Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ai.llude: Encouraging Rewriting AI-Generated Text to Support Creative Expression

David Zhou, Sarah Sterman

TL;DR

This paper interrogates how AI-generated text influences the creative writing process and proposes intermediate, imperfect AI outputs as a mechanism to provoke rewriting and enhance authorial ownership. It introduces ai.llude, an instrumented editing environment that presents two suggestion modes—fluent continuations and intermediate fragments—and analyzes 27 writers' sessions to compare effects on rewriting, engagement, and reflective thinking. The findings show intermediate suggestions significantly increase rewriting and support ideation and planning, while fluent suggestions improve ease of use; writers employ both modes in diverse, context-dependent ways. The work highlights design implications for AI writing tools that prioritize user control, ownership, and the preservation of the writer’s creative process, and suggests future work in personalization and improved metrics for psychological ownership.

Abstract

In each step of the creative writing process, writers must grapple with their creative goals and individual perspectives. This process affects the writer's sense of authenticity and their engagement with the written output. Fluent text generation by AIs risks undermining the reflective loop of rewriting. We hypothesize that deliberately generating imperfect intermediate text can encourage rewriting and prompt higher level decision making. Using logs from 27 writing sessions using a text generation AI, we characterize how writers adapt and rewrite AI suggestions, and show that intermediate suggestions significantly motivate and increase rewriting. We discuss the implications of this finding, and future steps for investigating how to leverage intermediate text in AI writing support tools to support ownership over creative expression.

Ai.llude: Encouraging Rewriting AI-Generated Text to Support Creative Expression

TL;DR

This paper interrogates how AI-generated text influences the creative writing process and proposes intermediate, imperfect AI outputs as a mechanism to provoke rewriting and enhance authorial ownership. It introduces ai.llude, an instrumented editing environment that presents two suggestion modes—fluent continuations and intermediate fragments—and analyzes 27 writers' sessions to compare effects on rewriting, engagement, and reflective thinking. The findings show intermediate suggestions significantly increase rewriting and support ideation and planning, while fluent suggestions improve ease of use; writers employ both modes in diverse, context-dependent ways. The work highlights design implications for AI writing tools that prioritize user control, ownership, and the preservation of the writer’s creative process, and suggests future work in personalization and improved metrics for psychological ownership.

Abstract

In each step of the creative writing process, writers must grapple with their creative goals and individual perspectives. This process affects the writer's sense of authenticity and their engagement with the written output. Fluent text generation by AIs risks undermining the reflective loop of rewriting. We hypothesize that deliberately generating imperfect intermediate text can encourage rewriting and prompt higher level decision making. Using logs from 27 writing sessions using a text generation AI, we characterize how writers adapt and rewrite AI suggestions, and show that intermediate suggestions significantly motivate and increase rewriting. We discuss the implications of this finding, and future steps for investigating how to leverage intermediate text in AI writing support tools to support ownership over creative expression.
Paper Structure (31 sections, 4 figures, 4 tables)

This paper contains 31 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Creative Writing and Rewriting
  4. AI Writing Support Tools
  5. Motivating Rewriting
  6. Design of ai.llude
  7. Design Goals
  8. System Overview
  9. Methods
  10. Participant Recruitment
  11. Experimental Design
  12. Writing Environment
  13. Writing Evaluation Survey
  14. Rewrite Metrics
  15. Qualitative Analysis
  16. ...and 16 more sections

Figures (4)

  • Figure 1: Interface of ai.llude. The AI menu opens when the tab key is pressed and allows writers to request suggestions. User text is displayed in black and AI generated text in purple. The word count and writing time are displayed at the bottom of the editor. The editor provides basic styling controls such as bold, italic, and underline. Figure text: Alice in Wonderland (Lewis Carroll).
  • Figure 2: Writing Flows. Dots indicate activity in the text editor, either of the user typing or the user requesting an AI suggestion. The suggestions were disabled for the first 15 minutes of the session, or until 150 words had been written.
  • Figure 3: Suggestion A (fluent continuation) vs. Suggestion B (intermediate suggestion) average rewrites. Left: Rewrites of Suggestion A tend to have more AI text remaining compared to rewrites of Suggestion B. Right: Rewrites of Suggestion A tend to be more similar to their initially provided forms compared to rewrites of Suggestion B.
  • Figure 4: Writers were similarly motivated to read and evaluate both Suggestion A and B text, but more writers felt the need to rewrite Suggestion B.