Human–AI communication in initial encounters: How AI agency affects trust,liking,and chat quality evaluation

The present study

Adopting the goal/question/metric method (Wohlin et al., 2012), this study aims to achieve the following objectives:

AI agency in human–AI communication

The notion of agency finds its roots in philosophy, notably in the works of Aristotle and Hume. In human–computer interaction, human agency, sometimes called a sense of agency, refers to “the experience of controlling both one’s body and external environment” (Limerick et al., 2014: 1). In empirical studies, the concept has been confined to humans’ perceived sense of their own agency while interacting with machines or digital interfaces (Bennett et al., 2023). Some studies also touch upon the concept of attribution of agency in human–machine interaction. However, their focal point is still on how the attribution of collaborative partners (human vs machine) would affect humans’ perceived sense of agency of humans themselves (Obhi and Hall, 2011).

While previous studies on agency have been extensively centered on how humans perceive their own agency during interactions with machines, a notable shift has occurred due to the integration of communicative AI into people’s daily lives. Humans no longer merely regard AI as channels or mediators, but increasingly as social actors with which they can develop relationships (Spence, 2019). Therefore, how agentic AI behaves in human–AI communication is crucial in contributing to humans’ impressions of AI, and, by extension, their evaluations of AI from both relational and functional perspectives (Edwards et al., 2019; Westerman et al., 2020). Due to the lack of intentionality, which was widely acknowledged as the key criterion of defining agency (Dennett, 1987), so far it is still debatable to what extent AI chatbots are truly agentic. However, during human–AI communication, humans may perceive AI to have some levels of agency in their own behaviors such as the “thinking and acting capacities that they appear to have during H-M interaction” (Liu, 2021: 387).

Upon initial examination, the concept of AI agency appears to closely resemble the idea of machines’ proxy agency of humans, that the AI agents are programmed to follow rules designed by humans (Bandura, 2006), or machine apparent agency, that self-learning AI is capable of learning rules from data (Liu, 2021). Nevertheless, the extent of AI agency may not be solely defined by its ability to adhere to rules set by humans, or even to self-learn in ways that humans may not yet fully comprehend. In the context of human–AI communication, we define AI agency as AI’s exemplified freedom to act on its own. An agentic AI could be viewed as acting independently and freely on its own, while a less agentic AI would be perceived as pre-determined and lacking choices in its actions such as an automaton.

In the context of human–AI communication, AI agency can be manifested in multiple dimensions. It can be observed in its source agency, such as to what extent it is self-governed or controlled by humans (Liu, 2021; Sundar, 2020). In addition, as a conversational partner, AI agency can be expressed and realized through its language agency (Duranti, 2004), which pertains to the language and linguistic forms employed by AI that reflect its “socioculturally mediated capacity” to engage in intersubjective communication (Ahearn, 2001: 112). Furthermore, as an embodied agent, AI’s appearances can also indicate its embodied agency (Krämer et al., 2015; Lee et al., 2006). For the current study, we focus on AI’s source autonomy and its language subjectivity, as these aspects are more commonly seen in text-based chat interactions between humans and AI chatbots online.

AI’s autonomy

How do individuals perceive a communicator who appears to express thoughts and ideas on their own? How, if any, does this perception change when the communicator is an AI rather than a human? In interpersonal impression formation, the targets’ ability to regulate and control their own behaviors may play a key role in affecting how others form impressions of them (Endacott and Leonardi, 2022). Although impression formation usually gradually unfolds during interactions with the target, the degree to which the target is independent, capable of making its own divisions, and has control over its own behaviors would serve as important cues for the observers to form impressions. When humans interact with autonomous AI chatbots, the exemplified autonomy of AI may serve as cues to promote perceptions of AI’s social affordance (Fox and McEwan, 2017; Gambino et al., 2020). Therefore, humans may perceive autonomous AI chatbots to have more agency over their actions and behaviors.

Notably, the perception of AI agency may seem very similar to AI’s anthropomorphism, which focuses on AI’s projection of human-like qualities or behaviors. However, these two notions should be differentiated from each other. On one hand, an AI chatbot could behave very much in a human-like way, while still being evaluated as not agentic. On the other hand, a very agentic AI chatbot may project a strong sense of freedom in its action, while not resembling any human behaviors and responses. The perceptions of agency should be set apart from to what extent AI resembles humans, instead, the benchmark for comparison is the extent to which the AI is autonomous, independent, self-governing, and acting on its own rules.

The extended CASA framework argues that changes in media’s social affordance, people’s knowledge and experiences with media agents, and interactions between humans and media agents have changed drastically in the past three decades (Gambino et al., 2020; van der Goot and Etzrodt, 2023), humans have formed specific and unique perceptions and scripts of human–media interactions (Fox et al., 2015; Gambino et al., 2020). In this current study, the extent to which AI chatbots can operate autonomously, be guided by their own set of rules, and think in a way that may go beyond human imagination or expectations could serve as fundamental factors in shaping how humans perceive and assess their interactions with AI interlocutors.

AI’s subjective language

In human–AI communication, besides whom people communicate with (i.e., source autonomy), another crucial aspect through which AI agency manifests itself is how this communication is delivered. Linguistic patterns in everyday descriptions often contain pervasive and powerful cues at agency (Fausey et al., 2010). In text-based human–AI communication, language patterns used by AI communicators uncover the nuances of agency expression, attribution, and denial, particularly through varying degrees of subjectivity embedded within the language of the interlocutor (Ahearn, 2001).

Subjectivity refers to a person’s perspective or opinion, especially one’s feelings, beliefs, and desires (Honderich, 1995). Subjective language often indicates one’s personal opinion, in contrast to objective knowledge or fact-based beliefs. Subjectivity emphasizes that a person is not only responding to, or having a passive relationship with the world, but indicates their agency, an active engagement with the environment or material world. Research in human communication has shown that the use of first-person singular pronouns is always associated with self-focus (Ahearn, 2001; Chung and Pennebaker, 2007; Tausczik and Pennebaker, 2010). Moreover, literature in intercultural communication has also found that the use of first-person singular pronouns is associated with an independent self-construal (Chung and Pennebaker, 2007; Na and Choi, 2009). Consequently, AI chatbots that employ more subjective language may be perceived as more agentic compared to those that use objective language.

Therefore, language used by machines may serve as one type of primary cue which could promote human’s social perceptions and responses (Turkle, 1984). By using subjective language, AI may trigger humans to form social perceptions of AI and use social responses when interacting with AI (Molina and Sundar, 2022, 2024). The degree to which media can provide social affordances and personalization of messages or feedback would both promote humans to form perceptions and scripts specific to human-media agent interactions (Fox and McEwan, 2017; Gambino et al., 2020). When interacting with AI using subjective language, humans may perceive AI as independent and capable of providing subjective and value-laden judgments, which might lead to other relational and functional judgments of AI chatbots.

AI agency and human perceptions

Research in social cognition and interpersonal communication has long explored two fundamental dimensions of social perceptions: communal (seeing an actor as either hostile or friendly) or agentic (seeing an actor as assertive or submissive, Bakan, 1966; for a detailed review, see Schauf et al., 2023). Essentially, the two dimensions indicate the motivation and competence of an actor to pursue individual goals, while they may (or may not) engage in meaningful social relations with others (Abele et al., 2016). In the context of human–AI communication, these dimensions take on new relevance. When people engage in interactions with AI, they often assess the AI’s behavior in terms of agency—how AI pursues its objectives in a self-reliant and effective way—and sociability, which relates to the AI’s perceived friendliness and social engagement. Moreover, the framework from interpersonal communication underscores the enduring relevance where individuals apply these dimensions to understand and navigate their interactions with another social actor, highlighting the importance of considering agency in shaping perceptions of sociability, especially when the actor is an AI.

Specifically, our study focuses on three functional and relational aspects that characterize human perceptions of AI: trust, liking, and perceived chat quality. Trust involves “a cognitive process associated with one’s confidence in another’s goals or purposes, and the perceived sincerity of another’s word” (Tanis and Postmes, 2005: 413). Liking refers to the positive feelings and attitudes one holds toward another person (Rubin, 1970). Substantial literature has noted that perceptions such as trust and liking can induce positive affect and favorable cognitive responses (Feng and MacGeorge, 2010; Perloff, 2008).

In human–computer interaction (HCI), trust has been defined “as the attitude that an agent will help achieve an individual’s goals in situations characterized by uncertainty and vulnerability” (Lee and See, 2004: 51). Trust has been studied extensively in HCI and human–robot interaction (HRI) as beliefs, attitudes, intentions, and behaviors (e.g., Liu, 2021; also for detailed reviews, see Glikson and Woolley, 2020; Hancock et al., 2020; Lee and See, 2004). In the current study, we follow the interpersonal tradition to conceptualize and operationalize trust. That is, trust is not viewed solely as evaluations of the performance and competence of machines, but rather as people’s subjective perceptions and experiences of sincerity, candidness, and trustworthiness within the context of human–AI communication.

Liking is another factor that has been studied extensively in HCI and HMC. Early studies have explored how similarity between humans and computers/robots may influence humans’ psychological responses such as liking and social attraction to computers/robots (Isbister and Nass, 2000; Lee et al., 2006; Moon and Nass, 1996). For example, several studies have examined how humans’ self-disclosure to the chatbots/robots’ self-disclosure/reciprocal self-disclosure affect humans’ perceived intimacy and liking of chatbots/machines (e.g., Ho et al., 2018; Mou et al., 2024). Other research has focused on the perceived similarities in personalities between humans and chatbots/machines affecting humans’ liking and evaluations of the chatbots’ machines (e.g., Moon and Nass, 1996; Nass et al., 1996).

With trust and liking focusing on the relational aspect of human–AI communication, chat quality evaluation focuses on the functional aspect. Considering that AI chatbots are technological products designed to fulfill the role of conversation partners, Guzman and Lewis (2020) advocated that the differences between humans and machines (in our context, AI chatbots) in their functions as communicators should be studied to better understand the role of technology as a unique type of communicator.

Linking source autonomy with social perceptions

In human–AI communication, AI’s autonomy may affect humans’ perceived trust, liking, and chat quality evaluation of the AI chatbots. The extended CASA framework posits that through communicating with media, humans may develop mental models and scripts for media agents (Gambino et al., 2020). When interacting with AI chatbots, people would use media agents’ heuristics or cues to form impressions of and evaluate AI chatbots and may perceive that AI has high agency over their own actions as independent and accountable. Rather than focusing on the anthropomorphism of the media agents, the extended CASA framework argues that the social affordance, bandwidth, and interaction experiences between humans and media agents have promoted humans to adapt and develop criteria and scripts unique to media agents (Gambino et al., 2020).

In human–AI communication, AI is viewed as a counterpart of humans to make conversations with. Using interpersonal communication as an analogy, individuals view independent and self-reliant communication partners as more competent and capable of making decisions. Perceptions of trustworthiness, liking, and functional evaluations such as conversation qualities are closely connected with the targets’ competence and consistency/predictability (McKnight and Chervany, 2002). In interpersonal communication settings, self-reliant, and competent individuals would be judged as more trustworthy, more likable, and fulfilling the role of conversation partners more efficiently (Simpson, 2007). For example, a person would trust and like others who have agency over their own actions and do not require approval nor reply to others; and they would also rate conversations with self-reliant others as more efficient, since it doesn’t require third-parties’ involvement. Similarly, mindful or autonomous agents appear to be capable of controlling actions and better able to perform their intended functions (Waytz et al., 2014). Autonomous AI is independent and is responsible for its own actions and behaviors, while human-assisted AI may indicate that the AI chatbot is not competent or its performance has to be supervised by humans. Compared to human-assisted AI, autonomous AI’s conversations would be evaluated as more consistent or predictable, since its performance does not require humans’ intervention. Therefore, we propose:

Linking language subjectivity with social perceptions

By using subjective language, AI may trigger humans to form social perceptions of AI and use social responses when interacting with AI (Molina and Sundar, 2022, 2024). On one hand, although not identical, subjective language also coincided with the assertiveness dimension of agency in social perceptions (Abele et al., 2016), and it entails an actor’s subjectivity and independence. Although, as humans, we may have our distinct styles of language use; for example, some people prefer assertive language while others hold a tentative style of expression. Impression formation is a complicated process and the perceiver, the target, and the relationship may all play a part in this process. However, in social interactions, compared to objective language, which does not reflect any personal opinions or attachment, people generally prefer to talk to others who at least have a perspective. Subjective language typically contains value judgments, opinions, and assumptions and reflects a particular perspective. The reflected perspective in subjective language may not be necessarily more accurate or have better quality compared to objective language. Nonetheless, language-reflecting perspectives could prime a sense of subjectivity, which fundamentally differentiates AI chatbots from mindless machines (Waytz et al., 2014). This subjectivity reflected in language use may indicate that an agentic actor is actively engaged in and constantly responding to the environment.

On the other hand, the level of personalization through which media agents could offer tailored feedback to the user would encourage users to respond in a more personalized way with greater social responses (Fox and McEwan, 2017). In our case, compared to objective language, which shares more similarities with computer language, the use of subjective language largely resembles some levels of personalization in language use; therefore, it would promote people to feel affinity to the AI chatbots, and develop higher trust, and rate the chat with better quality.

Taken together, when interacting with AI using subjective language, humans may perceive AI as having a perspective and capable of providing subjective and value-laden judgments, leading to higher ratings of trustworthiness, likability, and chat quality for subjective AI chatbots.

The joint effect of AI’s autonomy and language use on human perception

Furthermore, AI’s source autonomy and language subjectivity may have an additive effect on humans’ perceptions of AI. Similar to earlier predictions on humans’ perceived trustworthiness, liking, and chat quality evaluation, both source autonomy and language subjectivity would serve as cues for humans to evaluate and form impressions. When language subjectivity interacts with source autonomy, autonomous AI chatbots using subjective language may be evaluated as the most agentic, trustworthy, likable, and having the best chat quality.

AI literacy in human–AI interaction

When examining how users form relational and functional evaluations of chatbots in human–AI interactions, one crucial predictive factor is the extent to which they are knowledgeable about the basic functioning mechanism of AI, aware of the advancement of artificial intelligence, and capable of integrating AI into their everyday life (Ng et al., 2021; Wang et al., 2023). The notion of AI literacy has been defined as “a set of competencies that enables individuals to critically evaluate AI technologies; communicate and collaborate effectively with AI; and use AI as a tool online, at home and in the workplace” (Long and Magerko, 2020: 599). Previous research has developed a multidimensional conceptual framework for AI literacy, encompassing aspects such as knowledge and understanding of AI technology, evaluation and application skills, and consideration of AI ethics (see Ng et al., 2021 for a detailed review).

In the context of human communication with AI chatbots, we argue that AI literacy could either enable or constrain users’ ability to assess and apply knowledge and concepts during conversations with AI interlocutors, and thus, literacy could act as a significant moderator influencing the impact of AI chatbots’ source autonomy and language style on users’ trust, liking and overall evaluation of chat quality. Furthermore, people who possess greater AI literacy in the latest advancements in machine intelligence may have higher expectations regarding how AI could potentially reason and learn. As a result, they may interpret interactions with chatbots displaying varying levels of agency differently. In particular, we focus on awareness, usage, and evaluation of AI applications within Wang et al.’s (2023) framework. We exclude ethics because we contend that ethics play a less significant role when individuals form initial impressions during their encounters with an AI interlocutor. Given that little empirical research has explored how individuals’ AI literacy would affect the associations between AI agency and their relational and functional evaluations of AI chatbots, we propose the following research question:

Participants

This study received IRB approval from the first and corresponding authors’ university. Participants were recruited from Amazon Mechanical Turk. The study was introduced as an AI chat experience study on Mechanical Turk and only participants who are located in the U.S. were allowed to participate. The final sample had 376 valid responses including 139 female participants, 236 male participants, and 1 participant chose “prefer not to answer” for the biological sex question. The average age of the participants is 34.15 (SD = 9.71). Most participants identified themselves as White or Caucasian (89.63%), followed by Asian (4.78%), Black or African American (3.46%), Latino/Hispanic (1.86%), and others (0.27%). In terms of education, most participants reported having a bachelor’s degree (64.10%), followed by 16.22% of participants with a graduate and professional degree, and the rest 19.69% indicated less than a bachelor’s degree.

Procedure

The experiment was a 2 (AI’s autonomy: autonomous AI vs human-assisted AI) × 2 (AI’s language subjectivity: subjective vs objective) × 2 (topics: traveling versus reading) factorial design. These two topics (i.e., traveling and reading) were included to increase the external validity, and they were deliberately chosen because they are common topics for initial social encounters in interpersonal communication. Participants were recruited from MTurk and redirected to a Qualtrics survey page to complete the online experiment. Each of them received $1.5 in compensation. The results indicated that the mean time for participants to finish the study is 902.64 seconds (approximately 15 minutes) with a standard deviation of 520.52 seconds (8.68 minutes).

In the beginning, participants were informed that the study involved chatting with an AI chatbot, and they were asked questions regarding the conversation afterward. Before being directed to the online chat interface, participants first answered questions in regard to their AI literacy, and their pre-existing attitudes toward AI. Then, participants were randomly assigned to either autonomous or human-assisted AI conditions and were presented with the corresponding introduction (exemplified in Figure 1). After reading through the introduction, they were randomly assigned to AI chatbots with either subjective or objective language styles under the topic of traveling or reading. Participants then engaged in online chatting with the chatbots. To facilitate the chat and resemble real-life interaction between humans and chatbots, during each turn of the conversation, the chatbots asked participants to make comparisons or provide their opinions or attitudes. After the online chatting, we conducted manipulation checks among participants, and then asked them to rate their liking and trust of the chatbots, and their overall evaluation of the conversation quality.

OPEN IN VIEWER

Figure 1.

Prompts for source agency manipulation (upper: fully autonomous; lower: human-assisted).

Stimuli

Measurements

Manipulation check

Preliminary analysis

Multivariate analyses of covariance (MANCOVA) with participants’ AI literacy and pre-existing attitudes toward AI included as covariates were conducted. SPSS 26 (IBM Corps) software was used for the statistical analyses. The MANCOVA results indicated that topic manipulation has no main effect on trust, F_trust(1, 366) = 0.52, p = .47, on liking, F_liking(1, 366) = .24, p = .63, or on chat quality, F_{chat_quality}(1, 366) = 0.62, p = .43. Also, there were no significant two-way interaction effects between topics and source autonomy on trust, F_trust(1, 366) = 1.26, p = .26, on liking, F_liking(1, 366) = 0.76, p = .38, or on chat quality, F_{chat_quality}(1, 366) = 0.04, p = .84. The two-way interactions between topics and language subjectivity were not significant for trust, F_trust(1, 366) = .01, p = .98, for liking, F_liking(1, 366) = .00, p = .89, or for chat quality, F_{chat_quality}(1, 366) = 0.003, p = .95. The three-way interactions among topics, source autonomy, and language subjectivity were not significant for trust, F_trust(1, 366) = 0.05, p = .83, for liking, F_liking(1, 366) = .12, p = .73, or for chat quality, F_{chat_quality}(1, 366) = 0.004, p = .95. Therefore, topics were not included in the final analyses.

In terms of the covariates, participants’ AI literacy (M = 4.06, SD = .57) was significantly related to their trust of the AI chatbots, F(1, 369) = 16.12, p < .001, partial η_p² = 0.04, their evaluation of chat quality, F(1, 369) = 27.54, p < .001, partial η_p² = 0.07, but not their liking of the AI chatbots, F(1, 369) = 3.74, p = .054. Participants’ pre-existing positive attitude toward AI (M = 4.08, SD = .62) significantly predict their trust of the AI chatbots, F(1, 369) = 34.60, p < .001, partial η_p² = 0.09, their liking of the AI chatbots, F(1, 369) = 33.14, p < .001, partial η_p² = 0.08, and their evaluation of chat quality, F(1, 369) = 9.50, p = .002, partial η_p² = 0.03. Participants’ pre-existing negative attitude toward AI (M = 3.65, SD = 0.97) was significantly associated with their trust of the AI chatbots, F(1, 369) = 5.99, p = .02, partial η_p² = 0.02, their liking of the AI chatbots, F(1, 369) = 40.23, p < .001, partial η_p² = 0.10, but not their evaluation of chat quality, F(1, 369) = 3.77, p = .053. Descriptive statistics, estimated marginal means, and standard error can be found in Table 1. MANCOVA results can be found in Table 2.

OPEN IN VIEWER

Table 1.

Observed means (SDs) and estimated marginal means (SEs) for trust, liking, and chat quality.

AI’s source agency	AI’s language agency	Observed means (SDs)			Estimated marginal means (SEs)
		Trust	Liking	Chat quality	Trust	Liking	Chat quality
Autonomous	Objective	3.93 (0.68)	6.55 (1.89)	4.03 (0.72)	3.97 (0.05)	6.67 (0.14)	4.07 (0.05)
	Subjective	4.23 (0.49)	7.11 (1.35)	4.20 (0.49)	4.27 (0.05)	7.25 (0.14)	4.24 (0.05)
	Total	4.08 (0.61)	6.83 (1.66)	4.11 (0.62)	4.12 (0.03)	6.96 (0.10)	4.16 (0.04)
Human-assisted	Objective	4.07 (0.61)	7.02 (1.68)	4.16 (0.70)	4.03 (0.06)	6.91 (0.14)	4.11 (0.05)
	Subjective	3.97 (0.61)	7.04 (1.73)	4.09 (0.63)	3.93 (0.05)	6.90 (0.14)	4.05 (0.06)
	Total	4.02 (0.61)	7.03 (1.70)	4.13 (0.66)	3.98 (0.03)	6.90 (0.10)	4.08 (0.04)

OPEN IN VIEWER

Table 2.

MANCOVA predicting trust, liking, and chat quality.

Variables	Trust			Liking			Chat quality
	F	p	ηp 2	F	p	η 2	F	p	ηp 2
AI literacy	16.12	< .001	.04	3.74	.054	-	27.54	< .001	.07
AI positive attitude	35.60	< .001	.09	33.14	< .001	.08	9.50	.002	.03
AI negative attitude	5.99	.015	.02	40.23	< .001	.10	3.77	.053	-
AI source agency	9.61	.002	.03	.16	.69	-	2.13	.15	-
AI language agency	4.68	.031	.01	4.46	.035	.01	.89	.35	-
AI source agency × AI language agency	18.34	< .001	.05	4.79	.029	.01	4.94	.027	.01
R 2	.45			.40			.36
Adjusted R2	.44			.39			.35

Hypotheses testing

H1 was regarding the main effects of AI chatbots’ source autonomy on participants’ trust, liking, and chat quality evaluation. The results showed that participants’ trust of the AI chatbots was higher for autonomous AI (M = 4.12, SE = .03), compared to human-assisted AI (M = 3.98, SE = .03), F(1, 369) = 9.61, p = .002, partial η_p² = .03. In terms of liking, participants who had conversations with autonomous AI (M = 6.96, SE = .01) did not differ from those who chatted with human-assisted AI (M = 6.90, SE = .10), F(1, 369) = .16, p = .69. Participants who chatted with autonomous AI (M = 4.16, SE = .04) did not differ from those who chatted with human-assisted AI (M = 4.08, SE = .04) in their evaluations of the chat quality, F(1, 369) = 2.13, p = .15. Therefore, H1(a) was supported, while H1(b) and H1(c) were not.

H2 was regarding the main effects of AI chatbots’ language subjectivity on participants’ trust, liking, and chat quality evaluation. The results showed that participants’ trust of the AI chatbots was higher for subjective AI (M = 4.10, SE = .03), compared to objective AI (M = 3.99, SE = .03), F(1, 369) = 4.68, p = .03, partial η_p² = 0.01. Participants who chatted with subjective AI (M = 7.07, SE = .10) indicated higher levels of liking toward AI chatbot compared to those who chatted with objective AI (M = 6.79, SE = .10), F(1, 369) = 4.46, p = .04, partial η_p² = 0.01. Participants who chatted with subjective AI (M = 4.15, SE = .04) did not differ from those who chatted with objective AI (M = 4.09, SE = .04) in their evaluations of the chat quality, F(1, 369) = 0.89, p = .35. Therefore, H2(a) and H2(b) were supported, while H2(c) was not.

H3 predicted two-way interactions between AI chatbots’ source autonomy and language subjectivity on participants’ trust, liking, and chat quality evaluation. The two-way interaction between source autonomy and language subjectivity was significant for participants’ trust of AI chatbots, F(1, 369) = 18.34, p < .001, partial η_p² = 0.05, for participants’ liking of the AI chatbots, F(1, 369) = 4.79, p = .03, partial η_p² = 0.01, and for participants’ evaluation of the chat quality, F(1, 369) = 4.94, p = .03, partial η_p² = 0.01.

In order to compare participants’ trust of AI in the autonomous AI chatbots using subjective language conditions to the other three conditions, planned contrast analysis was performed. The results showed that participants who chatted with autonomous AI using subjective language (M = 4.23, SD = .49) had higher trust of AI compared to the ones who chatted with autonomous AI using objective language (M = 3.93, SD = .68), t(372) = 3.43, p = .001, Cohen’s d = 0.50, and compared with the ones who chatted with human-assisted AI using subjective language (M = 3.97, SD = .61), t(372) = 2.88, p = .004, Cohen’s d = 0.46, but not different compared to the ones who chatted with human-assisted AI using objective language (M = 4.07, SD = .61), t(372) = 1.73, p = .08. Therefore, H3(a) was partially supported.

Another planned contrast analysis was performed to compare participants’ liking of AI in the autonomous AI chatbots using subjective language conditions to the other three conditions. The results showed that participants who chatted with autonomous AI using subjective language (M = 7.11, SD = 1.35) had higher liking of AI compared to the ones who chatted with autonomous AI using objective language (M = 6.55, SD = 1.89), t(372) = 2.32, p = .021, Cohen’s d = 0.34, but not different compared with the ones who chatted with human-assisted AI using subjective language (M = 7.04, SD = 1.73), t(372) = .31, p = .76, or the ones who chatted with human-assisted AI using objective language (M = 7.02, SD = 1.68), t(372) = .37, p = .71. Therefore, H3(b) was partially supported.

In order to compare participants’ evaluation of chat quality while interacting with autonomous AI chatbots using subjective language to the other three conditions, planned contrast analysis was performed. The results showed that participants who chatted with autonomous AI using subjective language (M = 4.20, SD = .49) rated the chat with higher quality compared to the ones who chatted with autonomous AI using objective language (M = 4.02, SD = .72), t(372) = 1.99, p = .048, Cohen’s d = 0.29, but not different compared with the ones who chatted with human-assisted AI using subjective language (M = 4.09, SD = .63), t(372) = 1.18, p = .24, or the ones who chatted with human-assisted AI using objective language (M = 4.16, SD = .70), t(372) = .47, p = .64. Therefore, H3(c) was partially supported.

To examine the research question with regard to the role of participants’ AI literacy toward AI, we used model 3 of SPSS PROCESS Macro (Hayes, 2013), and tested its moderating effects on the main and interaction effects of source autonomy and language subjectivity on the dependent variables. The results showed that the two-way interaction between AI literacy and source autonomy was not significant for participants’ trust of AI chatbots, b = -.42, t(366) = -1.67, p = .10, while significant for participants liking of AI, b = -1.65, t(366) = -2.24, p = .026, and their chat quality evaluation, b = -.59, t(366) = -2.04, p = .042. The effects of source autonomy on liking and chat quality evaluation were stronger for people with lower AI literacy. The two-way interaction between AI literacy and language subjectivity was significant for participants’ trust of AI chatbots, b = -.66, t(366) = -2.61, p = .009, participants’ liking of AI chatbots, b = -1.49, t(366) = -2.00, p = .045, and their chat quality evaluation, b = -.73, t(366) = -2.51, p = .013. The effects of language subjectivity on trust, liking, and chat quality evaluation were stronger for people with lower AI literacy. Participants’ AI literacy moderated the interaction effect of source autonomy and language subjectivity on trust, b = .33, t(366) = 2.01, p = .045, liking, b = .97, t(366) = 1.99, p = .047, and chat quality, b = .43, t(366) = 2.24, p = .03. As shown in Table 3, autonomous AI with subjective language was perceived as more trustworthy, likable, and with higher quality than human-assisted AI with subjective language, and the effect was stronger among people who possess a lower level of AI literacy.

OPEN IN VIEWER

Table 3.

Moderated moderation effect of AI literacy.

		Trust				Liking				Chat quality
		b	SE	t	p	b	SE	t	p	b	SE	t	p
Source × language × AI literacy		.33	.17	2.01	.045	.97	.48	1.99	.047	.43	.19	2.24	.03
	AI literacy
Objective language	M − SD	.09	.09	1.02	.31	.62	.26	2.36	.02	.12	.10	1.21	.23
	M	.03	.67	.51	.61	.16	.20	.84	.40	.02	.08	.20	.84
	M + SD	.01	.08	.06	.95	−.06	.24	−.27	.79	−.04	.09	−.42	.68
Subjective language	M − SD	−.48	.10	−4.89	.00	−.51	.29	−1.79	.07	−.34	.11	−3.05	.01
	M	−.32	.07	−4.70	.00	−.32	.20	−1.66	.09	−.17	.08	−2.20	.03
	M + SD	−.23	.08	−2.78	.01	−.23	.25	−.94	.35	−.08	.10	−.85	.40

Practical implications

Past literature on human–AI interaction has focused on the importance of involving end-users in iterative design processes to foster their understanding of AI systems (Long and Magerko, 2020). The findings of our study also shed light on important design considerations. Our results indicate that autonomous AI chatbots articulating using language reflecting their own perspective are considered the most trustworthy and likable, with notably higher interaction quality. This result underscores the importance of designing chatbots to use language that suggests a form of perspective or intentionality. This could involve crafting responses that imply thoughtfulness or consideration, even if the underlying AI lacks true agency. While maintaining accuracy and objectivity of interactions, we advocate for a nuanced approach that aligns AI chatbots’ language style with its perceived autonomy to optimize user engagement and satisfaction. For instance, guidelines could be developed for matching more objective or subjective language styles to different levels of system autonomy across AI applications. Moreover, our conceptual and methodological approach highlights that clearly communicating the extent of the AI’s autonomy and capabilities to users is essential in managing expectations and building trust. Alongside design improvements, our results on the conditional effect of AI literacy also suggest that promoting the understanding of how AI chatbots function could reduce frustrations arising from misaligned expectations and enhance user trust and acceptance.

Limitations and future directions

There are several limitations pertaining to the current study. The first limitation is that the semi-scripted responses from AI chatbots may not fully resemble natural conversations people have with AI chatbots. In order to ensure the internal validity of the experiment, AI chatbot’s responses were presented to the participants in chronological sequence across eight conversational turns. While efforts were made to design the chat interface to replicate an actual online chat environment and to use ChatGPT to generate responses representing different language styles, future studies could consider using more ecologically valid approaches such as ecological momentary assessment, or longitudinal observations of humans interacting with AI chatbots in real life scenarios.

Another limitation in our research design is that the chatbot led the conversation by initiating a question at each turn, whereas in real life, people may not choose to suspend, resume, or redirect the topic of the conversation at their convenience. In addition, the current study was conducted as a one-time chat session between the participants and the AI chatbots, with each session lasting approximately 10 to 15 minutes. The limited and less sustained duration of the chat session could have implications for the depth and complexity of human–AI interactions. Future studies can build upon the current research design and incorporate observational designs or interviews to replicate the findings or to find boundary conditions that influence the dynamics of human–AI communication.

The third limitation pertains to the AI chatbots’ embodiments. In order to maintain consistency and control for possible confound, we utilized the same visual designs for the AI chatbots in different experimental conditions. The visual design of cartoon robots with less anthropomorphism features may prime participants in the machine or robot identity of the chatbots, compared to the more popular human-like avatars with customizable features. This may limit the generalizability of the findings. Future studies could adopt different visual designs of the chatbots with various features of embodiment to test the conditional effects of the AI agency.

Last but not least, the effect sizes observed in this study were rather small, which could be due to the short interaction time and the online nature of the experimental design. Considering that the sample was acquired from MTurk the study was conducted online, longer interaction time or multiple conversation sessions could introduce other confounding variables, and the experimental fatigue could also affect the results. Future studies could also utilize multiple online chat sessions or in-lab experiments to extend the findings.