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

📝 Paper Summary

Human-AI Interaction AI-Assisted Writing Algorithmic Influence

Inline AI suggestions transform writing from an active process of memory retrieval into 'Reactive Writing,' where users prioritize evaluating AI-generated text over generating their own ideas.

Core Problem

While it is known that opinionated AI can shift writers' views, we lack a descriptive understanding of the procedural cognitive changes—the moment-by-moment interactions—that cause this influence.

Why it matters:

Biased language models can subtly steer public discourse and entrench specific worldviews in the collective information ecosystem
Writers may experience 'cognitive offloading,' reducing their own critical engagement and ideation capabilities
Current understanding focuses on outcomes (final text) rather than the process (how the text was composed), limiting our ability to design for writer agency

Concrete Example: When a writer pauses to think about social media's impact, an inline AI suggests a specific argument (e.g., 'social media harms mental health'). Instead of retrieving their own memory, the writer simply evaluates and accepts this suggestion, allowing the AI to seed the direction of the essay.

Key Novelty

The Theory of Reactive Writing

Conceptualizes AI co-writing not as augmentation but as a fundamental shift in cognitive mode: from 'Proposer' (generating ideas) to 'Evaluator' (checking suggestions)
Identifies that rapid inline suggestions interrupt natural ideation loops, causing writers to build upon AI-seeded concepts rather than their original plans

Evaluation Highlights

Analysis of 1,291 AI co-writing sessions reveals that engaging with suggestions becomes the central activity, displacing traditional ideation
Qualitative interviews with 19 participants confirm writers feel in full control even while systematically adopting AI-suggested topics and opinions
Exposure to suggestions on specific topics strongly predicts the presence of those topics in the final text, even beyond directly accepted phrases

Breakthrough Assessment

8/10

Provides a crucial theoretical framework ('Reactive Writing') that explains the *mechanism* of AI persuasion, moving beyond simple outcome measurement to process analysis.

⚙️ Technical Details

Problem Definition

Setting: Human-AI co-writing of opinionated essays (topic: social media impact)

Inputs: User keystrokes and context in a rich-text editor

Outputs: Inline text completions suggestions (opinionated)

Pipeline Flow

User Interface (Frontend) -> Interaction Logger -> AI Model (Backend) -> Suggestion Display

System Modules

AI-Cowriting Research Platform (AI-CRP)

Reddit-style discussion mockup with rich-text editor

Model or implementation: Web-based application

Opinionated Suggestion Engine

Generate contextually relevant but biased text completions

Model or implementation: GPT-3-based language model

Novel Architectural Elements

Integration of explicitly opinionated suggestion prompts within a standard autocomplete interface to study persuasion mechanics

Modeling

Base Model: GPT-3-based language model

Training Method: Prompt Engineering (Opinionated Prompting)

Compute: Not reported in the paper

Comparison to Prior Work

vs. Standard Autocomplete: This study uses an *opinionated* model to deliberately introduce bias, whereas standard tools aim for neutrality or personalization
vs. Jakesch et al. (2023): Expands on prior outcome-focused work by providing a *process-level* account (Reactive Writing) of how the influence occurs [cited in paper]
vs. CoAuthor (Lee et al.) [not cited in paper]: Focuses specifically on the cognitive shift in *ideation* caused by inline suggestions, rather than just collaboration patterns or user satisfaction

Limitations

Study focuses on a specific controversial topic (social media impact), which may not generalize to neutral tasks
Inline interface design specifically triggers the 'reactive' behavior; other interfaces (e.g., chat sidebars) might not
Results rely partially on self-reported retrospective interviews, which can be subject to recall bias
Analysis snippet provided does not quantify the magnitude of the opinion shift (e.g., percentage change in sentiment)

📊 Experiments & Results

Evaluation Setup

Mixed-methods study: Controlled writing experiment followed by retrospective analysis

Benchmarks:

Social Media Opinion Task (Argumentative Writing) [New]

Metrics:

Suggestion engagement rates (accept/edit/ignore)
Opinion alignment (text vs. AI bias)
Qualitative codes (from verbal protocols)
Statistical methodology: Quantitative analysis of interaction logs; Qualitative coding of interview transcripts

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
The provided text contains the Introduction and Methods but cuts off before the Results section. Therefore, specific numeric values for interaction rates or opinion shifts cannot be extracted.

Main Takeaways

Process Shift: Writing with inline AI fundamentally changes the cognitive loop from 'think-then-type' to 'read-evaluate-accept/edit' (Reactive Writing).
Illusion of Control: Writers maintain a strong sense of agency and believe they are the authors of the ideas, even when the AI seeds the primary arguments.
Agenda Setting: AI suggestions function as algorithmic agenda-setters; even if a user edits a suggestion, the *topic* introduced by the AI often remains the focus of the subsequent text.
Disrupted Ideation: Rapid inline suggestions (after 2s pauses) interrupt the user's natural memory retrieval process, leading to reliance on external (AI) cues.

📚 Prerequisite Knowledge

Prerequisites

Cognitive Process Model of Writing (Flower and Hayes, 1981)
Basics of Large Language Models (LLMs)
Human-Computer Interaction (HCI) evaluation methods

Key Terms

Reactive Writing: A mode of writing defined by the authors where the writer's primary cognitive task shifts from generating ideas to evaluating and modifying AI-generated suggestions

Cognitive Offloading: The use of external tools (like AI) to reduce the mental effort required for tasks like memory retrieval or planning

Inline Suggestions: AI-generated text completions that appear directly at the cursor position (ghost text) while the user is typing

Cued Retrospective Verbal Protocol: A qualitative method where participants watch a replay of their task performance and verbalize their thought processes after the fact

Ideation: The cognitive process of retrieving information from memory and forming new concepts or arguments

AI-CRP: AI-Cowriting Research Platform—an open-source experimental environment used to log keystrokes and manage AI suggestions