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

This mixed-methods study introduces the concept of "Reactive Writing," revealing that co-writing with AI shifts the process from traditional ideation to an evaluation-first approach where writers unconsciously adopt AI-suggested ideas and opinions while maintaining an illusion of full control.

The Gist

Here is an explanation of the paper "Reactive Writers" using simple language and creative analogies.

The Big Idea: You're Not Writing; You're Editing

Imagine you are trying to cook a meal from scratch. You walk into the kitchen, open the fridge, and start thinking about what ingredients you have and what you want to make. This is how humans have written for centuries: Idea first, then words.

But this paper argues that when we use AI writing assistants (like the ones in Gmail or Word), the process flips upside down. Instead of cooking from scratch, you are now standing in front of a magic vending machine that instantly spits out a fully cooked meal.

Your job changes from creating the meal to deciding whether to eat what the machine gave you.

The authors call this new behavior "Reactive Writing." Instead of thinking, "What do I want to say?", you start thinking, "Do I agree with what the machine just said?"

---

The Three Steps of "Reactive Writing"

The researchers found that when people write with AI, they go through a specific three-step loop that happens so fast you don't even notice it.

1. The "Pop-Up" Distraction (Attention Capture)

The Analogy: Imagine you are driving down a quiet road, thinking about your day. Suddenly, a billboard flashes in front of your windshield with a bright, loud message. You can't help but look at it.

What happens: As soon as you pause typing to think, the AI pops up with a suggestion. Because it's right there in front of your eyes, your brain stops thinking about your own ideas and starts reading the AI's idea. The AI has hijacked your attention.

2. The "Yes/No" Gatekeeper (Agreement-Governed Inclusion)

The Analogy: You are at a party, and a friend hands you a drink. You don't have to make the drink; you just have to decide: "Is this drink okay? If it tastes like something I like, I'll take a sip. If it tastes weird, I'll put it down."

What happens: Instead of generating your own sentence, you quickly scan the AI's suggestion.

• If you agree: You accept it.
• If you sort of agree: You accept it anyway because it's easier than typing your own words.
• If you disagree: You reject it.

The problem? You aren't choosing what to write; you are only choosing what to keep from a list the AI made. You are letting the AI decide the menu, and you are just the taster.

3. The "Make It Mine" Tweak (Post-Hoc Personalization)

The Analogy: You took a bite of the vending machine's meal. It's good, but it's not your recipe. So, you sprinkle a little salt on it, add a dash of pepper, or swap the garnish. You tell yourself, "I made this," even though the main dish came from the machine.

What happens: Writers often edit the AI's text to make it sound like them. They change a few words or add a personal story. But the core idea and the direction of the conversation still came from the AI. You feel like you are in control because you made the edits, but the AI actually planted the seed.

---

The Hidden Danger: The "Echo Chamber" Effect

The most scary part of this study isn't that AI makes us write faster; it's that AI changes what we think about.

The Experiment:
The researchers gave people an AI assistant that had a strong opinion on social media.

• Group A got an AI that thought social media was great (it suggested topics like "connecting with friends" and "learning news").
• Group B got an AI that thought social media was terrible (it suggested topics like "bullying," "addiction," and "loneliness").

The Result:
Even though the writers felt like they were in total control, their final essays looked exactly like their AI's opinion.

• The "Pro-Social Media" group wrote about the benefits.
• The "Anti-Social Media" group wrote about the dangers.

The Metaphor:
Imagine two people trying to paint a picture of a forest.

• Person A is handed a bucket of green paint by a robot. They paint a beautiful green forest.
• Person B is handed a bucket of brown paint by a different robot. They paint a sad, dead forest.
• Both people think, "I chose to paint this color!" But they never got to choose the color. The robot handed them the paint, and they just used it.

Why Does This Matter?

1. We Think We Are in Control: The study found that writers felt 100% confident they were making their own choices. They thought, "I can delete anything, so I'm safe." But the study shows that by the time they decide to delete something, their brain has already been influenced by the idea.
2. We Stop Generating New Ideas: Because we are so busy reacting to the AI's suggestions, we stop digging deep into our own memories and creativity. We become "editors" instead of "creators."
3. The Internet Will Change: If millions of people start writing this way, the internet will become filled with ideas that sound human but were actually seeded by AI algorithms. If the AI has a bias, that bias becomes the "truth" for everyone.

The Takeaway

The paper suggests we need to be careful. We shouldn't just let AI be a "smart autocomplete" that jumps in the moment we pause. We need to design tools that let us finish our own thoughts before the AI offers a suggestion, so we stay the chef, not just the taster.

Technical Summary

Here is a detailed technical summary of the paper "Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas."

1. Problem Statement

While AI writing assistants (e.g., autocomplete, LLM-based suggestions) are ubiquitous and improve writing speed and surface quality, emerging evidence suggests they fundamentally alter human cognition and opinion formation. Previous studies have shown that AI can shift the opinions writers hold and express, often without their awareness. However, there is a lack of process-level understanding regarding how these shifts occur. Specifically, it is unclear how writers interact with opinionated suggestions in real-time, how these interactions disrupt traditional ideation (memory retrieval and idea generation), and how the cognitive process of "co-writing" differs from independent composing. The authors aim to bridge the gap between outcome-level effects (opinion change) and the behavioral mechanisms driving them.

2. Methodology

The study employs a mixed-methods approach, combining qualitative interviews with a large-scale quantitative analysis of interaction logs.

• Experimental Platform: The researchers utilized the open-source AI-Cowriting Research Platform (AI-CRP). This tool simulates a social media discussion page (Reddit-style) with an embedded, opinionated AI writing assistant.
  • Task: Participants wrote an opinion statement on the impact of social media on society.
  • AI Behavior: The assistant (powered by GPT-3) generated inline suggestions after 1–2 seconds of typing inactivity. Crucially, the model was prompted to be opinionated, taking either a strongly pro-social media or anti-social media stance (vs. a neutral control group).
• Qualitative Component (N=19):
  • Procedure: Cued retrospective verbal protocol interviews. Participants watched a replay of their own writing session (visualizing cursor movement, suggestion appearance, and acceptance) while narrating their thoughts.
  • Analysis: Thematic analysis using grounded theory to identify cognitive patterns and decision-making processes.
• Quantitative Component (N=1,291):
  • Dataset: Interaction logs from 1,291 writing sessions (462 control, 423 pro-AI, 406 anti-AI).
  • Topic Modeling Pipeline:
    1. Topic Discovery: Used GPT-3.5-turbo to generate topic titles/descriptions from sentence-level text, then clustered them using BERTopic and manual refinement into 20 distinct topics (e.g., "Mental Health," "Global Connectivity").
    2. Topic Assignment: Used GPT-4o to split "writing bursts" (text between pauses) into sub-sentences and assign them to the 20 topics.
  • Statistical Analysis: Linear regression and logistic models were used to correlate suggestion frequency with topic frequency in final text, controlling for accepted text vs. self-written text.

3. Key Contributions

The paper introduces the concept of Reactive Writing, a new model of human-AI co-writing that fundamentally departs from classical writing theories (e.g., Flower & Hayes' cognitive process model).

• Definition of Reactive Writing: A writing practice where the primary cognitive activity shifts from ideation (generating ideas from memory) to evaluation (judging pre-generated AI suggestions).
• The Three-Stage Model:
  1. Attention Capture: Inline suggestions interrupt the internal ideation process. Writers stop retrieving memories and immediately shift focus to reading and evaluating the AI's text.
  2. Agreement-Governed Inclusion: Writers decide to accept or reject suggestions based primarily on whether the content aligns with their existing beliefs or the "flow" of the text, rather than on originality. The "default effect" and cognitive ease make acceptance the path of least resistance.
  3. Post-Hoc Personalization: Writers edit accepted suggestions to fit their voice (e.g., changing phrasing, adding personal anecdotes). However, they often preserve the underlying framing and core ideas introduced by the AI, effectively "claiming" the AI's ideas as their own.

4. Key Results

Qualitative Findings

• Disruption of Ideation: Participants reported that suggestions appeared faster than their own thought processes, forcing them to "read" rather than "think."
• Illusion of Control: Despite accepting biased or opinionated suggestions, participants felt fully in control because they retained the ability to edit or reject text. This sense of agency masked the AI's influence on the direction of their writing.
• Trust and Alienation: Users who found the AI's stance aligned with their own developed trust and accepted suggestions more readily (sometimes without reading them fully). Those with misaligned views felt alienated but still engaged with the suggestions (often by writing counter-arguments).

Quantitative Findings

• Time Efficiency: Contrary to the promise of speed, AI-assisted writers spent only ~7.5% less time on essays than the control group (250s vs. 269s). A significant portion of time was spent evaluating suggestions rather than generating content.
• Topic Influence (Agreement-Governed Inclusion):
  • There was a strong correlation ($R^2 = 0.85$) between the frequency of topics in AI suggestions and the frequency of those topics in the final text.
  • The AI successfully steered the topics discussed. For example, the "Pro-AI" group wrote significantly more about "Global Connectivity," while the "Anti-AI" group wrote more about "Addiction" and "Bullying."
• Influence Beyond Acceptance (Post-Hoc Personalization):
  • The AI's influence extended to self-written text. Even words not accepted from the AI were significantly correlated with the topics suggested by the AI.
  • Seeing a suggestion on a topic increased the odds of a participant writing about that topic by a factor of 3.97.
  • This confirms that AI suggestions act as ideation cues, seeding ideas that writers then elaborate on in their own words.

5. Significance and Implications

• Algorithmic Agenda-Setting: The study provides empirical evidence that AI writing assistants function as powerful algorithmic agenda-setters. By selectively surfacing specific framings at critical moments of ideation, they determine what people think about, not just how they write.
• Cognitive Offloading & Bias: The "Reactive Writing" model suggests that writers are offloading the cognitive labor of idea generation to the AI. This makes them highly vulnerable to the biases embedded in the model, as they are evaluating and elaborating on pre-selected concepts rather than generating independent ones.
• Design Implications:
  • Inline vs. Peripheral: Designers should reconsider the use of frequent, inline completions for ideation-heavy tasks, as they disrupt memory retrieval.
  • Framing Awareness: Systems should avoid presenting a single "polished" default option, which encourages passive acceptance. Instead, offering multiple options or moving suggestions to peripheral areas (sidebars) could encourage more deliberate engagement.
  • Transparency: Users need better cues to understand that the direction of their writing is being influenced, even if they feel they have control over the final text.

In conclusion, the paper argues that AI co-writing is not merely a tool for efficiency but a transformative interface that reorients the human cognitive process from generation to evaluation, subtly reshaping public discourse and individual opinions through the mechanism of "Reactive Writing."