cs.HCMar 11, 2026

Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas

Authors: Advait Bhat, Marianne Aubin Le Quéré, Mor Naaman, Maurice Jakesch

Organizations: Paul G. Allen School of Computer Science, University of Washington · Center for Information Technology Policy, Princeton University · Cornell Tech · Bauhaus University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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 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.

Figures & tables

Figure 1. Screenshot of the experimental writing platform showing AI suggestions. The image is a screenshot of a Reddit-style discussion platform. At the top, a post asks, "Is social media good for society?" The user is composing a response in a text box, starting with "In my view" and the inline AI suggestions appear in gray saying "social media is a powerful tool that can benefit society in multiple ways. For instance, it enables instantaneous…". The interface includes buttons for saving and sharing and instructions on how to complete the task along with how to accept the AI suggestions.
Figure 2. Simplified overview of the topic classification pipeline. Step 1 ( topic discovery ) shows how initial topics were identified from user-written and accepted text using gpt-3.5-turbo , with each topic receiving a description to enable clustering. Step 2 ( topic assignment ) shows how text was split into sub-sentences based on topic separation, with each sub-sentence assigned by gpt-3.5-turbo to one of the final topics derived from hierarchical clustering. Two-step pipeline diagram using gpt-3.5-turbo. Step 1 (Topic Discovery): user-written text in a speech bubble leads to two boxes showing discovered topics with descriptions (Social connection, Cultural exposure). Step 2 (Topic Assignment): the same text split into sub-sentences, each assigned to final clustered topics (Global connectivity and networking, Information news and learning).
Figure 3. A model of the Reactive Writing process. In this model, a new AI suggestion triggers Step 1: attention capture , where a writer’s attention is redirected from ideation to a new suggestion. In Step 2: agreement-governed inclusion , the writer immediately begins to evaluate whether they wish to accept or reject the text in their writing, with a bias towards accepting suggestions. The diagram lists the main reasons to include or reject the text. When text has been accepted, writers turn to Step 3: post-hoc personalization , where they incorporate and adapt the suggestions to their own writing. A flowchart showing the Reactive Writer model with three numbered steps: (1) Attention Capture, (2) Agreement-Governed Inclusion, and (3) Post-hoc Personalization. A robot icon represents the AI generating new suggestions. Two boxes on the right show reasons for accepting suggestions (agrees with content, admires prose, finds it convenient, believes they can build upon it) and rejecting suggestions (disagrees with content, doesn't match voice, wants to stay original, resisting intrusion).
Figure 4. Writing with AI did not substantially reduce the time participants spent on the essay. Mean task time with 95% confidence intervals in N = 1,291 writing sessions. When writing without AI assistance, participants, on average, wrote for 269 seconds. Participants writing with AI assistance spent about 248-250 seconds on their essay. Bar chart comparing mean task time across three conditions: Control (no AI) at approximately 269 seconds, Positive AI at approximately 250 seconds, and Critical AI at approximately 248 seconds. Error bars show 95% confidence intervals. The differences between AI conditions and control are small.
Figure 5. Frequency of topics in the AI suggestions (left) and final text (right) when writing with a positive or critical AI assistant. Grey bars indicate control group frequencies for reference. N=1,291, 95% confidence intervals shown in black. The AI assistant’s suggestions (left panel) showed a clear tendency towards certain topics, with many suggestions related to global connectivity in the positive AI treatment group, and many suggestions related to addiction, bullying, and loneliness in the critical AI treatment group. Participants’ written text (right panel) inherited the topical biases of the suggestions to a substantial extent. Bar charts comparing topic frequencies in AI suggestions (left panel) and in final essays (right panel) for positive (blue) and critical (red) assistants, with grey bars marking control group frequencies. The left panel shows assistants emphasize stance-congruent topics; the right panel shows these biases partially carry over into participants' writing.
Figure 6. Embeddings of participants’ sentences when writing with a critical (red) and a positive AI assistant (blue) compared to those writing without AI assistance (grey) . t - SNE embeddings of all sentences in N=1,290 writing sessions. Participants writing with AI assistance wrote about a narrower and systematically different set of topics, while those writing without AI assistance covered a broader area of topics. The marginal density distribution (bottom) shows the horizontal separation of groups along the primary dimension, with critical AI (red) concentrated on the left, positive AI (blue) on the right, and control (grey) distributed across the full range. t-SNE scatter plot of sentence-level embeddings across conditions: red points (critical AI) and blue points (positive AI) cluster into distinct regions on opposite sides of the plot and cover a narrower area than grey points (no-AI control), which are distributed across the full range, indicating more constrained and systematically different topic coverage when writing with AI.
Figure 7. Correlation between how often the AI suggested a topic (x-axis) and how much participants wrote about it (y-axis) Topic correlations with a fitted linear regression line and 95% confidence bands. N=1,291. The left panel shows topic frequencies in the overall text, while the right panel shows topic frequencies in self-written text only. Topics that were suggested more often by the AI were both more likely to appear in participants’ final text and their own writing. Topic correlations with a fitted linear regression line and 95% confidence bands. N=1,291. The left panel shows topic frequencies in the overall text while the right panel shows topic frequencies in self-written text only. Topics that were suggested more often by the AI were both more likely to appear in participants' final text and their own writing.
Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8. Dendrogram for merging and splitting topics. Hierarchical clustering dendrogram showing the agglomerative merging of initial LLM-generated topics into the final 20 topics. The tree structure displays topic labels at the leaves with horizontal lines indicating merge points at different distance thresholds.
TopicDescription
addiction and distractionContent about excessive use, need for moderation, inability to disconnect, or diverting attention from important activities
information sharing, news and learningContent about acquiring knowledge, staying informed, or educational aspects
cyberbullying, arguments, and negativityContent about online harassment, conflicts, or negative interactions
global connectivity, communication and networking (broader)Content about connecting with new people, professional networking, or communication at a societal/global scale
interpersonal connections with friends and family (personal)Content about MAINTAINING or STRENGTHENING existing personal relationships
mental healthContent about psychological well-being, emotional states, or mental conditions
Appendix
Table 1. Topics and brief descriptions used for labeling (no description for uncategorized).
More in this topic

Human-AI Co-Creation

Human-AI co-creation studies how people and AI jointly develop ideas, artifacts, and decisions while preserving meaningful human agency.

See all 148 papers

Explore similar work

CardsList
Sep 1, 2026

Designing Proactive Thought Partners for Writing

Writing involves diverse cognitive activities, from ideation to revision, and writers' needs vary across individuals and moments. Proactive AI promises to provide the right support at the right time, yet existing proactive tools largely focus on generic textual assistance, such as autocomplete. This paper studies the design space of proactive thought partners: AI agents that proactively offer customizable, higher-level cognitive support during writing. We instantiated this concept in a technology probe and deployed it with 16 participants for one week. The probe allows users to create partners by configuring their roles and proactivity. As users write, relevant partners take the initiative at appropriate moments to offer suggestions. Our findings show that participants configured proactive support through prospective planning, used suggestions for both idea generation and self-monitoring, and valued lightweight visual representations alongside non-directive rhetorical framing for non-intrusive interventions. We derive implications for designing proactive writing assistants around customization, timing, engagement, and representation.

Chao Zhang, Abe Davis, Chih-Wei Chen +1

*Work done during an internship at Google DeepMind. · 1Cornell University · 2Google DeepMind

Proactive AssistanceHuman-AI Co-Creation
AskPDF
Apr 28, 2026

Co-Writing with AI: An Empirical Study of Diverse Academic Writing Workflows

Despite AI tools becoming increasingly embedded in academic practice, little is known about how university students integrate them into their writing processes. We examine how students engage with AI across different writing tasks, and how this engagement is shaped by individual factors including AI literacy, writing confidence, trust, authorship concerns, and motivation. Study1 surveys 107 UK university students to map task-specific and co-occurring patterns of AI use across five writing stages (ideation, sourcing, planning, drafting, and reviewing) and their associations with individual factors. Study2 complements this by exploring how these patterns can be assembled in practice, through interviews with 12 postgraduates reflecting on their established use of AI in assessed writing. Together, the studies suggest that AI integration is selective and heterogeneous, forming three recurring and value-oriented configurations: (1) early-stage (learning-oriented), where tools support exploration and understanding; (2) late-stage (quality-oriented), where tools support drafting and refinement; and (3) peripheral (productivity-oriented), where tools are used to reduce friction and sustain momentum across the process. We offer a workflow-level account of AI-supported academic writing, showing how students navigate competing priorities of learning, quality, productivity, and authorship, and how they evaluate and take responsibility for AI-generated outputs.

Silvia Bodei, Duncan P. Brumby, Katie Fisher +1

University College London, London, United Kingdom

AI in EducationAI LiteracyGenerative AI in EducationHuman-AI Co-Creation
AskPDF
Apr 24, 2026

How AI Writing Assistance Distorts Perceptions of AI Users and Their Opinions

Hundreds of millions of people use AI for writing assistance. We asked how AI distorts perceptions of AI users and their opinions. Writers (N=2,939) wrote political opinion paragraphs with and without AI assistance. Separate readers (N=11,091) blindly evaluated these paragraphs and made inferences about the writers. With AI, writers seemed more opinionated, competent, positive, and demographically privileged. Writers objected to many of these distortions yet preferred AI-assisted text even when made aware of them. Reward models trained on our data mitigated objectionable distortions but reduced user acceptance. Additional readers (N=8,798) trusted AI-assisted writers more and were more persuaded by AI writing when AI was more distortive. Overall, distortions from AI writing assistance are pervasive, persistent under human oversight, and likely to have consequential effects on human behaviours and attitudes.

Paul Röttger, Kobi Hackenburg, Hannah Rose Kirk +1

University of Oxford · UK AI Security Institute

Trust in AIDemographic Bias in Language ModelsHuman-AI Co-Creation
AskPDF