[论文解读] Beyond Prompts: Exploring the Design Space of Mixed-Initiative Co-Creativity Systems

Zhiyu Lin, Upol Ehsan+4 more|arXiv (Cornell University)|May 3, 2023

Creativity in Education and Neuroscience被引用 11

一句话总结

本文定义了混合主动协同创作（MI-CC）系统的设计空间，实施了7种系统变体来覆盖该空间，并显示较广的设计空间覆盖能提升感知的创造力支持与成就感，且用户偏好随专业水平而异。

ABSTRACT

Generative Artificial Intelligence systems have been developed for image, code, story, and game generation with the goal of facilitating human creativity. Recent work on neural generative systems has emphasized one particular means of interacting with AI systems: the user provides a specification, usually in the form of prompts, and the AI system generates the content. However, there are other configurations of human and AI coordination, such as co-creativity (CC) in which both human and AI systems can contribute to content creation, and mixed-initiative (MI) in which both human and AI systems can initiate content changes. In this paper, we define a hypothetical human-AI configuration design space consisting of different means for humans and AI systems to communicate creative intent to each other. We conduct a human participant study with 185 participants to understand how users want to interact with differently configured MI-CC systems. We find out that MI-CC systems with more extensive coverage of the design space are rated higher or on par on a variety of creative and goal-completion metrics, demonstrating that wider coverage of the design space can improve user experience and achievement when using the system; Preference varies greatly between expertise groups, suggesting the development of adaptive, personalized MI-CC systems; Participants identified new design space dimensions including scrutability -- the ability to poke and prod at models -- and explainability.

研究动机与目标

• 提出并把 MI-CC 系统的设计空间框架付诸实施（三个轴线：发起者、 elaboration/reflection、global/local）。
• 调查不同通信配置如何影响用户感知的创造力支持与故事生成中的目标达到。
• 评估更广的设计空间覆盖是否在不同专业水平的用户中提升体验。
• 识别用户报告的新设计空间维度（如 explainability/scrutability）及其对协作的影响。

提出的方法

• 构建一个具有三个轴线的假设 MI-CC 设计空间（Human vs. Agent-initiated, Elaboration vs. Reflection, Global vs. Local）。
• 实例化 7 个 MI-CC 故事系统变体，代表设计空间的子集以进行探索性研究。
• Leverage the Creative Wand 框架 with four components (Creative Context, Experience Manager, Communications, Frontend) to run experiments.
• Use two AI storytelling systems (Plug and Blend with topic control and GPT-J, and CARP for critique) to implement communication modules.
• Measure outcomes with the Creative Support Index and task performance on a constrained 10-line story.
• Conduct a between-subjects study with 185 participants recruited via Prolific, randomizing system condition and counter-balancing presentation order.

实验结果

研究问题

• RQ1 Does broader coverage of the MI-CC design space improve perceived creative support and goal achievement?
• RQ2 How do different communication types (agent-initiated, human-initiated, elaboration, reflection, global, local) influence user experience across expertise levels?
• RQ3 What new dimensions (e.g., scrutability, explainability) emerge from user studies of MI-CC systems?

主要发现

Overall	Agent-Initiated Only	Human-Initiated Only	Elaboration Only	Reflection Only	Global Only	Local Only
Q1: Expressiveness	62.2%*	74.2%*	46.9%	56.7%	78.1%*	63.0%	54.5%
Q2: Enjoyment	60.5%*	74.2%*	43.8%	50.0%	81.2%*	59.3%	54.5%
Q3: Exploration	62.7%*	71.0%*	46.9%	56.7%	71.9%*	70.4%*	60.6%
Q4: Immersion	62.2%*	71.0%*	50.0%	60.0%	75.0%*	59.3%	57.6%
Q5: Collaboration	59.5%*	71.0%*	40.6%	56.7%	81.2%*	59.3%	48.5%
Q6: Result worth effort	60.5%*	64.5%+	53.1%	60.0%	71.9%*	66.7%+	48.5%
Q7: Better responses	61.6%*	67.7%*	56.2%	63.3%	78.1%*	59.3%	45.5%

• Wider design-space coverage yields equal or better ratings on creative support and perceived achievement across multiple metrics.
• Removing dimensions such as agent-initiated or elaboration communications degrades the creative experience across measures of expressiveness, enjoyment, exploration, immersion, collaboration, and perceived result quality.
• Global-only communications reduce exploration and perceived results worth the effort, highlighting the importance of local changes for creativity exploration.
• Participant preferences vary by expertise and familiarity with AI, suggesting the need for adaptive, personalized MI-CC configurations.
• Users value controllability and scrutability, and express a desire for explainability to build trust and improve mental models of AI behavior.