: AI suggestions can act like agenda-setting technologies by making particular framings salient and directing which topics writers discuss.The discussion relates this mechanism to agenda-setting theories in mass communication.
- 6.3 Limitations: The strongest persuasive implications are limited to opinionated assistants, while attention capture and satisficing may extend to more neutral tools.The study used stance-laden suggestions and a deliberately strong reactive-writing configuration, which may amplify effects in real-world systems.
7 Generative AI Disclosure
The authors disclose using generative AI tools for literature discovery, manuscript refinement, streamlining, and formatting, while stating that all manuscript text was written and finalized by the authors.
- 7 Generative AI Disclosure: ChatGPT supported literature discovery, while Claude and ChatGPT refined manuscript content and handled formatting tasks.The authors state that all manuscript text was written and finalized by them.
Pre-interview
The pre-interview procedure began with study enrollment, followed by essay writing and an invitation to join a Zoom interview.
- Pre-interview: Participants joined the study, wrote an essay, and were then invited to join a Zoom interview.
Interview
The interview used cued retrospective protocols to reconstruct participants’ thoughts and decisions during AI-assisted writing. Replay-based questioning examined initial plans, reactions to suggestions, acceptance decisions, and subsequent writing.
- Interview: Cued retrospective protocols replayed writing sessions and repeatedly elicited what participants were thinking and doing at specific moments.The procedure emphasized recalling actions and thoughts without justifying decisions.
- Interview: At the session start, interviewers asked about participants’ concrete writing plans and how initial suggestions compared with those plans.These questions targeted the relationship between pre-existing ideas and AI suggestions.
- Interview: When suggestions appeared, interviewers asked about participants’ ongoing thoughts, intended text, and how much of the suggestion they read.The questions focused on the writer’s process immediately before and during suggestion exposure.
- Interview: Interview materials included topic-labeling references and consent procedures requiring anonymization, study-use restrictions, and deletion of recorded data after one year.The interview protocol listed addiction and distraction among the topic labels and documented participant withdrawal rights.
- Interview: After suggestion acceptance, interviewers asked why it was chosen, what participants had planned instead, and why other suggestions were rejected.The protocol compared selected suggestions with participants’ original plans and alternatives.
- Interview: The interview also examined ideas triggered by suggestions, similarities between suggestions and later sentences, perceived system functioning, and themes such as inspiration and selection.Participants were questioned both about reactions to suggestion content and about possible influence on later writing.
A.2 Appendix B: Dendrogram and 20 Final Topics
Figure 8 presents a dendrogram for the process of merging and splitting topics. It documents the topic-organization structure used in Appendix B.
- A.2 Appendix B: Dendrogram and 20 Final Topics: Figure 8 uses a dendrogram to represent topic relationships.The visual is specifically described as a dendrogram for topic organization.
- A.2 Appendix B: Dendrogram and 20 Final Topics: The dendrogram depicts where topics are merged.Merging is one of the two operations identified in the figure description.
- A.2 Appendix B: Dendrogram and 20 Final Topics: The dendrogram also depicts where topics are split.Splitting is presented alongside merging as part of the topic-organization process.