AI as Social Glue, CHI 21′ – Paper Summary

Paper Review

Paper title: AI as Social Glue: Uncovering the Roles of Deep Generative AI during Social Music Composition
Author: Minhyang (Mia) Suh, Emily Youngblom, Michael Terry, and Carrie J. Cai
Published on: CHI 21′, Honorable Mention
DOI: https://doi.org/10.1145/3411764.3445219

TL;DR

“AI increased the ease of human-human collaboration but decreased its depth”

1. Introduction

• Current Deep Generative AI models are able to synthesize new text, images, and artwork (e.g.: ChatGPT for Text, DALL-E 2 for Images, and MuseNet for music).

• It enables an individual to do ‘creative’ work by co-creating with AI.
• People’s creativity needs to be more nuanced and complex in real-world contexts, where the creative practice is often experienced socially—for example, in co-composing and band jamming sessions.
• The previous work research focused only on (solo) human-AI collaboration, with less knowledge about the role of AI in (social) human-human collaboration.
• It raises the question: “What kind of roles can AI play in human-human collaboration for creative work?”

2. Related Work

To introduce the reader to this topic, some examples of Human-AI co-creation that were mentioned in the paper will be presented as the related work for this article.

• Human-AI co-creation on design Ideation[1]

• Human-AI co-creation on game-level design[2]

• Human-AI co-creation on music generation[3]

3. Methodology

3.1 Participant

This paper conducted a qualitative laboratory study on 30 participants (15 pairs), consisting of 5 novices, 5 hobbyists, and 5 serious composers who did not know each other before the experiment. The gender is balanced with 5 sets of female-female, 5 sets of male-male, and 5 sets of mixed gender.

Each participant was asked to co-compose 2 music, one with AI and one without AI.

3.2 Platform

In this paper, the previously published and developed algorithm, Cococo Music Generation [3] is used as the platform for collaborative human-AI music co-creation

3.3 Procedure

The experiment of this paper was conducted during COVID-19 using an online video conference application. All process was recorded and transcribed with Think Aloud Procedure.

Think aloud procedure is a method of user testing in which participants use the system while continuously thinking out loud – Verbalizing their thought as they move through the user interface[4].

The procedure was as follows

1. Each participant was asked to use the program as a tutorial for 10 minutes
2. Each pair co-composes 2 songs, one with AI and one without AI, 20 minutes for each.
3. In co-compose processes, each pair will take one card from Dixit Boardgame [5] and composes the music based on the card picked
4. Conduct a semi-structured interview after the experiment

4. Findings

4.1 Co-Creating Process

At the starting point of the co-creating process, they found that the participants usually start by setting a “social contract”,

What kind of feeling do you want to have? Any specific key that you want?

In the co-creation without AI, pairs usually:

• Splitting the task by voice (3 pairs), bar (1 pair), and/or took turns from time to time (5 pairs)
• One leading and one give feedback, while constantly asking for confirmation

In the co-creation with AI, pairs usually use AI for these 3 tasks

1. Open-end inquiries – “Let’s see what AI says”
2. Goal-directed request – “Let’s do it more minor”
3. Get some idea – “Let’s ask AI to generate this part for us and review it”

4.2 Role of AI in Social Collaboration

The second finding of this paper was the 5 primary roles of AI in social co-creative collaboration

1. AI implicitly seeds a common ground
2. AI acting as a psychological safety net
3. AI provides a force for progress
4. AI mitigating interpersonal stalling and friction
5. AI altering the creative and collaborative roles of humans

AI seeding a common ground

AI enables participants to have a mutual understanding and an initial momentum for getting the composition started

AI acting as a psychological safety net

AI provided a psychological safety net, which helped pairs to be less socially pressured, more creative, and more playful in co-composition

AI provides a force for progress

AI played a role as a force of process, providing an initial starting point, generating new ideas when stuck, and filling the unfinished work that helped the co-composition process move quickly

AI mitigating interpersonal stalling and friction

AI played a role as a mitigator of interpersonal friction, making the human-human partnership smoother and more positive

altering the creative and collaborative roles of humans

However, AI could alter the creativity and collaboration of humans, reducing the depth of co-creation and interpersonal engagement

5. Implications

AI as a Social Glue

• AI plays a meaningful role in mediating nuanced social dynamics
• Raise a new question: “How AI could be designed more intentionally to facilitate social collaboration?”. For example, AI could generate the solution that is “middle-ground” between collaborators

Trade-off: “Easing” vs “deepening” collaboration

1. AI could limit the depth of collaboration despite it enhancing the ease of collaboration
2. Participant role naturally shifted towards jointly evaluating the AI output, acting more as co-producer and less as co-composer
3. For experts, AI introduces some additional decision-making overhead, and the output was difficult to control

Future work that could be explored from this implication:

• How the depth of collaboration and interpersonal engagement can be preserved, or enriched, when AI is present
• The future design of deep generative AI may consider explicitly allowing collaborators to control how deep the collaboration is.

Human-ness of Deep Generative AI

• AI can act as a psychological safety net and mitigate mutual trust

The question that arises from this implication:

• How to define an objective function for deep generative AI?
• What defines the “humanness” of AI music generation?
• Which aspect impedes and enriches the quality of collaboration

Scope of Collaboration as a Function of AI training data

• The dataset in training data is also important because the dataset implicitly limits the creative scope of composition
• The collaborator skillset may not match the training set

6. Limitation and Future Work

Limitation

• The task is relatively a low-stakes task
• The participant may be more or less forthcoming depending on how high-stakes the task is, and how familiar they are with each other when conflict arises

Future work

• Aside from the future work defined in the previous implication,
• Examining the role of AI across a greater variety of human-human collaboration contexts (e.g: different tasks, social relationships) can be explored in the future

7. Conclusion

• AI can play meaningful roles in mediating social dynamics during collaborative composition
• AI has the potential to be a social glue during the co-creation process

Contribution

1. An understanding of the various roles of AI in mediating human-human co-creation
2. Description of how collaborators develop strategies together when co-creating with AI
3. Implications and recommendations for what to consider and how to utilize AI in supporting co-creative process

Reference

[1] Yuyu Lin, Jiahao Guo, Yang Chen, Cheng Yao, and Fangtian Ying. 2020. It Is Your Turn: Collaborative Ideation With a Co-Creative Robot through Sketch. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 1–14.  https://doi.org/10.1145/3313831.3376258

[2] Matthew Guzdial, Nicholas Liao, Jonathan Chen, Shao-Yu Chen, Shukan Shah, Vishwa Shah, Joshua Reno, Gillian Smith, and Mark O Riedl. 2019. Friend, collaborator, student, manager: How design of an ai-driven game level editor affects creators. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. 1–13. https://doi.org/10.1145/3290605.3300854

[3] Ryan Louie, Andy Coenen, Cheng Zhi Huang, Michael Terry, and Carrie J. Cai. 2020. Novice-AI Music Co-Creation via AI-Steering Tools for Deep Generative Models. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’20). Association for Computing Machinery, New York, NY, USA, 1–13. https://doi.org/10.1145/3313831.3376739 

[4] https://www.nngroup.com/articles/thinking-aloud-the-1-usability-tool/

P.S :

I (Ananda Phan) encourage the readers of this article to read the full paper in order to fully understand its content. If there are any mistakes or misinterpretations in the paper, please kindly leave your comments in the comment section below. If there are any discrepancies between this writing and the original paper, the interpretation provided in the original paper should take precedence.