Two-Player Alternate Uses Test: A Controlled Testbed for Interactive Human-AI and Human-Human Co-Creation

Jul 16, 2026

Upload by Miles

Two-Player Alternate Uses Test (AUT) introduces a controlled, two-participant platform to compare human–human and human–AI co-creation under matched, interactive conditions, while using calibrated non-interactive baselines. The testbed decomposes co-creative performance into participant traits, partner perceptions, and content dynamics. An in-person pilot with 62 participants shows GPT-4-backed originality equals human partnering under matched time limits. Moderation effects include BAS Drive and cognitive outsourcing, plus a seeding intervention.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

Contributions

Two-Player Alternate Uses Test: A Controlled Testbed for Interactive Human–AI and Human–Human Co-Creation Controlled, Interactive Human-AI Co-Creation Babak Hemmatian¹, Anita Keshmirian², Yijun Lin, Shravan Ramamoorthy³, Mostafa Jahadakbar⁴, Elena Khuri-Reid⁴, Jingyu Wang⁵, Sina Hadjarab⁶, Sigrid Veum², Pranav Gupta⁴, Deepak Somaya⁴, Lav Varshney⁴ ¹AI Innovation Institute, Stony Brook University ²Forward College Berlin ³Alinea ⁴University of Illinois Urbana-Champaign ⁵University of California San Diego ⁶University of Chicago Controlled research on AI ideation typically compares independent agents, while field studies of human–AI collaboration sacrifice experimental control. We introduce a controlled, two-player extension of the Alternate Uses Test (AUT) that enables comparison of human–human and human–AI co-creation under matched interactive conditions, alongside calibrated non-interactive baselines. The platform supports decomposition of performance into three typically confounded factors: participant traits, partner perceptions, and content dynamics. An in-person pilot (N = 62) demonstrates its utility. Under matched time limits, originality with a GPT-4 partner is statistically equivalent to that with a human partner. Approach motivation (BAS Drive) moderates whether interactive partnership benefits originality, and self-reported cognitive outsourcing predicts lower originality specifically in human–human dyads. Prior exposure to highly creative ideas improves later performance, suggesting a “seeding” intervention. We release the platform, code, and dataset as a shared testbed for controlled studies of human–AI co-creation. CCS CONCEPTS • Human-centered computing → Human computer interaction (HCI); Interactive systems and tools; Empirical studies in HCI; • Applied computing → Psychology. Additional Keywords and Phrases: human–AI co-creation; collaborative creativity; creativity support tools; creative cognition; generative AI; perspective taking; Alternate Uses Test 1 INTRODUCTION Large language models are deployed for creative tasks once thought exclusive to humans, and a psychology of AI creativity has emerged. Two research traditions dominate. Experimental work places humans and LLMs in parallel divergent-thinking tasks and compares their independent outputs (e.g., [14, 26]), yielding clean causal identification but no purchase on what happens during co-creation. Field studies of human–AI teaming in contexts like classrooms and workplaces (e.g., [3, 6]) preserve ecological validity but sacrifice experimental control over the interaction. The middle ground, controlled study of real-time co-creation between a human and an AI compared with human-human interaction, is less studied. A handful of recent studies begin to close this gap. Doshi and Hauser [8] show that one-shot exposure to AI-generated story ideas raises individual originality while narrowing collective diversity, without conversation between participant and AI. Ashkinaze et al. [1] scale this passive-exposure paradigm to a large dynamic experiment. Shaer et al. [23] embed an LLM in asynchronous group brainwriting and validate LLM-based idea evaluation. Closest in design, Maier et al. [17] randomize interaction structure in one-on-one human–AI chat. Each isolates a different slice of the problem, but none preserves a classical divergent-thinking task across matched human–human, human–AI, and non-interactive conditions in a single apparatus so that interactivity, partner identity, conversation structure, and content exposure can be disentangled. This gap connects to a long-running C&C concern with creativity as a situated, dynamically unfolding interactive process. Prior C&C work has quantified interaction dynamics in open-ended co-creation [5] and articulated design principles for socially interactive systems [16]; more recent work frames turn-taking, contribution type, communication, and feedback as the dimensions that shape human–AI co-creat

FAQs

AUT为什么需要从单人版本扩展到“双人”版本？

传统研究要么缺少互动控制，要么缺少现实性对比。双人AUT在同一装置下匹配交互条件，用于同时拆分互动性、伙伴身份与内容暴露对结果的影响。

平台如何实现对“互动”和“非互动”条件的公平比较？

平台在固定时间限制与匹配指令下运行两名参与者的共享聊天；第三种非互动条件用预先评分的创意集合替代现场伙伴，以便与校准基线进行对照。

试点结果表明人类与GPT-4的原创性表现是否存在差异？

在匹配时间限制下，与GPT-4伙伴的原创性与与人类伙伴的原创性在统计上等价。BAS Drive会调节互动伙伴是否带来原创性收益，而认知外包会在“人-人”对中降低原创性。

哪些因素会影响互动协作对原创性的作用？

BAS Drive会调节互动伙伴对原创性的收益；自报的认知外包能预测更低原创性，且该效果特定出现在人-人组合中。此前接触高度创意点子也会通过“种子”干预提升后续表现。

Jul 16, 2026

Upload by Miles