Costly “Greetings” from AI: Effects of Product Recommenders and Self-Disclosure Levels on Transaction Costs
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsManuscript Review
Dear Yasheng Chen & Yuhong Tu:
Your manuscript “Costly “Greetings” from AI: Effects of Product Recommenders 2 and Self-disclosure Levels on Transaction Cost” requires some further minor changes:
· The authors commented on the manipulation in the Results section. The definition of the variable is missing in the Introduction and Background. This is important because it is being considered as a hypothesis test.
o Recommended authors and/or perspectives:
§ Jowett & O’Donnell (Propaganda & Persuasion / Coons & Weber (Manipulation) / Jongepier & Klenk (The Philosophy of Online Manipulation) refer, specially to the digital Manipulation.
· We consider enhancing the Introduction section with the discussion/perspective of the authors supporting the non-full digitalization and dependence on AI in algorithms to suggest and recommend products/services online and offline. (Lines 23-32)
· Include some context on how NLP AI tries to replicate the human way of communication, as it attempts to learn the intent and sentiment of the user (consumer) towards LLMs, voice assistants, and chatbots.
· The authors should discuss limitations in sample size (Methods section) and ethical considerations related to AI in consumer interactions.
o Check Kotler et al., 2024, Marketing 6.0, Wiley.
· It would be interesting to include a discussion section to compare the views of other articles/authors regarding the use of AI in consumer behavior studies.
· The authors should address potential biases in data collection, cultural differences, and long-term effects of AI-driven emotional support on consumer trust and loyalty.
· A section regarding ethical considerations, privacy, data security, and the potential manipulation of emotional responses is needed.
· Please add a separate section for ‘Data’ and explain what data has been used in detail, how the data was obtained, the nature of the data, the source of the data, the size of the data, the details relating to the accessibility of the data, and more importantly, details of rights and permissions required to obtain them.
The manuscript is considered an important contribution to the study of digital marketing, especially in the influence that AI has on the modern consumer.
Author Response
Comments 1: The authors commented on the manipulation in the Results section. The definition of the variable is missing in the Introduction and Background. This is important because it is being considered as a hypothesis test. Recommended authors and/or perspectives: § Jowett & O’Donnell (Propaganda & Persuasion / Coons & Weber (Manipulation) / Jongepier & Klenk (The Philosophy of Online Manipulation) refer, especially to the digital Manipulation.
Response 1: Thank you for pointing this out. We agree with this comment. Therefore, we have clearly defined the variables and their definitions, explained in footnotes 1-4: transaction cost, self-disclosure, emotional support, and AI recommender, respectively. Specifically:
Footnote 1. Drawing from transaction cost theory, transaction cost refers to the expenses involved when buyers (customers) and sellers (retailers) engage in trade (Celmons & Kimbrough, 1986; Devaraj et al., 2002). These costs encompass a variety of activities including the search for suitable trading partners, the acquisition of product information and pricing, the drafting of contracts, the actual purchasing process, and the enforcement of these contracts. In our setting, we concentrate on the supplier-side transaction costs incurred during interactions with different product recommenders.
Footnote 2. Self-disclosure refers to the communication of private information to another (Greene et al., 2006). This includes any information that refers to the self, such as personal dispositions, events in the past, or current or future plans of action (Greene et al., 2006; Al-Natour et al., 2021).
Footnote 3. Emotional support is a concept derived from social support and refers to the provision of care, concern, empathy, love, and trust (Kort-Butler, 2017).
Footnote 4. An AI recommender refers to any type of autonomous system that uses algorithms to produce recommendations for consumers (Wien & Peluso, 2021).
Meanwhile, following the references you provided, we have incorporated additional literature into our manuscript. This enhancement will aid in adhering to ethical and moral principles during the online experiment and in enhancing the validity of the manipulation of variables. Please see the updated manuscript and modifications made on lines 477-481. Specifically, “Additionally, our manipulation of independent variables online was guided by the ethical and moral principles outlined by Jongepier and Klenk [88] and Coons and Weber [89]. Adherence to these guidelines was imperative to ensure ethical compliance, fairness, and transparency throughout the execution of this online experiment”.
Comments 2: We consider enhancing the Introduction section with the discussion/perspective of the authors supporting the non-full digitalization and dependence on AI in algorithms to suggest and recommend products/services online and offline. (Lines 23-32)
Response 2: Thanks for your comment. We have, accordingly, modified our manuscript to emphasize this point. (see Lines 35-41). Specifically, However, several AI robotics firms, including Softbank’s Pepper unit [9] and Meta's M assistant, have recently scaled back or discontinued their AI service operations [10], which underscores the potential limitations and risks associated with overreliance on AI technologies. The constraints inherent in AI training data may lead to the oversight of the needs of diverse groups and raise significant concerns regarding privacy breaches associated with user data [11], underscoring the caution against sole reliance on AI recommendations.
Comments 3: Include some context on how NLP AI tries to replicate the human way of communication, as it attempts to learn the intent and sentiment of the user (consumer) towards LLMs, voice assistants, and chatbots.
Response 3: Thank you for pointing this out. We agree with this comment. We modified our manuscript to emphasize this point. (See Lines 26-35). Specifically, in the realm of e-commerce, AI-driven chatbots are meticulously engineered to replicate human communication by accurately discerning and responding to the underlying intent and sentiment in user interactions. This sophisticated emulation, facilitated by advances in Natural Language Processing (NLP) and Large language models (LLMs), substantially boosts operational efficiency and lowers transaction costs [3-5]. For instance, Amazon utilizes user-based AI technologies to match users with similar tastes and provide tailored product recommendations based on their historical behavior [6]. Similarly, cosmetics companies like Sephora and Kiehl's employ AI chatbots to recommend products tailored to customers' preferences and skin types [7, 8].
Comments 4: The authors should discuss limitations in sample size (Methods section) and ethical considerations related to AI in consumer interactions. Check Kotler et al., 2024, Marketing 6.0, Wiley.
Response 4: Thanks for your comment. We have, accordingly, modified our manuscript to emphasize this point.
For limitations in sample size, please see Line 426 and Footnote 5. Specifically, A power analysis was conducted to determine the required sample size for this study (Cohen, 1992), utilizing a power value of 0.8, an effect size (Cohen's f) of 0.4, equal group weights of 1 for four groups, and a significance level of 0.05. The calculated sample size was 76, with 19 participants per group. Our actual sample size is 78, which exceeds the calculated requirement of 76. This indicates that the study design has adequate statistical power, thereby enhancing the reliability and validity of the results. The increased sample size further ensures more precise detection of the anticipated effects and minimizes the risk of Type II errors.
For ethical considerations related to AI in consumer interactions, please see footnote 6, lines 492-496, and lines 542-548. Specifically, footnote 6: all experiments reported in this paper were reviewed and approved by the relevant IRBs. We adhered to the principles outlined by Jongepier and Klenk [87] and Coons and Weber [88] in the online experiment. Line 492-496: Additionally, our manipulation of independent variables online was guided by the ethical and moral principles outlined by Jongepier and Klenk [88] and Coons and Weber [89]. Adherence to these guidelines was imperative to ensure ethical compliance, fairness, and transparency throughout the execution of this online experiment.
Line 542-548: All experiments reported in this paper underwent review and approval by the relevant Institutional Review Boards (IRBs). The data collection process adhered to established legal standards and ethical guidelines, including the Declaration of Helsinki, the International Ethical Guidelines for Health-Related Research Involving Humans, and the General Data Protection Regulation (GDPR). The experiments posed no foreseeable risks or hazards to participants.
Comments 5: It would be interesting to include a discussion section to compare the views of other articles/authors regarding the use of AI in consumer behavior studies.
Response 5: Thanks for your comment. We have modified our manuscript to emphasize this point. Please see the Table 1 (Line 295). Specifically,
Table 1 The effect of AI on human behavior
|
Author |
Method |
Main findings |
|
Boudorf et al.[6] |
Randomized controlled trial |
When using digital advice, consumers were willing to pay 14% less for popular brand plans and 37% less for plans with higher star ratings, compared to having only basic product information. |
|
Kim et al.[53] |
Experiment |
Consumers are more likely to engage in unethical behaviors when interacting with AI agents, due to reduced anticipatory feelings of guilt. |
|
Leo and Huh [46] |
Experiment |
When service fails, people attribute less responsibility toward a service provider if it is a robot rather than a human. People attribute more blame toward a service firm when a robot delivers a failed service than when a human does. |
|
Huo et al. [47] |
Survey |
Patients' self-responsibility attribution is positively related to human–computer trust (HCT) and sequentially enhances the acceptance of medical AI for independent diagnosis and treatment. |
|
You et al. [48]
|
Experiment |
Individuals follow algorithmic advice more than identical human advice due to higher trust in algorithms, and this trust remains unchanged even when they are informed of the algorithm's prediction errors. |
|
Filieri et al. [54] |
Machine learning |
The majority of customer interactions with service robots were positive, and robots that moved triggered more emotional responses than stationary ones. |
|
Berger et al. [51] |
Experiment |
For an objective and non-personal decision task, human decision makers exhibit algorithm aversion if they are familiar with the advisor’s performance and the advisor errs. |
|
Commerfold et al. [17] |
Experiment |
Auditors proposed smaller adjustments to management's complex estimates when receiving contradictory evidence from an AI system rather than a human specialist. This effect was particularly pronounced when the estimates were based on relatively objective inputs. |
|
Filiz et al. [52] |
Experiment |
Algorithm aversion occurs more frequently as the seriousness of the decision's consequences increases. |
Comments 6: The authors should address potential biases in data collection, cultural differences, and the long-term effects of AI-driven emotional support on consumer trust and loyalty.
Response 6: Thank you for pointing this out. We agree with this comment. Indeed, cultural background can significantly influence individuals' attitudes toward AI, and this paper primarily examines a sample of Chinese university students, among whom cultural differences are relatively minimal. It is noteworthy that a potential research direction could involve collecting data from different countries (such as China and the United States) to explore variations in consumer attitudes towards AI under different cultural contexts.
Additionally, individuals' acceptance of AI may change over time, such as through learning effects. Currently, AI is widely accepted in the context of product recommendations, and the results of this study somewhat demonstrate that high self-disclosure can further enhance AI's positive impact on transaction costs. Additionally, maintaining participant engagement over long periods poses challenges, and participant dropouts can result in incomplete data, thus affecting the reliability of study outcomes. Long-term studies can be effectively conducted through field experiments.
Indeed, these two comments remain crucial influencing factors, and we will consider these points as directions for future research (Line 764-773).
Comments 7: A section regarding ethical considerations, privacy, data security, and the potential manipulation of emotional responses is needed.
Response 7: Thanks for your comment. We have modified our manuscript to emphasize this point. Please see in Section 3.6 Ethical Consideration (line 542-548), and Section 3.4 Data (line 491-496). Specifically, the Ethical consideration: All experiments reported in this paper underwent review and approval by the relevant Institutional Review Boards (IRBs). The data collection process adhered to established legal standards and ethical guidelines, including the Declaration of Helsinki, the International Ethical Guidelines for Health-Related Research Involving Humans, and the General Data Protection Regulation (GDPR). The experiments posed no foreseeable risks or hazards to participants.
Data: These data were collected anonymously and encrypted by the researchers to ensure the security and confidentiality of the information. Additionally, our manipulation of independent variables online was guided by the ethical and moral principles outlined by Jongepier and Klenk [88] and Coons and Weber [89]. Adherence to these guidelines was imperative to ensure ethical compliance, fairness, and transparency throughout the execution of this online experiment.
Comments 8: Please add a separate section for ‘Data’ and explain what data has been used in detail, how the data was obtained, the nature of the data, the source of the data, the size of the data, the details relating to the accessibility of the data, and more importantly, details of rights and permissions required to obtain them.
Response 8: Thanks for your comment. We have modified our manuscript to emphasize this point. Please see lines 486-497 and lines 786-788. Specifically,
Line 486-497: Before data collection, participants were furnished with a comprehensive experimental statement and an informed consent document, which clearly outlined the potential risks and benefits associated with their participation in the study. Data such as emotional support, intentions to return, and demographic information were gathered through an online questionnaire. These data were collected anonymously and encrypted by the researchers to ensure the security and confidentiality of the information. Additionally, our manipulation of independent variables online was guided by the ethical and moral principles outlined by Jongepier and Klenk [88] and Coons and Weber [89]. Adherence to these guidelines was imperative to ensure ethical compliance, fairness, and transparency throughout the execution of this online experiment.
Line 786-788: Data Availability Statement: Due to personal privacy issues, research data is available on request from the authors. The data that support the findings of this study are available from the corresponding author, Yuhong Tu, upon reasonable request.
Again, we sincerely appreciate the insightful comments and constructive suggestions provided by you. Your feedback has been instrumental in enhancing the quality and depth of our manuscript. We have carefully considered and incorporated your recommendations, which have significantly improved our study. Thank you for your valuable contributions to our research.
Reviewer 2 Report
Comments and Suggestions for AuthorsWhat is the main question addressed by the research?
This paper focuses on how to recommend the products using AI or using Human support they build an enhanced AI systems that effectively enhance product recommendation processes.
Dear author your research are more valuable thanks for your work and enhance your research with the following comments:
1. The main important issue in your research you should select only 1 reference to be mentioned in your paragraph as example in line 26 from [1-3] and in line 28 from [4-7] , inline 41 from [16-18] please reduce all references you mentioned In this paper you don’t need all this number.
This problem continue in all your research.
2. Please rewrite the abstract to add your results and the datasets used and the methods and techniques used in the paper
3. Also, we need to add the research goal only as points at the end of the introduction section.
4. We can add a paragraph named by the research issues or research gab start form the section To address these research questions line 71
5. From line 136 to 153 we can only add this part at the end of the research as conclusion in details as you write but you need to reduce this paragraph
6. At section 2 we need to add papers results in this research area with their finding in a table
7. We need to a reference to all your Hypothesis 1, 2, 3
8. Please explain how you divide your data into 4 groups as listed in table 1
9. You should add a figure that show your design steps and the use of data as input and output draw your design architecture
10. Write the difference between Panel A: Descriptives and panel B One-way ANOVA
11. If its possible to add a mathematical model how to calculate all your finding Hypothesis 1, 2, 3
12. You need to add the supported references "sparsity and cold start recommendation system challenges solved by hybrid feedback, adaptive learning systems based on ILOS of courses, hash semi cascade join for joining multi-way map reduce, music recommendation system used emotions to track and change negative users’ mood, predict student learning styles and suitable assessment methods using click stream, diabetic mellitus prediction with BRFSS data sets, “The importance of effective learning technology utilization, teacher leadership, student engagement, and curriculum in the online learning environment ","Game-based student e-learning experience: Empirical evidence from private universities in Jordan."
13. Please update your table as the journal format
14. Update the figures as journal format
15. You can rewrite your conclusion in the simple way and add your finding results
16. Also add your future work to this research.
17. Please add author contribution at the end of the paper and conflict of interest
thanks
Comments for author File: 
Comments.pdf
need revision
Author Response
Comments 1:The main important issue in your research you should select only 1 reference to be mentioned in your paragraph as example in line 26 from [1-3] and in line 28 from [4-7] , inline 41 from [16-18] please reduce all references you mentioned In this paper you don’t need all this number.
Response 1: Thank you for your valuable feedback regarding the use of references in our manuscript. We have pruned some references to streamline the text as per your suggestion. We initially included a comprehensive number of citations to robustly support our arguments and conclusions. However, we understand the need for conciseness and have adjusted accordingly. We appreciate your guidance on this matter.
Comments 2: Please rewrite the abstract to add your results and the datasets used and the methods and techniques used in the paper
Response 2: Thank you for your suggestion to enhance the clarity and completeness of the abstract. We have revised the abstract to include detailed descriptions of our research findings, the sample size, and the methods employed in our study. Please see in lines 10-17. Specifically, we recruited 78 participants and conducted a 2x2 online experiment in which we manipulated product recommenders (Human versus AI) and examined how self-disclosure levels (high versus low) affect consumers’ return intentions. We predicted and found a low level of self-disclosure from human recommenders instead of AI counterparts results in higher emotional support, which leads to lower transaction costs. However, under high levels of self-disclosure, consumers’ emotional support and subsequent transaction costs do not differ between human and AI recommenders.
Comments 3: Also, we need to add the research goal only as points at the end of the introduction section.
Response 3: Thank you for your suggestion to add the research goal at the end of the introduction section. In alignment with the structure of our paper, we have chosen to present the research question in a standalone paragraph. This format is intended to serve as a bridge that both introduces the research question and the subsequent sections. You can find this elaboration in lines 120-121 of the manuscript, where we clearly state the research question: "What is the effect of different product recommenders and self-disclosure levels on transaction costs?"
Comments 4: We can add a paragraph named by the research issues or research gab start form the section To address these research questions line 71
Response 4: Thank you for your suggestion to add a paragraph focusing on research issues or gaps. We have restructured the paragraph logic (see lines 120-146) to better align with the research questions of our paper. Using social exchange theory as the theoretical framework, we have constructed a relationship between product recommenders, self-disclosure levels, and transaction costs. This revision strengthens the flow and coherence of our introduction and provides a clear justification for our research focus.
Comments 5: From line 136 to 153 we can only add this part at the end of the research as conclusion in details as you write but you need to reduce this paragraph
Response 5: Thank you for your suggestion. We have relocated the section discussing the practical contributions to the conclusion of the paper, and we have also elaborated on how our research findings relate to sustainability (please see in lines 747-768). This reorganization helps to clarify the implications and enhances the overall coherence of the manuscript.
Lines 747-768: The research findings from this study offer significant practical implications for the deployment of artificial intelligence (AI) in product recommendation scenarios. Our re-sults show that even when the recommender system is AI-based, affixing a human label (e.g., naming it "Jonny Zhang") can substantially reduce transaction costs. This strategy of humanizing AI not only enhances the emotional support perceived by users but also contributes to more sustainable business practices by improving efficiency and reducing the resource drain typically associated with higher transaction costs. Moreover, our findings indicate that low self-disclosure—focusing solely on the functionality of the product recommendations rather than the recommender's personal accolades—further amplifies the positive impact of perceived human attributes on reducing transaction costs. This highlights the importance of how information is presented, emphasizing functional aspects over personal achievements or histories can lead to more effective, cost-efficient outcomes. This approach also leverages AI's ability to provide concise, functional in-formation without overloading the user with unnecessary details, which can lead to more streamlined and efficient consumer decisions. Such efficiency is crucial for sustainability as it reduces the waste of resources, both in terms of the cognitive load on consumers and the operational overhead for businesses. In essence, by focusing on the functional benefits of products rather than the intricacies of the AI itself, companies can foster a more sustainable interaction model that conserves resources while still meeting consumer needs effectively. That is, utilizing AI in emotionally driven interactions effectively manages user expec-tations and minimizes resource waste associated with transaction costs, thereby sup-porting sustainable practices that positively impact environmental and social outcomes.
Comments 6: At section 2 we need to add papers results in this research area with their finding in a table.
Response 6: Thank you for your suggestion. We have conducted a literature review and compiled the relevant studies in Table 1 (see line 295), titled "The Effect of AI on Human Behavior," which summarizes the key findings in this research area.
Comments 7: We need to a reference to all your Hypothesis 1, 2, 3
Response 7: Thank you for your valuable feedback. While we appreciate your suggestion, we believe that the relationships between the variables in Hypotheses 1, 2, and 3 are well-grounded in the theoretical framework outlined in the hypothesis development section, drawing upon existing research. We hope this approach adequately supports our hypotheses.
Comments 8: Please explain how you divide your data into 4 groups as listed in table 1
Response 8: Thank you for your suggestion. We have revised our manuscript to clarify the process of dividing the data into four groups (see lines 454-456). Specifically, we recruited participants from business school alumni through an online ques-tionnaire using Credamo. After clicking the link of this online questionnaire, participants were randomly assigned by Credamo to one of four experimental groups.
Comments 9: You should add a figure that show your design steps and the use of data as input and output draw your design architecture
Response 9: Thank you for your suggestion. We have created the figure as requested; please refer to lines 464-465.
Comments 10: Write the difference between Panel A: Descriptives and panel B One-way ANOVA
Response 10: Please see in lines 571-576. Specifically, Table 3, Panel A, presents the descriptive statistics for transaction costs, including sample sizes, means, standard deviations, standard errors, and the minimum and maximum values for each treatment group. Table 3, Panel B, reports the results of the ANOVA, which tests hypotheses regarding the means of the groups in our dataset. This analysis aims to determine whether there are significant differences in transaction costs between AI and human recommendation subjects.
Comments 11: If its possible to add a mathematical model how to calculate all your finding Hypothesis 1, 2, 3
Response 11: Please see in lines 539-541, Section 3.5 Mathematical model.
Transaction cost = β0 + β1 Product recommenders + β2 Self-disclosure Level + β3 (Product recommenders * Self-disclosure Level) +ϵ
Comments 12: You need to add the supported references "sparsity and cold start recommendation system challenges solved by hybrid feedback, adaptive learning systems based on ILOS of courses, hash semi cascade join for joining multi-way map reduce, music recommendation system used emotions to track and change negative users’ mood, predict student learning styles and suitable assessment methods using click stream, diabetic mellitus prediction with BRFSS data sets, “The importance of effective learning technology utilization, teacher leadership, student engagement, and curriculum in the online learning environment ","Game-based student e-learning experience: Empirical evidence from private universities in Jordan."
Response 12: Thank you for your thoughtful suggestions. While these articles present interesting insights, they are not directly relevant to the focus and scope of our study.
Comments 13: Please update your table as the journal format.
Response 13: Thank you for your suggestion. We will format the table as an image and submit it to the journal editor to meet the requirement of journal format.
Comments 14: Update the figures as journal format
Response 14: Thank you for your suggestion. We will format the figure as an image and submit it to the journal editor to meet the requirement of journal format.
Comments 15: You can rewrite your conclusion in the simple way and add your finding results
Response 15: Thank you for your suggestion. We have placed the conclusion in Section 6, where we have outlined three key practical implications based on our research findings. Additionally, we have discussed the limitations of our study and provided recommendations for future research. Please see lines 747-782.
Specifically, The research findings from this study offer significant practical implications for the deployment of artificial intelligence (AI) in product recommendation scenarios. Our re-sults show that even when the recommender system is AI-based, affixing a human label (e.g., naming it "Jonny Zhang") can substantially reduce transaction costs. This strategy of humanizing AI not only enhances the emotional support perceived by users but also contributes to more sustainable business practices by improving efficiency and reducing the resource drain typically associated with higher transaction costs. Moreover, our findings indicate that low self-disclosure—focusing solely on the functionality of the product recommendations rather than the recommender's personal accolades—further amplifies the positive impact of perceived human attributes on reducing transaction costs. This highlights the importance of how information is presented, emphasizing functional aspects over personal achievements or histories can lead to more effective, cost-efficient outcomes. This approach also leverages AI's ability to provide concise, functional in-formation without overloading the user with unnecessary details, which can lead to more streamlined and efficient consumer decisions. Such efficiency is crucial for sustainability as it reduces the waste of resources, both in terms of the cognitive load on consumers and the operational overhead for businesses. In essence, by focusing on the functional benefits of products rather than the intricacies of the AI itself, companies can foster a more sustainable interaction model that conserves resources while still meeting consumer needs effectively. That is, utilizing AI in emotionally driven interactions effectively manages user expec-tations and minimizes resource waste associated with transaction costs, thereby sup-porting sustainable practices that positively impact environmental and social outcomes.
We conclude with the following three caveats. First, given that product recom-mendation AI systems are already widely implemented across various sectors, conducting a field experiment would provide more authentic insights. By integrating real consumer interactions, a field study could capture genuine behavioral responses and deci-sion-making processes, which are often influenced by dynamic and complex real-world factors. Second, we run this online experiment on single-period performance, which may vary in a multi-period context. Future research should consider longitudinal studies to observe changes in consumer perceptions (e.g., emotional support) and behaviors (e.g., return behavior), as we did not account for the impacts of learning and experience, which they become more accustomed to AI and human interactions [51]. Third, future research should consider the heterogeneity (e.g., cultural differences) of employees. The algorithm aversion effect will lessen as human acquaintance with algorithms increases [101], re-sulting in a reduction in prejudices toward AI. We encourage future research to explore these issues.
Comments 16: Also add your future work to this research
Response 16: Thank you for your feedback. We have revised our manuscript to include three future research directions, highlighting areas where further exploration is needed to build upon our findings (See lines 775-782). Specifically, future research should consider longitudinal studies to observe changes in consumer perceptions (e.g., emotional support) and behaviors (e.g., return behavior), as we did not account for the impacts of learning and experience, which they become more accustomed to AI and human interactions [51]. Third, future research should consider the hetero-geneity (e.g., cultural differences) of employees. The algorithm aversion effect will lessen as human acquaintance with algorithms increases [101], resulting in a reduction in prejudices toward AI. We encourage future research to explore these issues.
Comments 17: Please add author contribution at the end of the paper and conflict of interest
Response 17: Thank you for your suggestion. We have added the author contributions and conflict of interest statements at the end of the paper as follows (see Lines 784-786, 795):
Author Contributions: Conceptualization, Y.C.; methodology, Y.T.; formal analysis, Y.T.; writing—original draft preparation, Y.T.; writing—review and editing, Y.T.; supervision, Y.C.; funding acquisition, Y.C. All authors have read and agreed to the published version of the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe topic of this study is interesting and overall the data analysis is standardized, but there is room for improvement in the logic of the introduction, here are some specific suggestions that I hope will be useful to the authors.
1. In the introduction, the authors mention the motivation for the study: “Therefore, it is imperative to investigate the underlying mechanisms, particularly…”. It is recommended that the underlying mechanisms be recapitulated to describe them more specifically. Also, the sudden appearance of transaction costs was a bit abrupt. I hope the author explains why transaction costs are used as an example.
2. The second and third paragraphs of the introduction do not follow closely together, and the author is expected to add a transitional statement before "self-disclosure".
3. The definition of transaction costs should be further clarified in the introduction, e.g., are they costs for sellers or buyers?
4. In the literature review, the authors mentioned: “AI technologies offer transformative potentials to optimize resource allocation and increase operational efficiencies”. The authors fully recognize the important importance of technologies such as AI and recommendation algorithms to enhance the transactional efficiency of online commerce. Authors could focus more on the role of recommendation affordance in facilitating user experience, e.g., the role of recommendation affordance in social e-commerce can be found in: Attention marketing in fragmented entertainment: how advertising embedding influences purchase decision in short-form video apps.
5. The way p=0.000 is written in Figure 3 is not rigorous.
6. Part 5 is suggested to be split into “5. Discussion” and “6. Conclusion”.
Author Response
Comments 1: In the introduction, the authors mention the motivation for the study: “Therefore, it is imperative to investigate the underlying mechanisms, particularly…”. It is recommended that the underlying mechanisms be recapitulated to describe them more specifically. Also, the sudden appearance of transaction costs was a bit abrupt. I hope the author explains why transaction costs are used as an example.
Response 1: Thank you for your feedback. First, as outlined in lines 41-53 of the manuscript, we discuss how the adoption of AI agents is influenced not only by their tangible benefits but also by subjective perceptions, especially the emotional and affective aspects of consumer interactions with AI-based service technologies. This leads us to explore the role of emotional perception, particularly emotional support, as a key underlying mechanism affecting individual behavior.
Secondly, we introduce the concept of transaction costs in footnote 1 on line 31, noting that while AI reduces labor costs, its impact on transaction costs is less clear. In our study, we specifically focus on the supplier-side transaction costs incurred during interactions with different product recommenders. This contextualizes the use of transaction costs as a pertinent example in our analysis. Specifically, drawing from transaction cost theory, transaction cost refers to the expenses involved when buyers (customers) and sellers (retailers) engage in trade (Celmons & Kimbrough, 1986; Devaraj et al., 2002). These costs encompass a variety of activities including the search for suitable trading partners, the acquisition of product information and pricing, the drafting of contracts, the actual purchasing process, and the enforcement of these contracts. While AI has replaced human roles and reduced labor costs for businesses, its impact on transaction costs, such as return rates due to product recommendations, remains unexplored. In our setting, we concentrate on the supplier-side transaction costs incurred during interactions with different product recommenders.
Comments 2: The second and third paragraphs of the introduction do not follow closely together, and the author is expected to add a transitional statement before "self-disclosure".
Response 2: Thank you for your feedback. To ensure coherence between the paragraphs, we transition from real-world corporate scenarios to the concept of self-disclosure. Please refer to lines 57-61, where we illustrate how companies often disclose AI’s capabilities and past performances (e.g., Google's BERT and IBM’s Watson), including its proficiency in textual analysis and processing accuracy, to build trust among users and enhance their overall experience. Similarly, akin to how humans foster relationships through self-disclosure, previous research highlights that AI self-disclosure is crucial for deepening these interactions [20, 21].
Comments 3: The definition of transaction costs should be further clarified in the introduction, e.g., are they costs for sellers or buyers?
Response 3: Thank you for your comments. We have revised the content and provided a comprehensive definition of transaction costs in footnote 1 on line 31.In our setting, we concentrate on the supplier-side transaction costs incurred during interactions with different product recommenders . Specifically，Drawing from transaction cost theory, transaction cost refers to the expenses involved when buyers (customers) and sellers (retailers) engage in trade (Celmons & Kimbrough, 1986; Devaraj et al., 2002). These costs encompass a variety of activities including the search for suitable trading partners, the acquisition of product information and pricing, the drafting of contracts, the actual purchasing process, and the enforcement of these contracts. While AI has replaced human roles and reduced labor costs for businesses, its impact on transaction costs, such as return rates due to product recommendations, remains unexplored.
Comments 4: In the literature review, the authors mentioned: “AI technologies offer transformative potentials to optimize resource allocation and increase operational efficiencies”. The authors fully recognize the important importance of technologies such as AI and recommendation algorithms to enhance the transactional efficiency of online commerce. Authors could focus more on the role of recommendation affordance in facilitating user experience, e.g., the role of recommendation affordance in social e-commerce can be found in: Attention marketing in fragmented entertainment: how advertising embedding influences purchase decision in short-form video apps.
Response 4: Thank you for your suggestion regarding the role of recommendation affordance and advertisement presentation methods in influencing user purchase intentions. While this is an important area of research, our current study specifically focuses on comparing the impact of different levels of self-disclosure on transaction costs between AI and human recommenders. This narrower focus allows us to deeply explore and isolate the effects of self-disclosure in the context of AI versus human interactions within transaction processes. However, your suggestion is indeed valuable and could be considered a promising direction for future research that builds on our findings, exploring how these dynamics operate in conjunction with advertising strategies in digital commerce.
Comments 5: The way p=0.000 is written in Figure 3 is not rigorous.
Response 5: Thank you for pointing out the issue with the presentation of statistical significance in our figure. We have addressed this by modifying the figures to reflect the more precise notation of p<0.01. This change ensures that our representation of statistical results is both rigorous and standardized across our figures.
Comments 6 : Part 5 is suggested to be split into “5. Discussion” and “6. Conclusion”.
Response 6: Thank you for your suggestion to organize the manuscript more effectively. In response, we have revised the manuscript by dividing Part 5 into two distinct sections: "Discussion" and "Conclusion."
In the "Discussion" section, lines 714-741, we present real-world business scenarios and our research findings, providing a thorough analysis of the implications of these results within the context of current industry practices.
In the "Conclusion" section, lines 747-782, we articulate three main practical implications derived from our findings and discuss their relevance to sustainability. We also address the limitations of our study and propose directions for future research. This structure enhances the clarity of our manuscript and better communicates the contributions and conclusions of our work.
Reviewer 4 Report
Comments and Suggestions for AuthorsThis paper explores the impact of human or artificial intelligence (AI) recommendations on transaction costs in product recommendation systems and analyzes the role of self-disclosure levels. The study found that AI recommendations lead to higher transaction costs compared to human recommendations. However, when the level of self-disclosure was low, there was no significant difference in transaction costs between AI recommendations and human recommendations. The article points out that in AI recommendation systems, even adding humanized labels can significantly reduce transaction costs, and stresses the importance of self-disclosure levels. This makes sense, but the following questions still need to be answered:
1) Although social exchange theory is mentioned in line 72 of the paper, its relationship with AI recommendation system is not explained in detail. It is suggested that the introduction of social exchange theory should include a detailed description of its specific application and impact in AI recommendation systems.
2) In Section 3.6 of the article, the operational definition of emotional support needs to be explained in more detail to ensure that the reader understands exactly how the variable is measured and manipulated in the experiment, while further enabling the reader to fully understand the experiment process.
3) In lines 463-465, the discussion and analysis of the results are not deep enough. There is no explanation for the insignificance of simple effects. It is recommended to add a discussion of this result, providing possible explanations and suggestions for further research.
4) The conclusion part should further detail the application and significance of the research results in practice, especially the specific recommendations for enterprises how to design and use AI recommendation systems. It is suggested to add relevant content as appropriate.
Author Response
Comments 1: Although social exchange theory is mentioned in line 72 of the paper, its relationship with AI recommendation system is not explained in detail. It is suggested that the introduction of social exchange theory should include a detailed description of its specific application and impact in AI recommendation systems.
Response 1: Thank you for your feedback regarding the connection between social exchange theory and AI recommendation systems. In response to your suggestion, we have expanded the explanation in lines 123-133 of the manuscript. Specifically, to address this research question, the social exchange theory provides a robust framework for identifying consumers' perceived feelings and subsequent behaviors. This theory suggests that relationships are established through a sequence of exchanges characterized by self-interest and interdependence [26, 27]. Social exchange theory pro-vides a framework for analyzing the distinct effects of artificial intelligence and human interactions on transaction costs by examining how consumers evaluate the costs and benefits associated with emotional engagement in both AI and human contexts. Com-pared with AI, humans are capable of offering profound emotional support, such as empathy and compassion, which is considered a high-value benefit that consequently reduces transaction costs. Additionally, this theory posits self-disclosure as a cognitive process that involves assessing rewards and costs, where the perceived value of disclosure is balanced against its potential risks [27].
Comments 2: In Section 3.6 of the article, the operational definition of emotional support needs to be explained in more detail to ensure that the reader understands exactly how the variable is measured and manipulated in the experiment, while further enabling the reader to fully understand the experiment process.
Response 2: Thank you for your comment regarding the need for a more detailed explanation of the operational definition of emotional support in our study. Please refer to lines 531-538 for a comprehensive description.
Specifically, emotional support refers to the provision of care, concern, empathy, love, and trust [92]. In line with the methodologies of Kessler et al. [93] and Lakey and Cassady [94], we utilized a 7-point scale to assess the extent to which participants believed their sales manager provided emotional support. In post experimental questionnaire, participants responded to the question, "To what extent do you believe your sales manager can provide you with emotional support?" The scale ranged from 1 (extremely unlikely) to 7 (extremely likely), enabling us to gauge the perceived emotional support under different experimental conditions.
Comments 3: In lines 463-465, the discussion and analysis of the results are not deep enough. There is no explanation for the insignificance of simple effects. It is recommended to add a discussion of this result, providing possible explanations and suggestions for further research.
Response 3: Thank you for your feedback regarding the depth of the discussion and analysis in lines 463-465. To address this, we have added a more detailed discussion of the results concerning the insignificance of simple effects. Specifically, in the contexts of high self-disclosure, both humans and artificial intelligence (AI) exhibit significant understanding and analytical capabilities. These capabilities prompt users to set high expectations, regardless of whether they are interacting with a person or an AI system. However, when product recommendations fail to meet these expectations, the resulting disappointment is substantial due to the significant cognitive resources invested, such as high expectations. That is, whether the recommender is human or artificial in-telligence, consumer return intentions are primarily driven by the negative emotions caused by the discrepancies between expectations and actual delivery. Thus, the trans-action costs will not differ between between AI and human.
Comments 4: The conclusion part should further detail the application and significance of the research results in practice, especially the specific recommendations for enterprises how to design and use AI recommendation systems. It is suggested to add relevant content as appropriate.
Response 4: Thank you for your suggestion regarding the need to detail the practical applications and significance of our research results, especially concerning the design and use of AI recommendation systems by enterprises. Please refer to lines 749-770 where we elaborate on this aspect. In this section, we discuss how our findings have significant implications for the deployment of artificial intelligence in product recommendation scenarios. Specifically, the research findings from this study offer significant practical implications for the deployment of artificial intelligence (AI) in product recommendation scenarios. Our re-sults show that even when the recommender system is AI-based, affixing a human label (e.g., naming it "Jonny Zhang") can substantially reduce transaction costs. This strategy of humanizing AI not only enhances the emotional support perceived by users but also contributes to more sustainable business practices by improving efficiency and reducing the resource drain typically associated with higher transaction costs. Moreover, our findings indicate that low self-disclosure—focusing solely on the functionality of the product recommendations rather than the recommender's personal accolades—further amplifies the positive impact of perceived human attributes on reducing transaction costs. This highlights the importance of how information is presented, emphasizing functional aspects over personal achievements or histories can lead to more effective, cost-efficient outcomes. This approach also leverages AI's ability to provide concise, functional in-formation without overloading the user with unnecessary details, which can lead to more streamlined and efficient consumer decisions. Such efficiency is crucial for sustainability as it reduces the waste of resources, both in terms of the cognitive load on consumers and the operational overhead for businesses. In essence, by focusing on the functional benefits of products rather than the intricacies of the AI itself, companies can foster a more sustainable interaction model that conserves resources while still meeting consumer needs effectively. That is, utilizing AI in emotionally driven interactions effectively manages user expec-tations and minimizes resource waste associated with transaction costs, thereby sup-porting sustainable practices that positively impact environmental and social outcomes.
Again, we sincerely appreciate the insightful comments and constructive suggestions provided by you. Your feedback has been instrumental in enhancing the quality and depth of our manuscript. We have carefully considered and incorporated your recommendations, which have significantly improved our study. Thank you for your valuable contributions to our research.
Round 2
Reviewer 2 Report
Comments and Suggestions for Authorsdear authors
you should reduce the number of refrences why in every citation you add from two to three refrences in the same paragraph only choose the most approiate one
please update all figures resolutions and tables as the journal format
please check the journal format and update your paper
thanks
Comments on the Quality of English Languageneed revision
Author Response
Comments 1: you should reduce the number of refrences why in every citation you add from two to three refrences in the same paragraph only choose the most approiate one.
Response 1: Thank you for your valuable feedback regarding the use of references in our manuscript. We appreciate your suggestion to streamline the citations within each paragraph.
In response to your comment, we have carefully reviewed the manuscript and made necessary adjustments by selecting only the most pertinent references for each citation. Additionally, while we have streamlined our references throughout the manuscript, we maintained dual citations in certain sections where multiple sources are essential to substantiate our contributions and the nuanced arguments presented (particularly noted in lines 135-142 of the contribution section). These instances were carefully considered to ensure they enrich the discussion and underscore the depth of our theoretical contributions.
This revision aimed to enhance clarity and ensure that each reference is directly relevant and supportive of our assertions. We believe these modifications have improved the manuscript’s readability and academic rigor, aligning with the journal's standards for citation practices.
Comments 2: please update all figures resolutions and tables as the journal format
Response 2: Thank you for your guidance on formatting requirements. We have updated all tables and figures to comply with the journal's specifications for resolution, ensuring each image is at least 1000 pixels in width/height in JPEG format. These updates are included in the revised manuscript, with specific changes made to tables 1-6 at lines 182, 319, 439, 456, 479, and 514. Additionally, figures 1-4 have been adjusted accordingly and can be found at lines 341, 457, 481, and 515. We have submitted these revisions in a compressed file format directly to the editor for review.
Comments 3: please check the journal format and update your paper
Response 3: Thank you for the reminder. We have carefully reviewed and adhered to the journal's template requirements and have made the necessary adjustments according to the guidelines provided at MDPI's formatting instructions. We have ensured that all formatting complies with the specified standards.
We appreciate the constructive feedback provided in the reviewers' comments, which have significantly helped in refining our manuscript. This process has not only strengthened the presentation of our research but also enhanced the overall rigor and impact of our findings, making them more accessible and relevant to the field. Please let us know if further modifications are required.
