PaperlayerView exact revision on arXiv

Source-linked AI summary

When Teams Embrace AI: Human Collaboration Strategies in Generative Prompting in a Creative Design Task

Yuanning Han, Ziyi Qiu, Jiale Cheng, RAY LC

arXiv:2509.21731v1cs.HC

TL;DR

Research on GenAI-assisted creativity has focused largely on individual prompting, leaving open how human teams collaborate through the prompting process. This paper studies co-prompting as people share and discuss prompts. It finds that GenAI can support consensus and creative work, while prompting introduces communication demands and participants tend to prioritize human collaborators’ ideas.

  • Problem

    Research has primarily examined individual prompting, leaving limited evidence on how diverse teams work together through the challenging prompting process in creative tasks.

    Source: §1 · p. 2, §2.1 · p. 2

  • Method

    The study qualitatively investigates co-prompting, defined as people sharing and discussing GenAI prompts.

    Source: §1 · p. 2

  • Results

    Co-prompting can help teams build and adjust prompts and support discussion and consensus, but it can also increase communication costs and divert attention from creative ideas.

    Source: §1 · p. 2, §6 · p. 13

  • Takeaways & Limitations

    Collaborative design systems should leverage human-human collaboration and shared human expertise in the GenAI creative process.

    Source: §1 · p. 2, §6 · p. 13

    The study used only online collaboration, which may yield dynamics and outcomes different from offline scenarios.

    Source: §5.4.1 · p. 12

Abstract

from arXiv · show

Studies of Generative AI (GenAI)-assisted creative workflows have focused on individuals overcoming challenges of prompting to produce what they envisioned. When designers work in teams, how do collaboration and prompting influence each other, and how do users perceive generative AI and their collaborators during the co-prompting process? We engaged students with design or performance backgrounds, and little exposure to GenAI, to work in pairs with GenAI to create stage designs based on a creative theme. We found two patterns of collaborative prompting focused on generating story descriptions first, or visual imagery first. GenAI tools helped participants build consensus in the task, and allowed for discussion of the prompting strategies. Participants perceived GenAI as efficient tools rather than true collaborators, suggesting that human partners reduced the reliance on their use. This work highlights the importance of human-human collaboration when working with GenAI tools, suggesting systems that take advantage of shared human expertise in the prompting process.

RAY LC†

The figure depicts two people using GenAI to iteratively draft a stage design, while the text characterizes GenAI as an efficient tool rather than a true collaborator.

RAY LC†

  • Two participants worked as a team with GenAI to iteratively draft a stage design inspired by a Dylan Thomas poem.The figure shows the process and outcomes of their work.Source: Figure 1 · p. 1
  • Participants saw GenAI as efficient tools rather than true collaborators, suggesting human partners reduced reliance on GenAI.Source: RAY LC† · p. 1

CCS Concepts

The section assigns the work a collaborative and social computing classification.

CCS Concepts

  • The paper is classified under empirical studies in collaborative and social computing.Source: CCS Concepts · p. 1

Keywords

This section lists the paper’s keywords and publication details.

Keywords

  • The listed keywords include human-AI collaboration, teamwork in prompting, GenAI engineering, and creative co-design.Source: Keywords · p. 1
  • The paper appeared at CHI ’24, held May 11–16, 2024, in Honolulu, Hawaii.Source: Keywords · p. 1

1 Introduction

The paper examines how teams collaborate while prompting GenAI for creative design, addressing a gap in research focused mainly on individual prompting. It reports that co-prompting can support experimentation and consensus while also adding communication and strategic-prompting demands.

1 Introduction

  • The study addresses limited evidence on how diverse human teams work together through prompting, a key challenge in GenAI-assisted creative tasks.Prior work primarily examined individual prompting, which can be challenging for non-experts.Source: §1 · p. 2
  • The study uses stage design as a case of multidisciplinary creative work that combines visual and textual elements and requires collective ideation.Source: §1 · p. 2
  • The researchers define co-prompting as people sharing and discussing prompts, and investigate it through a qualitative approach.Source: §1 · p. 2
  • Co-prompting helps teams build and adjust prompts, but can increase communication costs and divert discussion from creative ideas.Participants also reported experimenting, sharing ideas, and developing mutual understanding through the process.Source: §1 · p. 2
  • GenAI may support creative ideation and team collaboration, but strategic prompting can add mental demands that may reduce team performance.Participants still prioritized their own ideas and those of human teammates.Source: §1 · p. 2
  • The study identifies human co-prompting strategies and offers design insights for collaborative creative systems that serve multiple users.Source: §1 · p. 2

2 Background

Prior work documents prompting difficulties and AI perceptions largely in individual creative interactions, motivating study of human–AI team dynamics. Stage design provides a multidisciplinary setting for examining that collaboration.

2.1 Human-AI Creative Collaboration

  • Prior work finds non-expert users struggle to start, choose instructions, and build or modify prompts, while designers struggle to translate ideas into precise prompt words and ensure comprehensiveness and accuracy; most findings concern single users, not teams.Source: §2.1 · pp. 2–3

2.2 Perception of AI in Creative Processes

  • Prior creative-work studies describe GenAI as a friend, coworker, or inspiration source, but team-AI perception remains an open question.Other concerns include bias, plagiarism, diminished creative ideas, and AI’s limits as a decision-maker.Source: §2.2 · p. 3

2.3 Human Collaboration in Stage Design

  • Stage design combines interdisciplinary teamwork, collective ideation, and translating written material into visual art, making it a relevant setting for studying team-AI collaboration.The study focuses on collaboration from script reading to conceptual sketching, not promoting GenAI applications in the stage-design industry.Source: §2.3 · p. 3

3 Methods

The study paired students and recent graduates with design or performance backgrounds for an online stage-design workshop using GenAI. Interviews and qualitative analysis examined their collaboration and prompting processes.

3.1 Participants and Recruitment

  • The 18 participants were assigned to nine pairs; most were students or recent graduates with art/design backgrounds, and none had professional stage-art-design experience.Twelve participants majored in art or design, three had performance backgrounds, and all had design or performance experience.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Workshop Design

  • Pairs created at least five stage-design sketches or references inspired by Dylan Thomas’s “Do Not Go Gentle Into That Good Night,” using ChatGPT and Midjourney.The task drew on the poem as inspiration, and Midjourney was selected after pilot testing for its ease of use for non-experts.Source: §3.2 · p. 4
  • A warm-up introduced text prompting, after which participants used text prompts to generate Midjourney images and selected five outputs.Participants shared screens, discussed ideas, and could receive remote operating support when access or network problems arose.Source: §3.2 · p. 4, Figure 3 · p. 4

3.3 Interview Protocol

  • Researchers conducted individual remote interviews lasting 20–30 minutes to elicit participants’ accounts of teamwork, GenAI use, challenges, and strategies.Interviews were conducted in Mandarin, transcribed, and translated into English.Source: §3.3 · p. 4

3.4 Data Analysis

  • Two Chinese-speaking researchers independently coded interview transcripts, discussed disagreements, and iteratively refined codes and themes.They documented coding decisions and used regular meetings to compare decisions and refine the analysis.Source: §3.4.1 · pp. 4–5

3.4.2 Co-prompting process.

  • A researcher reviewed recordings to map prompting details and analyze users’ stages, interactions, challenges, and insights from prompts to generated images.The analysis used a user-journey map of interactions with ChatGPT and Midjourney.Source: §3.4.2 · p. 5

4 Results

Teams used two workflows: some developed a performance plot before scene details and images, while others established concepts before developing the plot. These approaches appeared related to participants’ backgrounds and patterns of ChatGPT use.

4.1 Overall Strategies

  • Four groups—2, 3, 8, and 9—followed a plot-based workflow, first collaboratively creating a comprehensive narrative and detailed performance plot-line, then elaborating the stage-scene specifics based on that narrative.Source: §4.1 · p. 5
  • Five groups used a concept-based workflow, defining themes, style, or central imagery before crafting the performance storyline and scenes.The groups were 1, 4, 5, 6, and 7.Source: §4.1 · p. 5
  • The workflows appeared related to participants’ professional backgrounds, and plot-based groups seemed to use ChatGPT more than concept-based groups.The authors also report an interplay between co-prompting and collaboration during design.Source: §4.1 · p. 5

4.2 Collaboration Influences Co-prompting

Participants used GenAI to rapidly visualize ideas while exchanging opinions about prompts and images. Co-prompting helped articulate concepts and adjust prompts, but communicating intent remained difficult and could make iteration burdensome.

4.2 Collaboration Influences Co-prompting

  • GenAI rapidly produced visuals of different concepts, while teammates exchanged opinions about prompt words and generated images.Participants sought collaborators’ opinions during co-prompting, and teams discussed both prompt wording and ideas represented in the outputs.Source: §4.2 · pp. 5–6
  • Teams developed abstract poem-inspired ideas such as “sad” into concrete prompts such as “dark room” and “whisper.”Co-prompting gave participants space to flesh out ideas that were still conceptual and hard to articulate.Source: §4.2.1 · p. 6

4.2.2 Co-prompting to overcome difficulties in adjusting the prompt.

  • Participants said GenAI often followed literal prompt meanings but missed underlying intentions or emotional nuance, requiring repeated clarification and modification.They also described its limited capacity for association as a hindrance.Source: §4.2.2 · p. 6
  • When an image missed the intended father-son relationship and sad atmosphere, participants revised prompt wording together and speculated about Midjourney’s interpretation.The intended son was a young man, but generations depicted a child and lacked the desired sadness.Source: §4.2.2 · p. 6

4.2.3 Communication with AI can still be difficult even with the help

  • Repeated clarification could feel draining and time-consuming; participants sometimes abandoned attempts after a few unsuccessful generations and reported greater caution in teams.Participants said the effort of precise group iteration could discourage further attempts and that they were careful to avoid mistakes that might reduce efficiency or let teammates down.Source: §4.2.3 · p. 6
  • Changing “Rock” to “Rock music” produced guitars but still failed to convey the intended rock-and-roll style, so the participants stopped trying.The initial prompt yielded a stone image.Source: Figure 7 · p. 7
  • Participants also sought teammates’ advice on prompt wording before entering or revising prompts.Both participants shown asked for opinions before finalizing their prompts.Source: Figure 8 · p. 7

4.3 Co-prompting Influences Collaboration

In these teams, GenAI outputs supported rapid experimentation and gave collaborators concrete material for sharing ideas, revising prompts, and reaching mutual understanding or consensus.

4.3.1 Co-prompting inspires discussion and leads to testing of ideas.

  • GenAI enabled rapid, low-labor trials that encouraged participants to experiment and iterate on ideas.Participants said image generation and revisions were faster and less laborious than manual production.Source: §4.3.1 · p. 7

4.3.2 Participants wanted to hear more from their teammates in co-prompting.

  • Participants invited teammates to help craft prompts, assess generated results, and refine prompts based on peer feedback.They sought input both while drafting prompts and after outputs appeared.Source: §4.3.2 · p. 7, Figure 9 · p. 8

4.3.3 Co-prompting can enhance mutual understanding in teams.

  • Co-prompting helped participants express concepts that were difficult to convey verbally, making abstract ideas more tangible to collaborators.Participants found GenAI useful for communicating core ideas even when its output did not fully align with the original concept.Source: §4.3.3 · p. 7
  • A collaborator’s interpretation of generated images changed P13’s view of a set she might otherwise have dismissed.The shared representations gave partners material to assess and discuss their interpretations.Source: §4.3.3 · pp. 7–8
  • When generated content dissatisfied both partners, continued discussion helped them iteratively modify prompts until they were mutually satisfied.This pattern is illustrated by P15 and P16.Source: Figure 10 · p. 8
  • For teams with conflicting ideas, ChatGPT-generated stories and other outputs offered a concrete basis for discussing feasibility and reaching consensus.P17 and P18 used narratives to deliberate on their ideas; their process moved from difficulty convincing each other to agreement through discussion.Source: §4.3.3 · pp. 8–9, Figure 11 · p. 8

4.4 Participants’ Perception Towards GenAI and Humans in Collaboration

Participants used GenAI for laborious content-generation tasks while retaining direction and evaluation. They valued both outputs aligned with their ideas and unexpected inspiration, while often trusting human collaborators more.

4.4 Participants’ Perception Towards GenAI and Humans in Collaboration

  • Participants generally treated GenAI as a helpful tool for laborious work, though some described it as a teammate or consultant.Many said it could not actively join discussions, understand emotional nuance, or participate in team decisions.Source: §4.4.1 · p. 9, §4.4 · p. 9

4.4.1 Human is responsible for guiding the GenAI to fulfill the laborious task.

  • Participants assigned GenAI content-generation and support tasks while retaining responsibility for direction and guidance.Uses included extracting keywords, translation, organizing materials, and deriving design recommendations; participants described the tool as saving time and labor.Source: §4.4.1 · p. 9

4.4.2 Participants reported a preference for GenAI’s output to match

  • Participants evaluated outputs, selected those matching their expectations, and revised prompts to steer subsequent image generation.They recognized aesthetic qualities and emphasized keywords to guide GenAI.Source: §4.4.2 · p. 9
  • Participants also welcomed unexpected outputs, including a broken moon and imagery that gained meaning through its connection to the poem.P14 selected the broken-moon image, while P3 and P4 ultimately chose an image after recognizing its connection to the poem’s lines.Source: §4.4.2 · p. 9, Figure 12 · p. 10, Figure 13 · p. 10
  • When GenAI did not provide hoped-for unexpected insights, P3 reported disappointment and said he might stop using it.He wanted ideas beyond what his prompts specified.Source: §4.4.2 · p. 9

4.4.3 Participants perceived the original ideas of humans to be at a higher category than those of GenAI.

  • Participants worried GenAI could diminish human creativity and often trusted teammates’ judgments more than AI output.They valued human ideas and reported that teammate suggestions often led to more diverse results; some also felt the workshop left too little time for creative discussion.Source: §4.4.3 · p. 9

5 Discussion

Co-prompting can support team collaboration and creative ideation, but prompt-related demands may distract teams from creative discussion. Participants also tended to place human collaborators above GenAI, informing the paper’s design recommendations and limitations.

5.1 RQ1: Challenges for Teams in Co-Prompting GenAI for Creative Design and Strategies for Overcoming Them

  • Co-prompting both supports creative collaboration and creates costs: prompting remains difficult and can divert attention from creative discussion.Participants reported communication complexity, trouble expressing intentions and adjusting prompts, and possible reduced diligence in ideation when GenAI outputs satisfied them.Source: §5.1 · p. 10, §5.1.1 · pp. 10–11

5.1.2 Strategies in collaboration could enhance prompting

  • Teams structured prompts, exchanged suggestions, and assigned laborious tasks to GenAI, using outputs to test ideas and make abstract concepts more tangible.Human partners also provided help with prompting; generated visuals supported mutual understanding and more efficient communication.Source: §5.1.2 · p. 11

5.2 RQ2: Perception towards the roles of GenAI and human collaborators in co-prompting

  • Participants placed human collaborators above GenAI, while expecting outputs to reflect their ideas and sometimes welcoming unexpected inspiration.They sometimes followed AI suggestions and later worried this could narrow exploration; observed workflows varied with participants’ backgrounds, and the paper speculates education may have influenced them.Source: §5.2 · p. 11

5.3 Design Implications

  • The authors recommend collaborative systems that support human-like interaction, adapt to design stages, and help users reflect rather than rely on GenAI suggestions.They propose varied outputs for brainstorming, more faithful visualization during concept validation, and interactions tailored to team members’ performance.Source: §5.3 · p. 12

5.4 Limitations

  • Because the study used online collaboration, its dynamics may differ offline, where prompt entry could shift more often between participants.The authors note that online and offline strategies could also be similar, since one person may enter prompts in either setting.Source: §5.4.1 · p. 12

5.4.1 Online collaboration setting.

  • English-trained models were used by native Chinese speakers prompting in English, so linguistic and cultural nuances may have affected interpretation and generation.The authors suggest exploring models trained specifically on Chinese-language data.Source: §5.4.2 · p. 12

5.4.3 Lack of control study of interaction without GenAI.

  • Without a GenAI-free comparison, the study describes team-AI engagement; its student sample and two-person teams also constrain generalization.The authors identify possible differences in how larger teams divide prompting and rely on GenAI.Source: §5.4.3 · p. 12, §5.4.4 · p. 12, §5.4.5 · p. 12

5.4.5 Limited team size and generalizability.

  • Findings are specific to paired, poem-inspired stage-design work involving text and image generation, and the analysis focused mainly on qualitative data.The authors propose analyzing generated content, prompts, and verbal behavior in future work.Source: §5.4.6 · p. 12, §5.4.7 · pp. 12–13

6 Conclusion

The study examines how teams collaborate while co-prompting GenAI for stage-design work. It finds that GenAI can aid creative work and consensus-building, but also presents communication and strategic-prompting challenges; participants prioritized human collaborators’ ideas.

6 Conclusion

  • The study paired students and designers with performance and design backgrounds to work with GenAI on stage-performance art design.The intervention was designed to observe collaboration dynamics.Source: §6 · p. 13
  • GenAI in creative teamwork can facilitate creative work and consensus-building while introducing communication and strategic-prompting challenges.Participants tended to prioritize human collaborators’ ideas and suggestions over GenAI’s.Source: §6 · p. 13
  • The authors emphasize human-human collaboration and recommend collaborative design tools that harness human perspectives and abilities alongside GenAI.They present the findings as practical guidance for researchers and practitioners developing tools in other domains.Source: §6 · p. 13
Loading 2509.21731v1…