TLDR: A study investigated how different interface metaphors (agentic, tool-like, magic-like) in AI co-writing systems affect writers’ perceptions of agency (control) and ownership (authorship). It identified three types of agency (explicit, implicit, writing process control) and three types of ownership (stylistic, conceptual, effort-based). The research found that tool-like interfaces foster expectations of explicit control, while agentic interfaces encourage a focus on conceptual contributions. The paper concludes with design recommendations for AI co-writing systems to better align with writers’ expectations and enhance their creative experience.
The world of creative writing is undergoing a significant transformation with the advent of Artificial Intelligence (AI) co-writing systems. While these tools offer exciting new possibilities for authors, they also bring forth complex questions about who is truly in control and who ultimately owns the creative output. A recent study delves into these very issues, exploring how the design of an AI system’s interface can profoundly impact a writer’s sense of agency and ownership.
Traditionally, a writer’s agency refers to their feeling of control during the writing process, while ownership relates to their self-identification as the author of the final piece. As AI becomes more integrated, these long-held ideals are challenged, potentially hindering the widespread adoption of these powerful tools.
Exploring Interface Metaphors
To understand these dynamics, researchers developed three distinct AI co-writing systems. Crucially, all three systems had identical underlying functionality, but their interfaces presented the AI in different ways, using what are called ‘interface metaphors’:
- Agentic: This metaphor personified the AI, making it feel like a collaborative partner, often through conversational interfaces.
- Tool-like: This approach presented the AI as a set of functions and controls, much like a software tool with buttons and sliders.
- Magic-like: This metaphor emphasized seamless, automatic functions with minimal controls, often using magical iconography like wands or sparkles.
Eighteen professional and non-professional writers were interviewed after using one of these prototype systems. The goal was to uncover how these different interface presentations influenced their perceptions of control and authorship.
A Taxonomy of Agency and Ownership
The study’s analysis revealed a nuanced understanding of agency and ownership, categorizing them into several subtypes:
Types of Agency (Control):
- Explicit Control: Writers felt in control through direct interface actions, such as adjusting parameters or giving instructions.
- Implicit Control: Control was perceived through the influence of the writer’s existing text on the AI’s output, even without direct commands.
- Writing Process Control: Writers maintained control by integrating AI contributions into their established workflows, treating AI-generated content as material to be edited and shaped.
Types of Ownership (Authorship):
- Stylistic Ownership: This was derived from a writer’s unique voice, language choices, and expressive elements in the text. Professional writers often emphasized this.
- Conceptual Ownership: Based on the writer’s contribution of core ideas, themes, and narrative direction. Non-professional writers often valued this more.
- Effort-based Ownership: A sense of ownership earned through the time, energy, and cognitive effort invested in developing and refining the text. This aligns with the idea that effort fosters a sense of possession, much like the famous ‘IKEA effect’ where assembling something yourself makes you value it more. One participant humorously likened it to the old pancake batter advertisements: people felt more ownership if they had to ‘add an egg,’ even if it wasn’t strictly necessary.
How Metaphors Shape Perceptions
The study found that interface metaphors significantly influenced these perceptions:
- Tool-like Metaphors: These interfaces fostered an expectation of explicit control. Writers using these systems looked for sliders and dropdowns to fine-tune the AI’s output. Even when these controls didn’t actually change the AI’s behavior (a necessary deception for the study), the *illusion* of control made some participants feel greater agency. Others, noticing the lack of impact, adopted an experimental approach, exploring what the AI ‘would do.’
- Agentic Metaphors: Systems presented as collaborative agents led participants to focus more on conceptual contributions. They expected to converse with the AI about ideas, plot points, and character sketches. This suggests that an agentic interface can encourage writers to value the AI’s conceptual input, even if the system’s actual capabilities are limited to text generation.
Also Read:
- Navigating the Future: Enhancing Trust and Safety in Autonomous Shipping Through Explainable AI
- Decoding Slang: How AI’s Informal Language Differs from Human Expression
Designing for Better Collaboration
The research offers valuable recommendations for designers of AI co-writing systems:
- Consider Different Points of Control: Designers should understand that writers seek control in various ways – explicitly through interface elements, implicitly through their text, or through their established editing processes. The interface should align with these expectations to prevent frustration.
- Design for Conceptual Relationships with Agentic AIs: If a collaborative dynamic focused on higher-order thinking (ideas, themes) is desired, agentic metaphors can be effective. However, if the goal is to encourage focus on sentence-level language, designers might want to modify or avoid overly agentic presentations.
- Play Into or Against Expectations of Control with Tool-Like Metaphors: If tool-like interfaces are used, ensure that the controls actually work as expected to meet user expectations of explicit control. Alternatively, if controls are ambiguous, users might treat the system as an experimental playground, generating inspiring but not necessarily final text.
This study, detailed further in the paper Where Do I ‘Add the Egg?’: Exploring Agency and Ownership in AI Creative Co-Writing Systems, highlights that the way an AI co-writing system is presented is just as important as its underlying technology. By thoughtfully designing interface metaphors, developers can create systems that not only assist writers but also enhance their sense of control and authorship, fostering more satisfying and effective human-AI collaboration in the creative process.
