Mutual Theory of Mind for Human-AI Communication

2025-05-02 0 0 1.16MB 7 页 10玖币

侵权投诉

Qiaosi Wang

qswang@gatech.edu

Georgia Institute of Technology

Atlanta, GA, USA

Ashok K. Goel

ashok.goel@cc.gatech.edu

Georgia Institute of Technology

Atlanta, GA, USA

ABSTRACT

New developments are enabling AI systems to perceive, recognize,

and respond with social cues based on inferences made from hu-

mans’ explicit or implicit behavioral and verbal cues. These AI

systems, equipped with an equivalent of human’s Theory of Mind

(ToM) capability, are currently serving as matchmakers on dating

platforms, assisting student learning as teaching assistants, and

enhancing productivity as work partners. They mark a new era in

human-AI interaction (HAI) that diverges from traditional human-

computer interaction (HCI), where computers are commonly seen

as tools instead of social actors. Designing and understanding the

human perceptions and experiences in this emerging HAI era be-

comes an urgent and critical issue for AI systems to fulll human

needs and mitigate risks across social contexts. In this paper, we

posit the Mutual Theory of Mind (MToM) framework, inspired by

our capability of ToM in human-human communications, to guide

this new generation of HAI research by highlighting the iterative

and mutual shaping nature of human-AI communication. We dis-

cuss the motivation of the MToM framework and its three key

components that iteratively shape the human-AI communication in

three stages. We then describe two empirical studies inspired by the

MToM framework to demonstrate the power of MToM in guiding

the design and understanding of human-AI communication. Finally,

we discuss future research opportunities in human-AI interaction

through the lens of MToM.

CCS CONCEPTS

•Human-centered computing

→

HCI theory, concepts and

models;Empirical studies in HCI;Natural language interfaces;•

Computing methodologies →Articial intelligence.

KEYWORDS

theory of mind, human-AI interaction, social intelligence

ACM Reference Format:

Qiaosi Wang and Ashok K. Goel. 2024. Mutual Theory of Mind for Human-AI

Communication. In Proceedings of Workshop on Theory of Mind in Human-AI

Interaction at CHI 2024 (ToMinHAI at CHI 2024). ACM, New York, NY, USA,

7pages. https://doi.org/XXXXXXX.XXXXXXX

1 INTRODUCTION

With new technology advancements, AI systems are increasingly

serving dierent social roles across contexts. For example, AI sys-

tems are acting as matchmakers to provide matches for our busi-

ness or life partners, as personal assistants to manage our daily

routines, as learning assistants to facilitate student learning, and

more. These AI systems are often able to perceive, recognize, and

ToMinHAI at CHI 2024, May 12th, Honolulu, Hawaii

2024. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00

https://doi.org/XXXXXXX.XXXXXXX

react to human characteristics, needs, and perceptions embedded

in our behavioral and verbal cues. This presents a new interaction

paradigm in Human-AI Interaction (HAI) that diverges from the

traditional Human-Computer Interaction (HCI)— people are ex-

pecting AI systems to possess social intelligence for diverse social

functions, yet are often uncertain about the AI’s capabilities and

social roles during interactions. Oftentimes during HAI, people

build dierent perceptions or mental models of the AI based on the

AI outputs [

]; at the same time, AI systems are also

constantly building dierent interpretations of human characteris-

tics, needs, and goals based on human’s input [

]. In

this emerging HAI paradigm, interpretations of each other, from

both the human side and the AI side, are playing an increasingly

crucial part in shaping human-AI interactions, but how should the

HAI community study it in a systematic way to enhance human-AI

interactions?

In this paper, we propose viewing this emerging HAI paradigm

through the lens of Mutual Theory of Mind (MToM). We draw in-

spirations from the human-human communication process, largely

enabled by our basic cognitive and social ability of Theory of Mind

(ToM). ToM is our ability to make conjectures about ourselves and

others’ mental states (e.g., emotions, intentions) [

], and it is

the key to enabling many human social behaviors such as com-

munication repair and making shared plans or goals [

]. Having

the capability of ToM enables us to construe a mental model of

others’ minds, which includes their thoughts, preferences, goals,

needs, plans, etc. [

]. In typical human-human communications,

having a Mutual Theory of Mind (MToM), meaning all par-

ties involved in the interaction possess the ToM, enables us

to continuously rene our interpretations of each others’ minds

through behavioral and verbal feedback, helping us to maintain

constructive and coherent communications.

Drawing on the parallel between the MToM in human-human

communications and the emerging HAI paradigm where both hu-

mans and AIs can construct representations of each other during

communications, we propose the framework of Mutual Theory of

Mind to guide the next generation of research in human-AI commu-

nications. We argue that MToM as a framework provides a process

and content account of human-AI communication that emphasizes

the iterative mutual shaping of each party’s interpretations and

feedback through dierent stages of the communication process.

We will rst review relevant literature on ToM and human-AI com-

munication, then describe the MToM framework in details. We

then summarize two of our empirical studies inspired by the MToM

framework, focusing specically on the second-level ToM— the

idea of “I can think about what you think about my mind”— in

human-AI communications. Finally, we discuss potential research

opportunities in human-AI interaction through the lens of MToM.

arXiv:2210.03842v2 [cs.HC] 25 May 2024

ToMinHAI at CHI 2024, May 12th, Honolulu, Hawaii Wang and Goel

2 RELATED WORK

2.1

Theoretical Perspectives of Communication

Communication is commonly dened as “the process of transmit-

ting information and common understanding from one person to

another.” [

] Scholars across disciplines have oered dierent per-

spectives to study and enhance communication.

In communication studies, researchers have focused on the dif-

ferent components at play during the communication process. The

classic Shannon-Weaver model of communication [

] outlines sev-

eral key components during the communication process [

]: sender

who initiates the communication process by sending messages en-

coded using symbols, gestures, words, or sentences through a cho-

sen channel to the receiver. While the message is transmitting

through the channel, there could be noises that could distort the

message. After receiving the message from the sender, the receiver

will decode the message into meaningful information, depending on

how the receiver interprets the message. Finally, the receiver will

provide feedback as a response to the sender. These key components

determine the quality and eectiveness of the communication.

The Cognitive Science perspective of communication highlights

the critical role of ToM [

]. ToM enables us to make suppositions

of other’s minds through verbal and behavioral cues, acting as

the foundation of human-human communication [

]. From this

perspective, both interlocutors during communication can form

interpretations of what’s on the other interlocutor’s mind based

on the implicit and explicit communication cues. For example, we

can often infer the interlocutors’ goals, plans, or preferences based

on what they said, their facial expressions, or their bodily expres-

sions [

]. Based on that interpretation we formed about the

other’s mind, we will act accordingly to correct, explain, or per-

suade. This cycle of building an interpretation of other’s minds and

then act upon that interpretation continues iteratively throughout

the communication process. Inferring about each other’s minds

through behavioral cues, according to this perspective, is therefore

crucial to a smooth and successful communication.

Communication process can also be interpreted from the social

science perspective through impression management [

]. In his

seminal work, Goman describes social interaction as an informa-

tion game between individuals and their audience to maintain the

“veneer of consensus” to keep the conversation going and to avoid

awkwardness. During social interactions, the audience usually try

to gather as much information as they could about the individu-

als they interact with in order to elicit a desirable response from

the individual; whereas individuals put up performances through

two kinds of expressions—– expressions that are intentionally per-

formed to leave a certain impression (expression given) or expres-

sions that are unintentionally given o that could inuence the

audience’s impressions of them (expression given o)—– to manage

impressions [

]. Throughout interactions, each party conveys their

denition of the situation through communications: individuals by

expressions and audience by reactions to the individuals.

These three perspectives on communication emphasize dierent

aspects of the communication process: the communication study

perspective focuses on the encoding and decoding process of mes-

sages; the cognitive science perspective discusses how behavioral

cues can inform our interpretations of interlocutor’s minds; the

social science perspective describes how interpretations of others’

minds could predict our behaviors. Our Mutual Theory of Mind

framework attempts to bring these dierent emphasis together into

one coherent framework to understand the mutual shaping process

of interpretations and feedback during communication.

2.2 Theory of Mind in Human-AI

Communication

Over the years, many researchers have recognized the crucial role

of ToM in HAI. In human-robot teaming research, ToM has been

intentionally built in as part of the system architecture to help

robots monitor world state as well as the human state [

], to con-

struct simulation of hypothetical cognitive models of the human

partner to account for human behaviors that deviate from original

plans [

], and to help robots to build mental models about user

beliefs, plans and goals [

]. Robots built with ToM have demon-

strated positive outcomes in team operations [

] and are perceived

to be more natural and intelligent [29].

Other research in HCI and human-centered AI has also been

exploring along the realm of ToM, focusing mostly on enhancing

user’s mental models and understanding of the AI systems. Prior re-

search has explored people’s mental model of AI systems— people’s

mental model of AI agents could include global behavior, knowl-

edge distribution, and local behavior [

]. People’s perception of

AI systems is instrumental in guiding how they interact with AI

systems [

] and thus serves as a precursor to their expectation of

AI’s behavior. Some recent research has also begun to examine how

to automatically infer user’s mental model of AI. Prior research sug-

gests the potential of leveraging linguistic cues to indicate people’s

perception of AIs during human-AI interactions. Researchers have

been able to infer users’ emotions towards an AI agent [

] and

signs of conversation breakdowns [26]from communication cues.

Given that AI’s behavior and output could also inuence user’s

mental model of the AI, and therefore how the user decides to

interact with the AI, we want to highlight that the interpretation-

feedback loop is mutual during the human-AI communication

process— user’s mental model of the AI can be informed by the AI’s

output, yet AI’s interpretation of the user can also be informed by

the user’s output, which is determined by the user’s mental model

of the AI. We propose the Mutual Theory of Mind framework to cap-

ture this mutual shaping process of interpretation-feedback during

human-AI communication.

3 MUTUAL THEORY OF MIND FRAMEWORK

Drawing from theoretical and empirical work, we posit the MToM

framework to guide the understanding and design of communica-

tions between humans and AI systems that exhibit social behaviors

enabled by ToM-like capability. The MToM framework provides

both process and content account of human-AI communication

by highlighting three elements that mutually shape the human-AI

communication process in three stages.

3.1 Three Elements of the MToM Framework

In the MToM framework, three elements are critical for humans

and AI to reach mutual understanding during the communication

process: interpretation, feedback, and mutuality.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MutualTheoryofMindforHuman-AICommunicationQiaosiWangqswang@gatech.eduGeorgiaInstituteofTechnologyAtlanta,GA,USAAshokK.Goelashok.goel@cc.gatech.eduGeorgiaInstituteofTechnologyAtlanta,GA,USAABSTRACTNewdevelopmentsareenablingAIsystemstoperceive,recognize,andrespondwithsocialcuesbasedoninferencesmadefro...

展开>> 收起<<

Mutual Theory of Mind for Human-AI Communication.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Mutual Theory of Mind for Human-AI Communication

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: