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

Paper Title

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

Publication Info

• Topic area: The impact of AI writing assistants on human ideation and composition processes.
• Keywords: AI writing assistants, reactive writing, human-AI collaboration, opinion shaping, ideation, cognitive processes, persuasion, topic modeling, algorithmic influence, inline suggestions.

Background and Problem

• Problem / challenge: Existing research shows that AI writing assistants can influence the content and opinions expressed by users, but there is limited understanding of how these tools reshape the behavioral and cognitive processes of writing.
• Significance: Understanding these changes is critical for designing AI systems that preserve writer agency and mitigate biases, as well as for addressing broader concerns about public discourse and democratic expression.
• Motivation and related work: Prior studies have demonstrated that AI suggestions can shift opinions and writing content, often without users’ awareness. However, these studies primarily focus on outcomes rather than the underlying processes. This paper seeks to fill this gap by examining how AI suggestions influence ideation, memory retrieval, and expression during writing.

Solution

• Proposed approach: The authors introduce the concept of Reactive Writing, a model describing how AI suggestions reorient the writing process from ideation to evaluation and elaboration on AI-generated content.
• Novelty:
  1. Identification of a three-stage reactive writing process: attention capture, agreement-governed inclusion, and post-hoc personalization.
  2. Mixed-methods study combining qualitative interviews and quantitative analysis of 1,291 AI-assisted writing sessions.
  3. Empirical evidence of how AI suggestions influence both the process and content of writing, including indirect effects on self-written text.
  4. Insights into the persuasive potential of AI writing assistants and their role as algorithmic agenda-setters.
• Procedure and key techniques:
  • Conducted 19 retrospective interviews with participants using an opinionated AI writing assistant.
  • Analyzed interaction logs from 1,291 writing sessions to study topic propagation from AI suggestions to final texts.
  • Developed a topic classification pipeline using GPT-3.5-turbo and hierarchical clustering to analyze thematic shifts in writing.

Results

• Concrete findings:
  • Participants writing with AI assistance spent only 7.5–10% less time on their essays compared to those without AI, despite accepting 25–31% of their text from the AI.
  • Topics suggested by the AI strongly influenced the final text, with a statistically significant correlation between the frequency of suggested topics and their presence in participants’ essays (R² = 0.85 for overall text, R² = 0.51 for self-written text).
  • Seeing at least one suggestion on a topic increased the odds of writing about that topic by a factor of 3.97.
• Advantage over baselines:
  • AI-assisted writing sessions showed narrower and more systematically biased topic coverage compared to the broader range of topics in the control group.
  • Participants writing with AI assistants were more likely to write about stance-congruent topics introduced by the AI.
• Experiments / evaluation:
  • Qualitative interviews revealed that AI suggestions interrupted ideation and redirected attention to evaluating pre-generated content.
  • Quantitative topic modeling confirmed that AI suggestions shaped both accepted and self-written text.
  • Statistical models demonstrated the influence of AI suggestions on topic frequency and self-written content.
• Limitations and future work:
  • The study’s experimental setting may amplify effects compared to real-world systems, limiting ecological validity.
  • Findings are based on a deliberately opinionated AI assistant; future work should explore less intrusive designs and broader populations.
  • Further research is needed to examine long-term impacts of reactive writing on cognitive engagement and public discourse.

Summary

This paper introduces the concept of Reactive Writing, a model describing how AI writing assistants reshape the writing process by shifting focus from ideation to evaluating and elaborating on AI-generated suggestions. Through a mixed-methods study, the authors demonstrate that AI suggestions significantly influence both the process and content of writing, often introducing new topics and framings that persist in final texts. While participants maintained a sense of control, the AI’s suggestions subtly shaped their writing direction and ideas. These findings highlight the persuasive potential of AI writing assistants and their role as algorithmic agenda-setters, raising important implications for system design and public discourse.