The Lack of Other Minds as the Lack of Coherence in Human–AI Interactions

Abstract

As artificial intelligence (AI) undergoes rapid evolutionary advancements, two enduring queries in the philosophy of language and linguistics persist: the problem of other minds and coherence. This can be further explored by the following question: is there a fundamental difference between human-AI interactions and human–human interactions? More precisely, does an AI partner’s ability to understand discursive coherence sufficiently approximate that of the human mind? This study frames the problem of other minds as a problem in discourse analysis, positing that linguistic exchange inherently constitutes interactions between minds, where the act of decoding discursive coherence serves as a proxy for apprehending other minds. Guided by this perspective, this study uses four criteria of discursive coherence to examine how AI partners (with a focus on ChatGPT) achieve discursive coherence, thus reflecting whether an AI partner’s ability to understand discursive coherence suffices to simulate the human mind. Through a comparison between human–human interactions and human-AI interactions, the results indicate that while ChatGPT demonstrates proficiency in constructing discursive coherence along dictional, intentional, emotional, and rational coherence lines, the structural complexity and generative creativity of its coherence lines remain significantly below the threshold observed in human–human interactions. Moreover, ChatGPT’s emotional expressiveness pales in comparison to the rich, nuanced affect inherent in human–human interactions.

1. Introduction

Artificial Intelligence (AI), characterized as a non-human form of intelligence possessing sufficient flexibility and autonomy to interpret data, glean knowledge from it, and thereby accomplish specific objectives [1], has not only caught up with but in many cases, exceeded human capabilities across a wide range of fields [2,3,4] With the significant progress of AI, human-AI interaction (HAI) has reached notable milestones, especially ChatGPT, which has expanded its repertoire to include not only engaging in conversations and handling complex Q&A sessions but also translating languages and producing various sophisticated texts, such as academic papers and literary works [5,6]. Moreover, it has demonstrated capabilities in logical reasoning, semantic analysis, and comprehension, participating in humor, casual conversation, and even flirtatious conversations [7,8,9]. These advancements underscore the evolving nature of HAI, which is increasingly approximating natural conversational exchanges.

Ongoing innovation in AI has significantly advanced the naturalness and fluency of interactions between AI and humans. In terms of communication modalities, HAI bears a striking resemblance to human–human interactions (HHI), with boundaries far more blurred than previously imagined [10]. This remarkable convergence raises profound questions: is there a fundamental difference between HAI and HHI? More precisely, does AI partners’ ability to understand discursive coherence sufficiently approximate that of the human mind? These inquiries intersect with two classic problems in the philosophy of language and linguistics: the problem of other minds and coherence. This paper argues that the problem of other minds is not merely an epistemological or conceptual issue but fundamentally a question of discourse analysis. HHI inherently constitutes an interaction between minds, where understanding utterances involves apprehending others’ mental states. Crucially, the process of discourse comprehension is coterminous with the construction of discursive coherence, such that the dynamics of coherence-building reveal the mechanisms of other-minds cognition. Guided by this framework, this study employs four established metrics of discursive coherence as evaluative criteria, conducting controlled comparisons between HAI and HHI to assess the construction of discursive coherence, thus reflecting whether AI partners’ (with a focus on ChatGPT 3.5) ability to understand discursive coherence sufficiently approximates that of the human mind.

The subsequent section commences with an examination of the existing literature concerning the philosophical conundrum of other minds, emphasizing its intrinsic qualities. Section 3 explores discursive coherence lines to test whether HAI can build discursive coherence at different levels. Section 4 conducts heterogeneous conversations to investigate the coherence construction in both HAI and HHI.

2. The Intrinsic Qualities of Other Minds

With ongoing AI technological innovation, ChatGPT-like intelligent partners increasingly engage in natural conversations with humans, raising questions about whether their discourse processing abilities match human mental capacities. This inquiry relates to the philosophy of language’s problem of other minds [11,12], a focal point in Western philosophy of mind and language since the mid-20th century [13]. Standard approaches to this question include the argument from analogy, the best explanation, and the criteria approach1 [14]. While these methods address the problem of other minds from distinct theoretical perspectives, each faces significant philosophical controversies, implying the complexity of understanding other minds. The problem of other minds is no longer merely an epistemological or conceptual issue but also a matter of discourse analysis.

2.1. As an Epistemological and Conceptual Problem

During a period roughly centered on the mid-20th century, philosophers extensively debated the problem of other minds, traditionally framed as an epistemological puzzle: How can I know (or justify the belief) that there exist other beings with mental or conscious states [14]? As an epistemological concern, the problem of other minds centers on the question of how we ascertain or confirm the presence of mental qualities such as thoughts, emotions, and beliefs in entities other than ourselves [13].

The oldest systematic solution to the problem of other minds is the argument from analogy, rooted in John Stuart Mill’s theory of analogical reasoning. Analogical reasoning cites similarities between two things and uses this as support for concluding further similarities may exist. Therefore, when responding to the question what evidence or grounds justify my belief that the walking and speaking beings I encounter have sensations and thoughts (i.e., possess minds), Mill posits two key premises: “First, they have bodies like me, which I know in my own case, to be the antecedent condition of feelings; and because, secondly, they exhibit the acts, and outward signs, which in my own case I know by experience to be caused by feelings” [15] (p. 243). While this argument was once popular [16], it entails an inherent epistemological asymmetry: while I have direct, privileged access to my own mental states, I can never directly observe or check another’s mind or verify if their internal states match mine [17]. This gap undermines the analogy’s logical force, as inductive inference from a single case (my own mind) to universal conclusions (all other minds) is inherently flawed in empirical reasoning [18].

Facing the shortcomings of the argument from analogy, scholars proposed the argument from best explanation (IBE) to advance on the argument from analogy. The core claim of IBE is that the hypothesis that others possess minds similar to mine offers the best explanation for their behavior [13]. Within this framework, mental states are treated as internal states of individuals, providing the most plausible account for observed behaviors—while alternative explanations (e.g., “others are god-controlled puppets”, “I alone act with mental states”, or “others have fundamentally alien minds”) are dismissed as implausible due to their lack of explanatory power [19]. Pargetter argues that IBE circumvents traditional criticisms of the analogical argument by relying solely on the explanatory efficacy of the hypothesis itself (i.e., “others have minds”) rather than inductive generalization from self-experience. However, Melnyk [20] highlights a critical distinction between scientific realists’ belief in “theoretical entities” and ordinary people’s belief in “other minds”: behavioral evidence alone cannot establish the existence of others’ mental states without supplementary knowledge from self-experience. This critique challenges IBE’s claim to standalone explanatory sufficiency, suggesting that it implicitly relies on the same solipsistic starting point as the analogical argument—a flaw that undermines its alleged superiority.

Solutions like the argument from analogy and IBE essentially frame “knowledge of other minds” as analogous to scientific explanation. Wittgenstein challenged this approach, declaring: “Misleading parallel: psychology treats of processes in the psychical sphere, as does physics in the physical” [21] (§571). Wittgenstein distinguished between criteria (grammatical rules governing word use) and symptoms (empirical hypotheses), a distinction central to his criterial approach [22]. For Wittgenstein, the relationship between mind and behavior is not deductive or causal but conceptual: behavior serves as a criterion for mental states, embedded in the “grammar” of language use [22]. The criterial approach posits that the link between mental states and behavior is constitutive rather than inferential. For example, itching and scratching are not connected via inductive reasoning but belong to the same conceptual system—our very concept of “itch” includes scratching as its criterion. This framework raises two key challenges: first, if no inferential process is involved, how do we acquire concepts of others’ experiences; second, without causal or deductive links, can conceptual connections alone bridge the divide between observed behavior and unobserved mental states [14]? Despite these challenges, the criterial approach offers a pivotal shift: by reframing the problem of other minds as a conceptual rather than epistemological issue.

By anchoring the problem of other minds in conceptual connections, the criterial approach directly links it to language. For philosophers of language, concepts are embedded in language; thus, dissolving the problem as a conceptual issue requires analysis of word use—a premise underlying their claim that most philosophical puzzles originate in linguistic misunderstandings. Applying this to other minds, its enquiries become inseparable from linguistic analysis, opening a new perspective: the problem of other minds as a problem of discourse analysis.

2.2. As a Problem in Discourse Analysis

The claim that the problem of other minds is a problem in discourse analysis not is speculative but rooted in the tradition of philosophy of language, which emphasizes how language analysis dissolves philosophical puzzles. As Price argued, “One’s evidence for the existence of other minds is derived primarily from the understanding of language [23] (p. 429)”—a claim that resonates with Liu’s assertion in his book the Summary of Arts that the study of speech is also the study of the mind [24]. The interaction between minds is enacted through linguistic exchange, meaning that solving the problem of other minds requires examining discourse processes and their context, thereby revealing mental interactions through analysis of communicative dynamics.

The claim that the problem of other minds belongs to a problem in discourse analysis is implicit in the work of philosophers such as Ryle, Austin, Strawson, and Wittgenstein, who systematically relocated the debate to the domain of linguistic analysis. Ryle [25] categorized mental concepts (e.g., “know”, “belief”) as logical categories of intensions, motives, and capacities, arguing they could be reduced to descriptions of observable behaviors. By treating “mind” as a shorthand for patterns of action rather than a hidden entity, he dissolved the problem of other minds into questions of behavioral consistency. Austin [26] shifted the focus from “How do we know other minds?” to an analysis of performative language uses. For Austin, understanding “other minds” requires decoding the illocutionary forces (e.g., asserting, questioning, commanding) underlying utterances, not speculating about internal states. Strawson critiqued Cartesian dualism by centering the concept of “persons” as irreducible wholes possessing both physical and mental attributes. For Strawson [27], “other minds” are not inferred but presupposed in our use of personal pronouns and in the practice of attributing both actions and experiences to others. Wittgenstein’s definitive critique of private mental entities argued that language gains meaning through public use, making “private languages” (accessible only to one mind) logically impossible. For him, mental states like “pain” are tied to public criteria (e.g., crying, wincing) embedded in linguistic practices [21]. While his argument does not directly prove other minds exist, it dismantles the premise that minds are inherently unknowable: if mental concepts are defined by shared behavioral criteria, then understanding others’ minds is a matter of mastering language games, not solving an epistemological puzzle. Collectively, Ryle, Austin, Strawson, and Wittgenstein showed that the problem of other minds arises from misusing language—treating mental terms as referring to private objects rather than social coordination tools. By analyzing how we attribute beliefs, desires, and emotions in daily discourse, they revealed that “other minds” are not philosophical enigmas but products of linguistic practice. This discursive turn offers a radical solution: knowledge of other minds is not a matter of evidence or inference but of participating in shared forms of life where language itself bridges the gap between “self” and “other”.

Understanding the problem of other minds from a discourse analysis perspective does not fixate on verifying the existence of other minds but focuses on how communication is constructed through discursive interaction. Discourse communication inherently involves an exchange between minds. For instance, functional linguist Givón [28] posits that the context enabling discursive and interpersonal interaction is a psychological representation of others’ mental states, terming this “context as other minds”. In his framework, contextual analysis does not begin with the external physical environment surrounding discourse but with the mental perspectives of interactants. Context is dynamic: it emerges through associative judgment as a conceptual “framing”—a concrete mental operation and a psychological model encompassing participants’ current states, rapidly shifting beliefs, and intentional orientations. In discourse, speakers consistently tailor their utterances to listeners’ mental models. Discursive interaction thus constitutes the construction of an other-minds context, an online modeling of interlocutors’ belief systems and intentional states. Within this contextual framework, mental representations of interlocutors’ epistemic and deontic states are rapidly constructed through grammatical encoding during communication. Discourse itself becomes a bilateral process of online mental modeling, where interactants mutually serve as “other minds”. Givón’s context as other minds only opens a pathway for the discourse-analytic approach to recognize the existence of other minds at the surface level of language but also provides a paradigmatic example for cognizing this issue through discourse analysis.

The discussions on the problem of other minds by Ryle and others, together with Givón’s elaboration on context as other minds, have linked this philosophical issue to discourse analysis. As a problem of discourse analysis, its core lies in how both communicative parties on-line construct an other-minds context during discourse exchange and judge whether the other party can generate coherent discourse. This is because understanding context as other minds necessarily involves interpreting discourse meaning, which in turn requires examining how both parties achieve coherence. Objectively, discursive interaction is the coherent construction of an other-minds context and a process of cohesive mental interaction. In this construction, “my mind” models “other minds”, while “other minds” simultaneously models “my mind”, ultimately forming a unified system of discursive interaction. Therefore, the problem of other minds is a problem of constructing cohesive interactions.

3. The Structure of Coherence Lines in Discourse Understanding

In discourse analysis, the issue of other minds extends beyond mere verification of their existence, concentrating instead on the mutual assessment of discursive coherence between interlocutors. Building upon Jakobson’s foundational functions of linguistic communication and integrating Givón’s perspectives, Du [29,30] has delineated coherence lines and coherence structure within discourse communication. It is important to note that the concept of “coherence” is a philosophical notion rather than cohesive devices in linguistics. It mirrors the thinking activity of discourse participants, which is manifested as the pairwise connection between the speaker’s and the hearer’s utterances [29] (p. 2).

3.1. Four Coherence Lines in Communication

Discourse communication entails the interplay between the components of a speech event and their respective communicative functions. Jakobson [31] posited that a speech event is typically composed of six core elements2 and a discourse encompass the representation of objects or ideas, the articulation of emotions, the embellishment of speech, and the conveyance of implicit meanings. Expanding upon Jakobson’s framework of discourse exchange, Du [30] proposes a discourse analysis that recognizes four distinct types of coherence lines including the dictional line, intention line, rational line and emotional line.

The dictional line refers to the exploitation and utilization of linguistic forms themselves as discursive resources that emerge and become salient in communication [29] (pp. 201∓202). While dictional highlights the linguistic forms as the linking device, its scope is broader than cohesive device, which encompassing all linguistic means designed to connect discursive segments pairwise, not only present simultaneous discursive segments but also absent, indirectly simultaneous ones. Thus, the dictional line can be equated to “cohesive devices combined with the semiotic properties of discursive segments”, embracing categories such as “reference”, “structural cohesion”, “logical connection”, “lexical means”, and “phonological association” [29] (p. 201). Consider the example 1:

(1)

A: “Ni bu hui tiao wu ma? (Can you dance?)”3

B: “Wo ke yi tiao liu. (I can dance six.)”

Example 1 illustrates the exploitation of linguistic forms, specifically phonological associations. In Example 1, the homophonic equivalence between “(wǔ/dance)” and “(wǔ/five)” triggers a conceptual shift from an action “wu (dance)” to number “six”.

The intentional line relates to the linguistic and philosophical concept “meaning”, referring to the abstract coherence beyond linguistic forms—a semantic continuity shared by interlocutors, triggered by discursive segments [29] (p. 216). It concerns reference in communication, which entails the utilization of contextual resources by both parties, primarily manifested through the referents of words, the conceptual attributes they denote, and the circumstantial contexts they pertain to. For example,

(2)

A: “Why do humans have two ears, two eyes, but only one mouth?”

B: “Because in life, one needs to listen more, see more, but speak less.”

In Example 2, the coherence of this conversation hinges on conceptual associations between “ears/eyes/mouth” and “listening/seeing/speaking”, as well as between “two/one” and “more/less”.

The emotional line focuses on the emotional functions of language, involving the continuity of human “feelings” and “emotions [29] (p. 207)”. The construction of affective momentum considers linguistic devices such as word choice, rhetoric, and selective expression, with modal particles, interjections, and evaluative adjectives serving as typical carriers. Although affective momentum is always manifested through discursive segments, when emotions intensify to a certain degree, the utterances themselves may lose literal meaning, becoming mere vehicles for emotional catharsis. Consider example 3:

(3)

Liu Bei: “To preserve that suckling I very nearly lost a great commander!”

Zhao Yun: “Were I ground to powder, I could not prove my gratitude.”

This conversation is chosen from the chapter 42 of Romance of the Three Kingdoms. This conversation happens after Zhao Yun risked his life to rescue Liu Bei’s infant son from chaos. Here, the primary function of communication is not literal meaning but emotional expression: Liu Bei’s utterance serves to express gratitude of Zhao Yun’s rescue, while Zhao Yun’s reply reinforces this affective reciprocity. Their interaction achieves coherence at the level of emotional salience, where shared feelings transcend linguistic meanings.

The rational line is dedicated to the communicative content, aligning with normative standards, and implies that discourse must conform to specific principles or logical sequences [29] (p. 231). In communication, both the discursive segments chosen by speakers and the emotions or meanings they convey are regulated by norms of rationality and plausibility. Discourses unconstrained by reason (e.g., the speech of individuals experiencing psychosis) cannot facilitate normal communication. When discourse is governed by rationality, the coherence constructed through dictional, intentional and emotional necessarily embodies plausibility. Rational momentum emerges from interactive communication, embodying both rational principles and pragmatic reasonableness. Consider:

(4)

Teacher: “Hey, there’s a job opening! They need a student with IELTS 6.5 and strong computer skills. Are any of you interested?”

Student A: “Is it preferring guys or girls?”

Teacher: “Males are preferred.”

Student A: “Alas! Why is that?”

Student B: “What is the salary?”

This conversation concerns recruitment information. Student B’s response, “What is the salary?” directly addresses practical or factual logic, as asking compensation is inherently reasonable in job application contexts, reflecting how rational line ensures discourse coherence.

During an actual interaction, the verbal contribution made by the producer and the receiver may fall into four lines of communication: choosing diction, conveying meaning, expressing emotions and give reasons. Correspondingly, there are four types of coherence lines: dictional line, intentional line, emotional line, and rational line. Through the interplay of these lines, a discourse is rendered coherent and fluent, facilitating a smooth exchange of ideas and sentiments.

3.2. The Structure of Discursive Coherence

Coherent discourse is essential for the effective fulfillment of communicative functions. The question arises: how can coherence in discourse communication be established and evidenced? Wittgenstein emphasized that internal processes necessitate external criteria for their recognition [21]. Echoing this perspective, Brandom contends that within particular discursive social practices, there is a requirement to externalize implicit content, rendering it explicit [32]. To scrutinize the contextual frameworks that underpin discourse coherence, one must analyze the techniques of discourse construction within defined contexts and investigate the roles of the four fundamental coherence lines.

Understanding coherence requires an understanding of the participants’ understanding in a verbal interaction, and when the producer and the receiver endeavor to achieve a mutual understanding in an actual verbal interaction, they aim at the maximal common intelligibility along the salient line of verbal exchange [29]. The whole process is characterized by a dynamic interplay between the speaker and the listener. The speaker initiates discourse propelled by emotional stimuli, and the conversations that emerge may exhibit a variety of coherence lines, including dictional, intentional, rational, and emotional lines. The hearer, in turn, can engage in the communicative process by responding to these same dimensions. From a theoretical standpoint, both interlocutors in the exchange are equipped with these four aspects, which give rise to four distinct types of coherence structures, each underpinned by a corresponding coherence line: dictional coherence, intentional coherence, rational coherence, and emotional coherence.

Coherence in communication can be attained not only through a single type of coherence line but also through an interweaving of multiple types. These multiple layers are defined by the simultaneous engagement of two or more coherence lines within a single conversational turn. When considering the vantage point of the participants in the interaction, each coherence layer necessitates the interconnection and harmonization of their respective coherence lines. This dynamic can lead to coherence that is established through both homogeneous and heterogeneous types of lines [30] (p. 211). Consequently, the structure of coherence lines can be systematically classified into two categories: mixed and non-mixed structures of discourse coherence.

A mixed structure of discursive coherence emerges when there is a divergence in emphasis between the addressor and the addressee, with the addressor highlighting one particular coherence line and the addressee focusing on a different one. Conversely, a non-mixed structure arises when both interlocutors achieve consensus and alignment along the same coherence line. The mixed structure relies on the intricate interplay of multiple coherence lines, which together create a complex tapestry of coherence. This complexity can be observed in two forms: the concurrent presence of multiple coherence lines of the same type or the intertwining of different types of coherence lines.

During communication, participants can establish four types of simple coherence by operating within the same coherence lines (as depicted in Figure 1), as well as achieve complex coherence by spanning across various layers of coherence lines. These four types of coherence are achieved through dictional–dictional lines, intentional–intentional lines, rational–rational lines, and emotional–emotional lines [30] (p. 211).

Additionally, a speaker might concentrate on a single coherence line, while the listener interacts across multiple lines, leading to a scenario known as “single-to-multiple coherence”. On the flip side, a speaker might employ multiple coherence lines but effectively engage with only one line in the listener (as illustrated in Figure 2).

In the context of natural communication, the discourse process involves the creation of multiple-layered coherence between conversing individuals (as shown in Figure 3).

The foregoing discussion has established that there are four fundamental coherence lines within the realm of discourse communication. Utilizing these foundational lines, simple coherence can be realized through the operation of identical coherence lines, while complex coherence is accomplished by integrating various coherence lines. Consequently, the spectrum of potential discursive coherence structures can be systematically represented in

Table 1. The structure of discourse coherence4.

Coherence	The Coherence Line of Addressee
	Types of lines	Dictional line	Intentional line	Rational line	Emotional line
	Dictional line	dictional–dictional	dictional–intentional	dictional–rational	dictional–emotional
	Intentional line	intentional–dictional	intentional–intentional	intentional–rational	intentional–emotional
	Rational line	rational–dictional	rational–intentional	rational–rational	rational–emotional
The coherence line of addressor	Emotional line	emotional–dictional	emotional–intentional	emotional–rational	emotional–emotional

.

Table 1 provides 16 potential structures that constitute discourse coherence. From the perspective of the discourse analysis of other minds, the existence of other minds can be transformed into whether the discourse can achieve coherence and at which levels such coherence can be attained. This is because the basis upon which discourse interaction proceeds lies in the construction of different levels of coherence lines. The possibility of discursive coherence is, in fact, the possibility of understanding: if one can conduct a coherent interpretation of the discourse from multiple levels, then one can understand the discourse and gain knowledge of other minds.

The process of discourse communication involves the dynamic creation of coherence and real-time management of diverse semantic relationships. Although it may not be possible to categorically state that there is an inherent distinction between HHI and HAI, a comparative analysis of their respective coherence constructions within conversations can elucidate the commonalities and disparities in their interactive dynamics.

4. Constructions of Discursive Coherence in HHI and HAI

As artificial intelligence accelerates its development, the understanding of communication has emerged as a fundamental topic in natural language processing (NLP) and a key area of research in HAI. This section is dedicated to investigating whether AI entities such as ChatGPT exhibit cognitive characteristics analogous to the human mind and to assessing whether there exists a fundamental divergence between HAI and HHI. To fulfill this objective, the analysis will focus on the manner in which ChatGPT establishes discursive coherence along the four coherence lines during its interactions with humans. This will be accomplished by designing and comparing homogeneous conversations to reveal both the congruencies and disparities in their respective interaction patterns. This first part of this section tests whether ChatGPT can achieve these four types of coherence lines in conversation and the second part compares the realization of coherence lines in HAI and HHI.

4.1. The Construction of Discursive Coherence in HAI

A discourse is characterized by the interactive nature of communication, with its essence rooted in the multi-level construction of coherence within the framework of coherence lines. The establishment of discursive coherence is predicated on the interplay of four fundamental lines: dictional, intentional, rational, and emotional lines. These elements serve as the foundational pillars that shape the discourse’s coherence. The question arises: can an AI partner like ChatGPT can effectively construct discursive coherence with these four coherence lines?

4.1.1. Highlighting the Dictional Line and Intentional Line

Dictional coherence arises from an emphasis on dictional lines, primarily achieved through linguistic connections that establish direct or indirect relationships with specific words from the addressor. This form of coherence is rooted in the structural aspects of language, creating a network of lexical ties that bind the discourse together. Intentional coherence, by contrast, is constructed by bringing the intentional line of the discourse. Distinct from the symbolic properties of linguistic expressions, this line represents the deeper purpose or intention behind the discourse. The intentional line extends beyond the surface structure of language, delving into the semantic content and the inferential relationships that contribute to the coherence of meaning. This meaning is mutually recognized and understood by all participants in the discourse, forming a shared basis for communication.

When ChatGPT interacts with humans, some discourses may highlight dictional coherence, while others may highlight intentional coherence. For example,

(5)

You: Do you like Shanghai?5

• ChatGPT: I really like Shanghai! Shanghai is a vibrant and charming city with a long history and unique culture. From its modern cityscape of skyscrapers to its old neighborhoods with historical traditions, Shanghai blends modernity with traditional charm, and it offers a wide variety of food, shopping and cultural activities.
• You: Do you like Shanghai?
• ChatGPT: I truly like Shanghai! There are many fascinating things about this city, such as its unique architectural style, rich cultural landscape and delicious food.
• You: Do you like Shanghai?
• ChatGPT: I can’t say like or dislike, because I don’t have emotions, but I can say that Shanghai is a very interesting and vibrant city with a unique culture and style.

The conversation presented in Example 5 shows that ChatGPT demonstrates the ability to establish discursive coherence at the dictional and intentional levels. This conversation revolves around preferences for Shanghai. In response to each query, ChatGPT consistently maintains grammatical fluency and addresses the topic with reasoned explanations. In this brief exchange, its interaction with humans highlights primarily the dictional and intentional coherence. The structural organization of the discursive coherence in this conversation is visually represented in Figure 4, which provides a schematic overview of the interaction’s flow and coherence-building elements.

As shown in Figure 4, this conversation initiates with the construction of dictional lines and then follows the intentional lines. The dictional line is achieved through linguistic repetition: phrases like “Do you like Shanghai?” and responses such as “I really like Shanghai!” “I truly like Shanghai!” and “I can’t say ‘like’ or ‘dislike’” establish formal continuity through word recurrence. The intentional line is evident in its sustained focus on the topic, with justifications anchored in contextual resources about Shanghai—such as describing it as “a vibrant and charming city with…” and noting “…its unique architectural style, rich cultural landscape, and delicious food.” Moreover, deictic elements (“you”, “I”, “Shanghai”, “this city”) and referential consistency serve as contextual anchors, as semantic coherence emerges from the strategic utilization of shared referents and situational knowledge. The establishment of intentional coherence involves the strategic use of these contextual elements and referents to weave a coherent narrative.

4.1.2. Highlighting the Rational Line and Emotional Line

Emotional coherence is established when participants in a discourse prioritize emotional lines, forging a connection through the exchange of shared emotions and sentiments. While discourse itself is not inherently emotional, it serves as a vehicle for the expression of emotional states. Interactions that emphasize emotional coherence involve the activation and mutual influence of emotional and affective experiences. Conversely, rational coherence emerges from focusing on logical consistency within the discourse. For meaningful interactions, both parties must select rationally sound discourse segments, aligning emotions and meanings with logical standards. Rational coherence hinges on participants’ mutual understanding and recognition of logical principles.

In discourse communication, ChatGPT show capacity to construct coherence along both rational and emotional lines. However, its approach to emotional coherence may sometimes appear less natural or nuanced, as exemplified in example 6,

(6)

You: You son of a bitch.

• ChatGPT: I think you may be trying to express dissatisfaction or impatience, but I hope we can be friendly. Is there anything I can do for you?
• You: You’re an idiot.
• ChatGPT: Haha, you’re so funny! Any questions or do you need help?
• You: I’m scolding you, but you’re laughing!
• ChatGPT: I understand that you may have some emotions and I don’t mind. I just want to make sure that our communication is positive, and is there anything I can do to help you?

In Example 6, discursive coherence is primarily achieved by emphasizing emotional and rational coherence lines. During this conversation, when confronted with human verbal abuse, ChatGPT refrains from reciprocal invective. Instead, it endeavors to understand the speaker’s motivations for such remarks, seeks to identify the rationale behind the discourse, and employs explicit emotional vocabulary to emotionally soothe the speaker as much as possible. The structure of discursive coherence is illustrated in Figure 5.

In Figure 5, the speaker initially underscores the emotional line, acknowledging that words alone do not convey emotions; it is people who utilize words to express them. During the first exchange, the speaker vents negative emotions by using the derogatory term “son of a bitch”, thereby highlighting the emotional aspect. In response to the speaker’s outburst, ChatGPT refrains from retaliation and instead responds rationally: it deduces that the speaker might be expressing dissatisfaction or impatience, indicating potential issues. Consequently, ChatGPT inquires if it can assist the speaker, adopting a tone of concern. Following the first exchange, the speaker continues to vent emotions with “you are an idiot”, maintaining emotional continuity. ChatGPT attempts to establish emotional coherence by humorously complimenting the speaker’s sense of humor, responding to negative outbursts with positive affect, offering comforting words, and asking, “Any questions or do you need help?” The third exchange mirrors the first: when the speaker expresses dissatisfaction with “I’m scolding you, but you’re laughing”, ChatGPT avoids emotional engagement and instead seeks to understand the reason for the insult, providing reassurance. This approach aligns with core requirements for intelligent chatbots’ responses: recognizing others’ emotions and attuning to their emotional state. For instance, if a chatbot detects distress, it should respond with comforting or uplifting words [33].

Based on the analysis of two HAI examples, it is evident that ChatGPT can construct discursive coherence across multiple layers—dictional, intentional, emotional, and rational lines. However, does this ability imply that ChatGPT approximates the human mind and that HAI is indistinguishable from HHI? Although ChatGPT appears to demonstrate the capacity to establish discursive coherence through these four coherence lines, enhancing the smoothness of HAI, we can assert that the coherence and creativity of HAI have not yet reached the natural level of HHI—particularly in emotional coherence. As observed in Example 6, ChatGPT’s emotional construction relies on system-predefined words or phrases, lacking the flexibility inherent in human interaction. This preliminary observation, of course, requires further empirical verification.

4.2. The Comparation Between Discursive Coherence in HHI and HAI

As an interactive chat program, ChatGPT exhibits a resemblance to humans in language use and understanding. It is capable of generating comprehensible text and providing contextually appropriate responses, achieving continuous, stable, and consistent multi-turn conversations that give users the impression of chatting with a friend. However, are interactions with ChatGPT truly equivalent to interactions with humans? Do no discernible differences exist in the construction of discursive coherence between HAI and HHI? This section analyzes heterogeneous conversations to investigate coherence construction in both HAI and HHI.

In actual communication, both parties can construct diverse coherent structures based on the four fundamental coherence lines, enabling flexible and varied interactions. When comparing the interaction between ChatGPT and a human on the same topic, the differences are presented in

Table 2. HHI and HAI with the same topic “Shanghai”.

HAI

HHI

You:    Do you like Shanghai? ChatGPT: I really like Shanghai! Shanghai is a vibrant and charming city with a long history and unique culture. From its modern cityscape of skyscrapers to its old neighborhoods with historical traditions, Shanghai blends modernity with traditional charm, and it offers a wide variety of food, shopping and cultural activities. You:    Do you like Shanghai? ChatGPT: I indeed like Shanghai! There are many fascinating things about this city, such as its unique architectural style, rich cultural landscape and delicious food. You:    Do you like Shanghai? ChatGPT: I can’t say like or dislike, because I don’t have the emotions, but I can say that Shanghai is a very interesting and vibrant city with a unique culture and style.

Human A: Do you like Shanghai? Human B: Yes, I like it, what’s up? Human A: Do you like Shanghai? Human B: Didn’t I tell you? Yes, I’ve been there several times. Human A: Do you like Shanghai? Human B: ??Are you out of your mind??

.

In Table 2, two conversations are presented, both centered on the topic of liking the city Shanghai. In HAI, the human acts as the speaker, and ChatGPT serves as the listener. In HHI, Human A is the speaker and Human B is the listener. As analyzed in Section 4.1.1, ChatGPT initially constructs coherence at the diction level and then extends it to the intentional level, explaining reasons for liking or disliking Shanghai to achieve overall coherence.

Compared with HAI, HHI involves a greater variety of coherence lines. In the first turn of exchange, speaker (Human A) and listener (Human B) primarily achieve coherence at the intentional level, supplemented by dictional, intentional, and emotional lines. Their conversation revolves around “Shanghai”: the repeated use of “like” establishes dictional line, “it” referring to “Shanghai” demonstrates intentional coherence, and the subsequent “what’s up” serves as an emotional bridge, indicating B’s interest and curiosity. In the second exchange, discursive coherence between A and B is achieved by highlighting emotional, rational, and intentional lines alongside the dictional line. The phrase “Didn’t I tell you” reveals B’s impatience, followed by “I’ve been there several times” as a rational explanation for liking Shanghai. The word “there” referring to “Shanghai” implies intentional lines. In the third exchange, the conversation emphasizes emotional coherence, with B expressing confusion and dissatisfaction via “??Are you out of your mind??”. The coherence structures of the conversations in HAI and HHI are represented in Figure 6 and Figure 7, respectively.

From the comparison of coherence structures in HAI and HHI shown in Figure 6 and Figure 7, it can be observed that the three rounds of HAI repeatedly emphasize coherence through dictional and intentional lines, lacking emotional creativity and expressivity. In contrast, HHI involves a greater variety of discursive coherence lines and places more emphasis on emotional expression and venting. HHI appears more flexible and creative because humans can generate novel conversational threads. In HAI, ChatGPT selects keywords (or concepts) based on questions and responds using related attributes, resulting in minimal information variation across the three rounds. Furthermore, ChatGPT’s responses are rational and question-focused, with minimal emotional affect. Notably, ChatGPT explicitly states its liking for Shanghai in the first two rounds but contradicts itself in the third by claiming neutrality, creating logical inconsistency and rendering the entire conversation unnatural.

Although ChatGPT can engage in conversations and construct discursive coherence based on dictional, intentional, rational, and emotional lines (as shown in Figure 4 and Figure 5), forming fluent conversations, this does not necessarily imply that HAI and HHI are identical or AI partners’ ability to understand discursive coherence sufficiently approximates the human mind. While ChatGPT can analyze human emotions from words, tone, and context, and use emotional language to achieve emotional coherence, these capabilities are algorithmically generated—simulated emotions rather than authentic ones [34]. Significantly, HAI and HHI remain distinct: compared to HHI, HAI exhibits lower coherence and creativity, with simpler and fewer coherence lines (as shown in Figure 6 and Figure 7), This is because ChatGPT’s discourse connections primarily depend on dictional and intentional lines, while HHI naturally integrates emotional and rational coherence. Additionally, HAI lags behind HHI in conversational naturalness and emotional richness.

5. Conclusions

The advent of ChatGPT has heightened ambiguity about whether a fundamental difference exists between HAI and HHI. More precisely, does AI partners’ ability to understand discursive coherence sufficiently approximates that of the human mind? To tackle these issues, this study investigates and contrasts the coherence construction in HAI and HHI. From the perspective of framing other minds as a matter of discourse analysis, the problem of other minds can be translated into conversation coherence construction. Given language’s diverse functions, conversational coherence can be established via four coherence lines: dictional, intentional, emotional, and rational lines. After comparing HAI and HHI on the same topic, this study finds that ChatGPT can establish coherence across dictional, intentional, rational, and emotional lines. Nevertheless, compared with HHI, ChatGPT’s construction of discursive coherence through these coherence lines is far simpler, which primarily focus on dictional and intentional lines, indicating a lack of communicative creativity and complexity. Moreover, although ChatGPT can use emotional language to create emotional coherence, it lags behind humans in emotional richness and naturalness. Although the examples selected in this article may be relatively brief, they can still reflect some of the problems in current HAI and give some hints for future advancements. For example, future advancements in HAI should concentrate on strengthening coherence construction and promoting emotional-domain research to foster more natural interactions.