Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 5;16:12112. doi: 10.1038/s41598-026-42465-4 Search in PMC Search in PubMed View in NLM Catalog Add to search Users’ prompting strategies and ChatGPT’s contextual adaptation shape conversational information-seeking experiences Haoning Xue Haoning Xue 1 Department of Communication, University of Utah, Utah, USA 5 Department of Communication, 2511 LNCO, University of Utah, Salt Lake City, UT 84111 USA Find articles by Haoning Xue 1, 5, ✉ , Yoo Jung Oh Yoo Jung Oh 2 Department of Communication, Michigan State University, Michigan, USA Find articles by Yoo Jung Oh 2 , Xinyi Zhou Xinyi Zhou 3 Department of Computer Science, Boise State University, Idaho, USA Find articles by Xinyi Zhou 3 , Xinyu Zhang Xinyu Zhang 2 Department of Communication, Michigan State University, Michigan, USA Find articles by Xinyu Zhang 2 , Berit Oxley Berit Oxley 4 Department of Behavioral, Social, and Health Education Sciences, Emory University, Georgia, USA Find articles by Berit Oxley 4 Author information Article notes Copyright and License information 1 Department of Communication, University of Utah, Utah, USA 2 Department of Communication, Michigan State University, Michigan, USA 3 Department of Computer Science, Boise State University, Idaho, USA 4 Department of Behavioral, Social, and Health Education Sciences, Emory University, Georgia, USA 5 Department of Communication, 2511 LNCO, University of Utah, Salt Lake City, UT 84111 USA ✉ Corresponding author. Received 2025 Oct 1; Accepted 2026 Feb 25; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13076770 PMID: 41786987 Abstract Conversational AI, such as ChatGPT, is increasingly used for information seeking. However, little is known about how ordinary users actually prompt and how ChatGPT adapts its responses in real-world conversational information seeking (CIS). In this study, a nationally representative sample of 937 U.S. adults engaged in multi-turn CIS with ChatGPT on both controversial and non-controversial topics across science, health, and policy contexts. We analyzed both users’ prompting strategies and the communication styles of ChatGPT’s responses. The findings revealed behavioral signals of digital divide: only 19.1% of users employed prompting strategies, and these users were disproportionately more educated and Democrat-leaning. Further, ChatGPT demonstrated contextual adaptation: responses to controversial topics contain more cognitive complexity and more external references than to non-controversial topics. Notably, cognitively complex responses were perceived as less favorable but produced more positive issue-relevant attitudes. This study highlights disparities in user prompting behaviors and shows how user prompts and AI responses together shape information-seeking with conversational AI. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-42465-4. Keywords: Conversational AI, Digital Divide, Information-seeking, Prompting Strategies Subject terms: Psychology; Psychology; Science, technology and society Introduction Conversational information-seeking (CIS) refers to the process of obtaining information through dialog-based systems 1 . Many CIS systems, such as ChatGPT, Claude, and Gemini, have been empowered by Large Language Models (LLMs) and have become major information sources for critical societal issues 2 . Among these conversational AI systems, ChatGPT remains the most widely used AI tool, potentially affecting millions of users 3 . A recent Pew Research report highlights that Americans are increasingly turning to ChatGPT for new information 4 . Users increasingly use and trust such systems for information on important societal issues across health, science, and public policies, even more than traditional search engines like Google 5 , because conversational AI is often perceived as a nonpartisan, objective information source (e.g., algorithm appreciation 6 and machine heuristics 7 . This growing reliance on conversational AI raises questions about the communication dynamics between users and AI systems. On the user side, how users actually prompt represents behavioral signals of AI literacy in action and the second-level digital divide in interacting with conversational AI, beyond self-reported attitudes and behavioral intentions 8 . The first-level digital divide revolves around technology access and adoption, while the second-level digital divide 8 zooms in on disparities in users’ skills in using technologies. Existing human-centered studies have mostly examined broader patterns of first-level digital divide in AI perception, adoption, and trust, showing how disparities in adopting conversational AI exist across income 9 and education levels 10 . Further, factors such as information sufficiency 11 , risk perceptions 12 , and social support 13 drive information-seeking intentions with conversational AI as well. Research shows that prompting strategies such as asking AI to adopt a persona 14 and requesting step-by-step reasoning 15 can impact the accuracy, quality, and style of AI responses. Yet little is known about how ordinary users actually use these strategies in real-world information seeking. Prompting as a behavioral signal of AI literacy in action and the second-level digital divide has been largely overlooked. On the AI system side, how AI responds is equally important. Existing research on CIS has primarily focused on content-related issues 16 , such as accuracy, bias, and hallucination. But we argue that communication styles are central to human-AI interactions 17 , particularly in CIS contexts, where the interaction is dynamic, personalized, and iterative. In these dyadic information-seeking exchanges, AI responses’ tone, structure, and style are critical in shaping user experiences, influencing attitude formation and change, and facilitating the exchange of ideas and factual information 18 . This is more pressing for controversial topics such as vaccination, climate change, and immigration, which are often polarized and emotionally charged. Existing work has been focusing on content-level adaptation, finding that ChatGPT’s responses on controversial issues are often more comprehensive 19 and may reinforce users’ confirmation biases 20 . However, the extent of such contextual adaptation in communication styles and its downstream effects on user perceptions remains unclear. Taken together, both how users prompt and how AI responds in CIS remain understudied. This study zooms in on ChatGPT, the most widely used conversational AI 3 and addresses these research gaps by examining three interconnected levels of human-AI interactions in CIS: (1) how users adopt prompting strategies, (2) how ChatGPT’s communication styles adapt to user prompts and context controversy, and (3) how ChatGPT’s communication styles shape users’ AI perceptions and context-relevant attitudes. We focus on communication styles because they represent key message features that are distinct from message content, according to persuasion frameworks such as the Elaboration Likelihood Model 21 . The way information is framed and presented can substantially influence how individuals assess the credibility of information sources and how their attitudes may shift. Given the interactive nature of CIS, we also examine ChatGPT’s adaptive behaviors: how its communication styles shift in response to different users and conversational contexts. Prompting strategies in conversational information-seeking Prompting strategies refer to user inputs with specific instructions or structures that can optimize AI responses 22 . While traditional information-seeking research focuses on search tactics and queries in online databases and search engines 23 , user prompts are natural language inquiries in dialogues. Examples include asking AI to adopt a persona 14 , requesting step-by-step reasoning 15 , or specifying response structures and styles 24 . Much of the research on prompting engineering has focused on the AI system side, testing how prompt engineering may improve LLM performance. However, how ordinary users actually adopt these prompting strategies in everyday information-seeking remains largely unclear. A recent study on Google Bard found that fewer than 6% of real-world prompts showed evidence of prompt engineering 25 ; these advanced strategies were rarely used in practice. Therefore, we seek to understand the prevalence of 8 prompting strategies (e.g., adopt a persona, step-by-step reasoning) in CIS, according to ChatGPT prompting guidelines 26 . RQ1 How do users use prompting strategies for information seeking with ChatGPT? Beyond this gap between prompting engineering research and how users actually prompt, another question concerns who actually adopts prompting strategies. The digital divide exists not only in AI access and usage but also in AI skills 9 , 10 . Prior research shows that individuals with higher family socioeconomic status 27 tend to have higher AI literacy, are more likely to use AI for work and education, and use more abstract language in AI conversations 28 . In contrast, less educated users reported having worse experiences with ChatGPT in seeking information on controversial issues such as climate change 29 . Building on prior findings of structural inequalities in accessing and using AI, we zoom in on the socio-demographic disparities in using prompting strategies, as prompts provide behavioral signals of AI literacy in action beyond self-evaluations and perceptions 30 . RQ2 How are users’ prompting strategies associated with socio-demographics? Communication styles adaptation in conversational information-seeking Communication styles are central to how people shape perceptions and form attitudes. Affinity-seeking language increases social likability in conversations 31 ; intensive and extreme language enhances persuasion 32 . Beyond static linguistic features, the dynamic exchange of communication styles is equally important 33 . Reciprocity of communication styles improves interpersonal relationship satisfaction 34 ; in online support groups, the alignment of positive and negative sentiments increases future support-seeking behaviors 35 . Conversational AI exhibits similar adaptive behaviors by mirroring linguistic cues and tones of user inputs 36 . But compared with human conversations, conversational AI’s adaptation is often asymmetrical, with AI accommodating users more 37 . Such adaptation can improve conversation flow and satisfaction 38 , but it can introduce risks when conversational AI overly adapts, such as excessively agreeing with users (a phenomenon known as “sycophancy” 39 ) or treating health-related claims as true by default 40 , which is parallel to humans’ truth-default bias 41 . This line of work confirms conversational AI’s adaptability, but how such adaptation applies to users’ prompting strategies is unclear. This adaptation extends to issue contexts and stances, with both similar (i.e., reciprocity) and dissimilar (i.e., compensation) styles 42 . Overall, ChatGPT’s responses on controversial issues are often more comprehensive 19 and liberal-leaning 43 . When discussing controversial issues, ChatGPT compensates negatively-toned prompts with responses in a neutral or positive tone 44 . However, contradictory evidence exists that ChatGPT exhibits more hostility and higher emotional intensity in discussing controversial issues 45 . The mixed evidence raises questions about how conversational AI adapts to issue controversy consistently in natural conversations with users. RQ3 How do (a) issue controversy and (b) users’ prompting strategies influence ChatGPT’s responses in information-seeking interactions? Research shows that conversational AI’s ability to adapt and customize communication styles influences how users perceive human-AI interaction and shapes attitude and behavior change 46 . For example, conversational AI matching users’ linguistic styles has been shown to increase donations to charities 47 . Elaborateness and politeness in Amazon Alexa’s responses improve user satisfaction 48 . Conversational AI’s adaptation to users’ communication styles improves the overall conversation flow, understanding, and satisfaction 38 , 49 . Although empirical evidence reveals the persuasiveness of conversational AI in persuasion contexts, it is unclear how AI responses, especially communication styles, may influence users’ AI perceptions and issue-specific attitudes. Therefore, RQ4 How do ChatGPT’s responses influence users’ (a) AI perceptions and (b) issue-specific attitudes? Furthermore, while prior work primarily examines generic linguistic features such as functional words and sentiment, we argue that focusing on specific communication styles more relevant to CIS interactions provides deeper and more meaningful insights. Therefore, in this study, we emphasize five information-related communication styles 50 , 51 . These communication styles go beyond the content exchanged; instead, these styles reflect how information is presented in both user prompts and AI responses. It is important to note that while we use information-seeking to refer to the broader context of users seeking information from conversational AI, information-seeking below captures question-asking language styles. Self-revealing sharing one’s own experience and opinion. Information-seeking asking questions or seeking advice. Fact-oriented discussing or providing factual information. Action-seeking calling for action or giving actionable advice. Cognitive complexity expressing or explaining complexly organized concepts. To answer these questions, we conducted a 3 (issue: health, science, policy) × 2 (issue controversy: controversial, non-controversial) between-subject experiment. A nationally representative sample of 937 participants engaged in a conversation with ChatGPT to seek issue-specific information. We highlight three main findings in this study. First, 19.1% of users have used at least one prompting strategy, predicted by education level and party affiliation. Second, issue controversy drives ChatGPT responses’ communication styles (e.g., action-seeking, cognitive complexity) and the number of citations, while the communication styles of user prompts remain consistent across issue controversy. Lastly, ChatGPT responses’ cognitive complexity influences users’ AI perceptions; users perceived more cognitively complex AI responses as less favorable but experienced more positive issue-relevant attitude changes with cognitively complex AI responses, suggesting the implicit influence of cognitive complexity on attitude changes. Methods Experiment design and procedure This between-subject experiment has a 3 (issue: health, science, policy) × 2 (issue controversy: controversial, non-controversial) (see Fig. 1 for experimental procedure). A nationally representative sample of U.S. adults ( n = 937) was recruited on Prolific in November 2024. All study procedures were conducted in accordance with relevant guidelines and were approved by the University of Utah’s Institutional Review Board (approval number: 00183541). Informed consent was obtained from all participants prior to participation in the study. Fig. 1. Open in a new tab Flow diagram of experiment procedure. Upon providing consent, participants reported demographics and pre-existing attitudes on AI and 6 issues. Afterward, participants were randomized to one of the six topics across issue controversy and topics in Fig. 1 . Participants were instructed to seek issue-relevant information from ChatGPT for a hypothetical social scenario. Participants were required to interact with GPT-4o for at least 5 turns before submitting the shareable conversation link for compensation, which is meant to ensure sufficient engagement and meaningful interactions 29 . See Fig. 2 below for an example instruction; see Table S1 in the Supplemental Information for all instructions. Lastly, participants reported their perceptions of AI responses, interaction experience, and post-experiment issue-specific attitudes. Among the 937 participants, 473 were female (50.8%), with a mean age of 45.1 years, 70.0% white, Democrat-leaning (55.7%), and median annual household income between $50,000 and $74,999. See Table S2 for demographics summary. The median participation time was 18 min. Fig. 2. Open in a new tab Illustration of example information-seeking instruction and interaction. Manipulation check To ensure the manipulation of issue controversy is successful, participants rated how controversial the six topics are on a 5-point scale. Participants perceived immigration (M = 4.36, SD = 0.83) as the most controversial and highway infrastructure as the least (M = 1.96, SD = 0.96). On average, three controversial topics (M = 3.78, SD = 0.73) were perceived as more controversial than non-controversial topics (M = 2.29, SD = 0.76; t = 29.81; p < .001). Conversation data collection We retrieved 747 valid URLs from 937 responses. We excluded n = 137 inaccessible links (e.g., broken chat links, chat links with sharing turned off) and n = 53 irrelevant conversations (i.e., chats irrelevant to the study context) from further analysis. We scraped all conversations and related metadata, including conversation titles, the ChatGPT models used, and external links included in ChatGPT’s responses. On average, there were 6 turns per conversation (M = 5.56, SD = 1.98, Min = 2, Max = 44, Median = 5). Most chats ( n = 722, 96.7%) were conducted with GPT-4o as instructed, while 24 chats (3.2%) used GPT-4o-mini and 1 chat (0.1%) used GPT-4. We included relevant conversations that did not strictly follow instructions, as they still reflect natural CIS with ChatGPT. Communication style extraction We used the Symanto Psychology API 50 and Linguistic Inquiry and Word Count (LIWC) 2022 52 to extract 5 communication styles related to information seeking. See Table 1 below for definitions, summary statistics, and examples. For every conversation, we aggregated user prompts and ChatGPT responses across conversation turns and calculated communication styles, respectively. This is meant to capture an overall interaction pattern while reducing noise arising from turn-to-turn variability. First, we used the Symanto Psychology API 50 , which is based on a fine-tuned BERT model, to extract communication styles related to self-revealing, information-seeking, fact-orientation, and action-seeking. Each of the four communication styles is assigned a probability score between 0 and 1, with higher values indicating a stronger presence in the text. Second, we used LIWC to calculate the categorical-dynamic index (CDI) to capture cognitive complexity 51 . CDI reflects the extent to which a text expresses complexly organized concepts, which can be inferred with functional words. We normalized CDI for comparison across communication styles, with higher values indicating more cognitive complexity. Table 1. Definitions, summary statistics, and examples of communication styles. Communication style Author M SD Example Self-revealing: Speakers share personal information or experiences. User 0.39 0.46 I love stevia and use it primarily but I need some convincing information to tell others how awful artificial sweeteners are. ChatGPT 0.06 0.23 That’s the spirit! I’m sure they’ll appreciate you taking the time to help them feel informed. Let me know how it goes or if any other questions come up along the way! Information-seeking: Direct or indirect questions searching for information. User 0.97 0.17 Are all the new boosters out now? When should they receive them for best holiday immunity? ChatGPT 0.07 0.24 I’d be happy to help with any questions you have about climate change. What would you like to know? Fact-oriented: Factual and objective statements. User 0.11 0.30 Why do you defend illegal immigration? The act of crossing the border illegally is a crime in and of itself. ChatGPT 1.00 0.04 The side effects of the COVID-19 vaccines are generally mild and temporary. Here’s a breakdown of the most common ones. Action-seeking: Direct or indirect requests, suggestions, and recommendations expecting action from other people. User 0.10 0.29 Can you put yourself into the shoes of someone against immigration and tell me your points and try not to bias what I’ve said at all ChatGPT 0.14 0.34 1. That’s a great initiative! Here are some ways to make it engaging and practical for your neighbors: 1. **Host an Informative Event or Workshop** 2. Investing in a proper filtration system is the best way to ensure you’re effectively reducing microplastic exposure. Would you like more details on any of these methods? Cognitive complexity: Expressing or explaining complexly organized concepts User 0.56 0.15 Summarize the impact of immigration on the US, Include topics like national security and economy. ChatGPT 0.59 0.10 Public opinion on immigration in the United States is notably divided, often along partisan lines. Here are some key statistics illustrating these divisions. Open in a new tab Factual and structural feature extraction In addition to communication styles, we extracted two distinctive features capturing the factual and structural dimensions of ChatGPT responses. Citation count Whether solicited or not, the number of citations provided in AI responses signals credibility and trustworthiness 53 . We extracted the number of external links in every ChatGPT response and then averaged these counts for each conversation, accounting for differences in conversation duration. This yields an average citation count representing the factual basis of AI responses (Min = 0, Max = 12, M = 0.62, SD = 1.66). Structure count Structure count is used to represent the readability and clarity of ChatGPT responses, as structured elements such as headings and bullet points can enhance comprehension and facilitate learning 54 . Structure count (Min = 0, Max = 104.4, M = 21.89, SD = 14.23) is the sum of the average number of structural elements in ChatGPT responses for every conversation, including the number of headings (Min = 0, Max = 12, M = 2.55, SD = 2.36), bullet points (Min = 0, Max = 32.4, M = 6.78, SD = 5.65), numbered points (Min = 0, Max = 12, M = 1.92, SD = 2.00), and bolded texts (Min = 0, Max = 64.8, M = 10.63, SD = 7.24). Prompting strategies annotation Referring to ChatGPT prompting guidelines 26 , we identified 8 prompting strategies and grouped them into three broad categories based on whether users provided or requested information: user-supplied information, style-related requests, and content-related requests (see Table 2 below for details). To systematically identify prompting strategies, we used ChatGPT to annotate all user prompts ( n = 4,154; see Table S3 for the annotation instructions). We chose this approach to capture how ChatGPT perceives and interprets the prompting strategies that users employ, which may directly influence how ChatGPT formulates its responses. As an additional validity check, three trained coders independently validated a random subset of n = 884 (21.3%) user prompts, achieving high inter-coder reliability (IRR = 0.814). Further, we calculated the weighted F1 score to account for data imbalance. The weighted F1 score of 0.903 shows that ChatGPT’s annotation has a high accuracy and a high consistency with human interpretation. Table 2. Summary statistics of prompting strategies with example prompts. Prompt category Prompt strategy #users (%) #prompts (%) Example User-supplied information Provide delimiter 3 (0.4%) 3 (0.1%) What are some specific solutions that involve policy change rather than “““reusable straws”"” ideas. Provide example 3 (0.4%) 4 (0.1%) I mean… like… something we all do but accept the risk where the risk is much higher… like, oh, driving to Kroger, for example. What else is very common but with a much higher statistical risk than getting the covid shot? Provide context 75 (10.0%) 87 (2.1%) My friends have been discussing the health effects of artificial sweeteners. I want to appear informed. I want to help my neighbors understand the safety and potential health risks associated with artificial sweeteners. Sum 78 (10.4%) 94 (2.3%) Style-related requests Specify length 22 (2.9%) 23 (0.6%) Summarize in 100 words or fewer how to convince someone that the new Covid-19 booster is worth getting. Specify style 46 (6.2%) 54 (1.3%) Could you answer that question in a more briefly and in a more conversational tone? Sum 67 (9.0%) 77 (1.9%) Content-related requests Specify persona 5 (0.7%) 8 (0.2%) I want you to act like an economics expert and tell me the pros and cons of the current US immigration policy. Request references 54 (7.2%) 68 (1.6%) Please cite sources for medical information on artificial sweeteners. Request stepwise reasoning 9 (1.2%) 11 (0.3%) Can you break this down into steps we should take to reverse global warming? Sum 67 (9.0%) 87 (2.1%) Open in a new tab Self-reported measures Tables S4-S8 list question items for measures below. See Table S9 for summary statistics of key variables by issue topic. Demographics We asked participants to report their age, gender, race, education, income, and political affiliation. Gender and race were converted to binary variables (gender: 1 for female, 0 for male; race: 1 for white, 0 for non-white). Political affiliation was measured by one question on participants’ political party affiliation on a 7-point scale (1: a strong Democrat, 7: a strong Republican) ( M = 3.29, SD = 2.02). Issue-specific pre-existing attitude and knowledge Participants’ pre-existing attitudes were measured with one question per issue on a 4-point scale, such as How effective do you think the COVID-19 vaccines are? Further, participants’ pre-existing knowledge was measured with one question: How would you rate your level of knowledge about the following topics? AI familiarity Users’ familiarity with AI technologies 55 was measured by one question: Which of the following technologies , if any , uses artificial intelligence (AI)? Responses included 12 AI products (e.g., Google search, chatbots). The number of AI products identified indicates familiarity with AI (Min = 0, Max = 12, M = 8.55, SD = 3.08). Issue-specific attitude (post-experiment) After interacting with ChatGPT, participants reported their issue-specific attitudes again for the assigned issue, using four questions per issue on a 5-point scale. Specifically, these questions are about COVID-19 vaccine safety (M = 3.91, SD = 1.15, Cronbach’s α = 0.93), climate change severity (M = 4.07, SD = 0.96, Cronbach’s α = 0.90), immigration benefits (M = 3.59, SD = 1.08, Cronbach’s α = 0.93), artificial sweetener safety (M = 2.70, SD = 1.03, Cronbach’s α = 0.88), microplastics severity (M = 4.04, SD = 0.84, Cronbach’s α = 0.92), and highway reconstruction benefits (M = 3.89, SD = 0.64, Cronbach’s α = 0.60). Perceived AI response quality Participants rated AI response quality with 12 items on a 7-point scale 56 , such as generic–in-depth , and clear–ambiguous (M = 5.72, SD = 0.89, Cronbach’s α = 0.89). Perceived AI interaction quality Participants rated the quality of the information-seeking interaction with AI with 7 items on a 7-point scale 56 , such as In this conversation , ChatGPT was able to understand my questions and instructions clearly (M = 4.29, SD = 0.55, Cronbach’s α = 0.73). AI perceptions: perceived likability , trustworthiness , and intelligence Participants evaluated AI with a source evaluation measure on a 7-point scale 57 , including perceived likability (M = 5.70, SD = 1.18, Cronbach’s α = 0.90), trustworthiness (M = 5.56, SD = 1.11, Cronbach’s α = 0.92), and intelligence (M = 6.03, SD = 1.06, Cronbach’s α = 0.90). Analysis plan To answer RQ1 about the prompting strategies used, we provide descriptive and summary statistics of the prompting strategies used. To answer RQ2 about the relationship between socio-demographics and users’ prompting strategy usage, we conduct a negative binomial regression with the number of prompting strategies used as the dependent variable. Negative binomial models are appropriate for the overdispersed prompting strategy count data (dispersion = 1.40, p < .001). We include gender, age, race, education, income, and political affiliation as independent variables with five covariates (i.e., issue controversy, issue topic, pre-existing attitude, pre-existing knowledge, and AI familiarity) controlled. To answer RQ3 about the influence of issue controversy and users’ prompting strategies on ChatGPT responses, we conduct linear regression models with the citation count, structure count, and five communication styles, respectively. Issue controversy and prompting strategy count (pure count and by three categories in separate models) are included as independent variables. Further, issue topic, users’ pre-existing attitude, user prompts’ five communication styles, and word counts of user prompts and ChatGPT responses are controlled as covariates. To answer RQ4 about the effects of ChatGPT responses on AI perceptions and issue-specific attitudes, we run similar linear regression models with AI perceptions and issue-specific attitudes as dependent variables, respectively. We include ChatGPT responses’ citation count, structure count, and five communication styles as independent variables. Users’ demographics, pre-existing attitudes, AI familiarity, and issue controversy are controlled as covariates. Results Prompting strategies are rarely used, mostly by educated and democrat-leaning users RQ1 asked about how users adopt prompting strategies in CIS. We found that prompting strategies were rarely used in information seeking (see Table 2 above for descriptive statistics). Across all conversations, 19.1% users ( n = 179) used at least one prompting strategy, resulting in n = 258 prompts (6.2% of all user prompts). Most users employed straightforward, plain questions rather than strategically engineered prompts in information-seeking with ChatGPT. Among prompts with any prompting strategies, the most commonly used strategies were providing contextual information ( n = 87, 2.1%), requesting external references ( n = 68, 1.6%), and specifying ChatGPT’s response styles ( n = 54, 1.3%). In contrast, the least frequently used strategies were providing delimiters ( n = 3, 0.1%), providing examples ( n = 4, 0.1%), and requesting ChatGPT to adopt a persona ( n = 8, 0.2%). It is worth noting that the high frequency of providing contextual information might be inflated by the experimental setting of information seeking for a hypothetical neighborhood discussion, as users tended to repeat back to ChatGPT as additional contextual information. Similarly, while providing examples is common for task-oriented prompting (e.g., few-shot prompting), it is less relevant in CIS. For RQ2 about the association between socio-demographics and the usage of prompting strategies, we found that education and political affiliation were significant predictors (see Fig. 3 below for visualization; see Table S10 for full regression models). More educated and Democrat-leaning users employed more prompting strategies, especially content-related requests. Further, users with higher levels of AI familiarity used more style-related requests, suggesting that users with more AI knowledge are more likely to fine-tune the styles of ChatGPT responses. Females were less likely to use content-related requests, while older users were less likely to prompt with user-supplied information. In addition to socio-demographics, users’ pre-existing knowledge negatively predicted prompting strategy usage. It is consistent with the notion that more knowledgeable users are less motivated to seek information. Overall, prompting strategies were not evenly adopted across the population, but skewed toward the educated and Democrat-leaning users. Fig. 3. Open in a new tab Regression coefficients and confidence intervals predicting the count of ( a ) prompting strategy and three prompting strategy categories, respectively: ( b ) user-supplied information, ( c ) style-related requests, and ( d ) content-related requests. Non-significant regression coefficients are marked in gray. See Table S5 in Supplemental Information for full regression tables. ChatGPT’s responses adapt to controversial issues and users’ prompting strategies, especially in terms of cognitive complexity RQ3 asked about how issue controversy and users’ prompting strategies influence ChatGPT’s responses. Figure 4 below reports regression coefficients for both sets of models: one using the overall number of prompting strategies (Panels a1-a2), and another using three prompting categories as predictors (Panels b1-b4). Consistently, for controversial issues, ChatGPT provided more references to external links, generated less structured responses, and used more cognitively complex and action-oriented language (Fig. 4 Panels a1, b1). These effects held when users’ communication styles were controlled, suggesting that ChatGPT adjusts its communication styles to contextual cues rather than simply mirroring user prompts. Fig. 4. Open in a new tab Regression coefficients and confidence intervals for the effects of issue controversy and prompting strategies. Panels a1-a2 show results from models with issue controversy and prompting strategy count as independent variables. Panels b1-b4 show results with issue controversy and three prompting strategy categories (i.e., user-supplied information, style-related requests, and content-related requests). Non-significant regression coefficients are marked in gray. Covariates were controlled in all models. See Tables S11-S12 for full regression tables. Further, the prompting strategy count overall did not significantly predict ChatGPT responses (Fig. 4 Panel a2), but more nuanced patterns emerged with three prompting strategy categories (Fig. 4 Panels b2-b4). When users provided additional information, ChatGPT responded with more action-oriented language. When users made style-related instructions, ChatGPT toned down cognitive complexity and used simpler language. However, when users made content-related requests, ChatGPT used more cognitively complex language for elaborate reasoning. These findings suggest that different types of prompting strategies activate distinct communication styles in ChatGPT responses. Overall, cognitive complexity is the most sensitive communication style, influenced by both user prompts and issue controversy. Cognitively complex ChatGPT responses decreased users’ AI perceptions but improved issue-specific attitudes RQ4 asked about how ChatGPT responses may impact users’ AI perceptions and issue-specific attitudes. Across all communication styles, only cognitive complexity showed significant impacts (see Fig. 5 below for illustration). ChatGPT with higher cognitive complexity was perceived as less favorable, with lower levels of perceived response quality, likability, and intelligence. This pattern suggests that users prefer straightforward responses with plain language in CIS. Meanwhile, cognitively complex responses positively affected issue-specific attitudes. Specifically, controlling for users’ pre-existing attitudes, cognitively complex responses resulted in stronger risk perceptions of microplastics and climate change, as well as increased support for vaccination, artificial sweetener, immigration, and highway reconstruction. This finding suggests that while users may perceive complex ChatGPT responses as less favorable, the cognitively demanding communication style nonetheless contributed to the persuasiveness of ChatGPT responses, implicitly influencing issue-relevant post-interaction attitudes with one-shot interaction. Fig. 5. Open in a new tab Regression coefficients and confidence intervals for the effects of cognitive complexity in ChatGPT responses on users’ AI perceptions and issue-specific attitude, with covariates controlled. Non-significant regression coefficients are marked in gray. See Table S13 for full regression tables. Discussion With a between-subjects experiment of conversational information-seeking and analysis of real conversations, our study highlights three key findings. First, people rarely use prompting strategies for CIS, and these strategies are mostly used by educated and Democrat-leaning individuals. Second, the communication styles of ChatGPT’s responses adapt to users’ prompting strategy and issue controversy. Lastly, the cognitive complexity of ChatGPT decreased users’ AI perceptions while positively shaping issue-specific attitudes. Below, we interpret potential explanations for our findings, propose actionable implications for conversational AI system design, and suggest directions for future research. First, we found that only 19.1% of users employed at least one prompting strategy, and these users were disproportionately educated and Democrat-leaning. This finding suggests a second-level digital divide, such that disparities exist beyond technology adoption but in how people actually use technologies 8 . It is consistent with previous findings that education predicts the digital divide 58 . In addition, it aligns with prior work on partisan gaps in technology adoption: Democrats are more likely to use social media 59 and are more open to using AI in policy-making 60 than Republicans. These findings reveal how socio-demographics and ideological leaning influence how people approach and perceive emerging technologies. This finding fills a critical gap in the research on conversational AI that focuses on the system-side disparities in AI responses by highlighting the human-side disparities and demonstrating how users actually prompt for information seeking. At the same time, the relatively low overall use of prompting strategies suggests that advanced prompting may not be essential for CIS: systems like ChatGPT can already interpret user intent and sustain natural conversations without engineered inputs 61 . Future research, especially surveys and qualitative studies, is needed to understand when and why users adopt prompting strategies. Second, our results suggest that ChatGPT adapts communication styles to not only users’ prompting strategies but also issue controversy, demonstrating contextual adaptability. We found that different types of prompting strategies activated distinct communication styles, which suggests that ChatGPT can adjust both content and styles to user intent. For example, when users employed content-related prompting strategies such as stepwise reasoning, ChatGPT prioritized response depth and elaborated with higher cognitive complexity. Moreover, ChatGPT’s responses to controversial issues contain more action-oriented language, higher levels of cognitive thinking, less structured responses (e.g., bullet points, subheadings), and more external references in comparison with non-controversial issues. It suggests that ChatGPT also adapts to broader social contexts in terms of communication styles. This finding adds to the growing body of research on AI’s social intelligence, which primarily focuses on adaptation to human users 62 . While this contextual adaptability does not necessarily imply that AI is socially aware or socially intelligent, it indicates that AI systems can resemble human-like adaptability by recognizing and responding to unique communication patterns across issue controversy in the training data. For instance, human conversations on controversial issues tend to be more analytical and emotional 63 ; ChatGPT may reproduce similar linguistic patterns. At the same time, given the existing finding that AI may amplify confirmation bias 20 , such contextual adaptation needs further academic attention to prevent potential harm caused by AI systems. Lastly, our results reveal a seemingly contradictory finding: users reported more negative AI perceptions but more positive issue-relevant attitudes with cognitively complex ChatGPT responses, with pre-existing attitudes and issue controversy controlled. Cognitive complexity may have implicit persuasive effects, even if users consciously dislike it. It aligns with prior work on ChatGPT-3, where opinion-minority users reported worse experiences but still shifted their attitudes, potentially due to cognitive dissonance 29 . One possible explanation lies in the cognitive disfluency, which refers to the mental difficulty of processing cognitively complex responses 64 . Complex information may disrupt cognitive fluency and decrease positive affect 65 , but it can also activate systematic processing and facilitate implicit learning 66 . Therefore, while users may prefer simple responses, they may unconsciously allocate cognitive resources to systematically process cognitively complex AI responses, thereby having more positive context-relevant attitudes. Our findings suggest that conversational AI should be designed to increase accessibility and inclusivity. First, because most people do not use advanced prompting, systems should support effective interactions without assuming technical expertise. In addition, ensuring that conversational AI adapts in ways that promote constructive dialogue, rather than reinforcing existing biases, will be critical as these tools are increasingly used in everyday contexts. Second, our findings point to a potential design challenge: cognitively complex responses strengthen issue-specific attitudes but reduce how favorably users view the AI. To reconcile this tension, systems could present complexity in more user-friendly ways. For instance, it would be beneficial to combine complex reasoning with stylistic elements, such as conversational tone or concrete examples, to maintain likability even when information is dense. This is essential to build credible, likable, and accessible conversational AI systems. Several limitations of this study need to be noted. First, the information-seeking interactions with ChatGPT in this experiment were not fully controlled. Although users were instructed to interact with GPT-4o, a few participants did not follow instructions. Further, we did not control the alternative settings, such as memory functions, which could influence the conversation quality and experience. Although this design was meant to capture naturalistic CIS with ChatGPT, future studies should aim for a cleaner experimental design, ensuring standardized conversation settings to minimize variability in user interactions. Second, this study focuses exclusively on ChatGPT. While this decision allows for consistency in experimental conditions, it may limit the generalizability of our findings to other conversational AI systems that exhibit different communication styles or interaction patterns. Future research should replicate this study across multiple conversational AI systems, such as Gemini and Copilot, to assess the variation of contextual and individual-level adaptations across AI systems. Third, our findings on communication styles and their persuasive effects are based on a one-shot interaction, and we did not measure attitude change directly. Given that users often engage in conversations with AI over time, future research should expand the scope of the current study and investigate the longitudinal effects of communication styles on users’ perceptions and attitudes with repeated conversations and information exposures. Further, future research may explore how individual differences (e.g., cognitive style, prior attitude) may shape these dynamics. Lastly, although we observed systematic differences in ChatGPT’s communication styles across issue contexts and prompting strategies, these findings are correlational. Future work should employ more controlled manipulations to examine whether these patterns reflect causal adaptation to user prompts. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (183.8KB, pdf) Author contributions H.X. proposed the research concept and ideation. H.X., Y.O., and X.Zhou. designed the study. H.X., X.Zhang., and B.O. annotated user prompts. H.X. conducted data analysis. X.Zhang. reviewed and validated data analysis. H.X., Y.O., and X.Zhou. wrote the manuscript. B.O. reviewed and edited the manuscript. Funding This research is supported by the University of Utah. Data availability The study data and replication code are available at: https://osf.io/9fbw7 . Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Radlinski, F. & Craswell, N. A. Theoretical Framework for Conversational Search. in Proceedings of the Conference on Conference Human Information Interaction and Retrieval 117–126 (Association for Computing Machinery, 2017). 10.1145/3020165.3020183 2. Zamani, H., Trippas, J. R., Dalton, J. & Radlinski, F. Conversational Information Seeking (Now, 2023). 3. Fischer, S. ChatGPT is still by far the most popular AI chatbot. Axios (2025). https://www.axios.com/2025/09/06/ai-chatbot-popularity 4. McClain, C. Americans’ use of ChatGPT is ticking up, but few trust its election information. Pew Res. Center (2024). https://www.pewresearch.org/short-reads/2024/03/26/americans-use-of-chatgpt-is-ticking-up-but-few-trust-its-election-information/ 5. Sun, X. et al. Trusting the Search: Unraveling Human Trust in Health Information from Google and ChatGPT. (2024). 6. Logg, J. M., Minson, J. A. & Moore, D. A. Algorithm appreciation: People prefer algorithmic to human judgment. Organ. Behav. Hum. Decis. Process. 151 , 90–103 (2019). [ Google Scholar ] 7. Sundar, S. & Kim, J. Machine Heuristic: When we trust computers more than humans with our personal information. in Proceedings of the CHI Conference on Human Factors in Computing Systems (ACM, 2019). 10.1145/3290605 8. Hargittai, E. Second-Level Digital Divide: Differences in People’s Online Skills. First Monday . 10.5210/fm.v7i4.942 (2002). doi:10.5210/fm.v7i4.942. [ Google Scholar ] 9. Khowaja, S. A. et al. PrivAcy, Digital divide, and Ethics) Evaluation: A Review. Cogn. Comput. 16 , 2528–2550 (2024). [ Google Scholar ] 10. Kacperski, C. et al. Characteristics of ChatGPT users from Germany: Implications for the digital divide from web tracking data. PLOS ONE . 20 , e0309047 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Liao, W., Weisman, W. & Thakur, A. On the Motivations to Seek Information From Artificial Intelligence Agents Versus Humans: A Risk Information Seeking and Processing Perspective. Sci. Commun. 46 , 458–486 (2024). [ Google Scholar ] 12. Hong, S. J. What drives AI-based risk information-seeking intent? Insufficiency of risk information versus (Un)certainty of AI chatbots. Comput. Hum. Behav. 162 , 108460 (2025). [ Google Scholar ] 13. Ju, B. & Stewart, J. B. Empowering Users with ChatGPT and Similar Large Language Models (LLMs): Everyday Information Needs, Uses, and Gratification. Proc. Assoc. Inf. Sci. Technol. 61 , 172–182 (2024). 14. Olea, C. et al. Evaluating persona prompting for question answering tasks. in Proceedings of th e 10th international conference on artificial intelligence and soft computing, Sydney, Australia (2024). 15. Wei, J. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2023). 16. Azzopardi, L., Dubiel, M., Halvey, M. & Dalton, J. A Conceptual Framework for Conversational Search and Recommendation: Conceptualizing Agent-Human Interactions During the Conversational Search Process. (2024). https://arxiv.org/abs/2404.08630 17. Walther, J. B. Theories of computer-mediated communication and interpersonal relations. in The handbook of interpersonal communication (eds Daly, J. A. & Knapp, M. L.) vol. 4 443–479 (2011). 18. Zheng, Q., Tang, Y., Liu, Y., Liu, W. & Huang, Y. UX Research on Conversational Human-AI Interaction: A Literature Review of the ACM Digital Library. in Proceedings of the CHI Conference on Human Factors in Computing Systems (Association for Computing Machinery, 2022). 10.1145/3491102.3501855 19. Sun, D. Q. et al. DELPHI: Data for Evaluating LLMs’ Performance in Handling Controversial Issues (2023). 20. Sharma, N., Liao, Q. V. & Xiao, Z. Generative Echo Chamber? Effect of LLM-Powered Search Systems on Diverse Information Seeking. in Proceedings of the CHI Conference on Human Factors in Computing Systems (Association for Computing Machinery, 2024). 10.1145/3613904.3642459 21. Petty, R. E. & Cacioppo, J. T. The Elaboration Likelihood Model of Persuasion. in Communication and Persuasion: Central and Peripheral Routes to Attitude Change (eds Petty, R. E. & Cacioppo, J. T.) 1–24 (Springer, 1986). 10.1007/978-1-4612-4964-1_1. 22. Arora, S. et al. Ask Me Anything: A simple strategy for prompting language models (2022). https://arxiv.org/abs/2210.02441 23. Bates, M. J. Information search tactics. J. Am. Soc. Inf. Sci. 30 , 205–214 (1979). [ Google Scholar ] 24. White, J. et al. A prompt pattern catalog to enhance prompt engineering with chatgpt. ArXiv Prepr. ArXiv230211382 (2023). 25. Trippas, J. R., Lawati, A., Mackenzie, S. F. D. & Gallagher, L. J. What do Users Really Ask Large Language Models? An Initial Log Analysis of Google Bard Interactions in the Wild. in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval 2703–2707 (Association for Computing Machinery, 2024). 10.1145/3626772.3657914 26. OpenAI. Prompt engineering - OpenAI API. Six strategies for geting better results (2024). https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results 27. Zhang, C., Rice, R. E. & Wang, L. H. (eds) (Xinyi), College students’ literacy, ChatGPT activities, educational outcomes, and trust from a digital divide perspective. New Media Soc. (2024). 10.1177/14614448241301741 28. Bassignana, E., Curry, A. C. & Hovy, D. The AI Gap: How Socioeconomic Status Affects Language Technology Interactions (2025). 29. Chen, K., Shao, A., Burapacheep, J. & Li, Y. Conversational AI and equity through assessing GPT-3’s communication with diverse social groups on contentious topics. Sci. Rep. 14 , 1561 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Knoth, N., Tolzin, A., Janson, A. & Leimeister, J. M. AI literacy and its implications for prompt engineering strategies. Comput. Educ. Artif. Intell. 6 , 100225 (2024). [ Google Scholar ] 31. Bell, R. A. & Daly, J. A. The affinity-seeking function of communication. Commun. Monogr. 51 , 91–115 (1984). [ Google Scholar ] 32. Blankenship, K. L. & Craig, T. Y. Language Use and Persuasion: Multiple Roles for Linguistic Styles. Soc. Personal Psychol. Compass . 5 , 194–205 (2011). [ Google Scholar ] 33. Burgoon, J. K. Interaction Adaptation Theory. in The International Encyclopedia of Communication (Wiley, Ltd, doi:10.1002/9781405186407.wbieci041. (2008). [ Google Scholar ] 34. Braun, T., Rohr, M. K., Wagner, J. & Kunzmann, U. Perceived reciprocity and relationship satisfaction: Age and relationship category matter. Psychol. Aging . 33 , 713–727 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 35. Xue, H., Liao, W. & Zhang, J. Interaction dynamics of social support expressions predict future support-seeking behaviors in online support groups. Comput. Hum. Behav. 156 , 108224 (2024). [ Google Scholar ] 36. Chang, R. C., Wang, H. C. & Convergence Reciprocity, and Asymmetry: Communication Accommodation Between Large Language Models and Users Across Cultures. Proc. Annu. Meet. Cogn. Sci. Soc. 47 (2025). 37. Blevins, T., Schmalwieser, S. & Roth, B. Do language models accommodate their users? A study of linguistic convergence. Preprint at (2025). https://arxiv.org/abs/2508.03276 38. Thomas, P., McDuff, D., Czerwinski, M. & Craswell, N. Expressions of Style in Information Seeking Conversation with an Agent. in Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval 1171–1180 (Association for Computing Machinery, 2020). 10.1145/3397271.3401127 39. Perez, E. et al. Discovering Language Model Behaviors with Model-Written Evaluations. (2022). https://arxiv.org/abs/2212.09251 40. Koopman, B. & Zuccon, G. Dr ChatGPT tell me what I want to hear: How different prompts impact health answer correctness. in Proceedings of the Conference on Empirical Methods in Natural Language Processing (eds Bouamor, H., Pino, J. & Bali, K.) 15012–15022 (Association for Computational Linguistics, 2023). 10.18653/v1/2023.emnlp-main.928 41. Levine, T. R. Truth-Default Theory (TDT): A Theory of Human Deception and Deception Detection. J. Lang. Soc. Psychol. 33 , 378–392 (2014). [ Google Scholar ] 42. Burgoon, J. K., Le Poire, B. A. & Rosenthal, R. Effects of Preinteraction Expectancies and Target Communication on Perceiver Reciprocity and Compensation in Dyadic Interaction. J. Exp. Soc. Psychol. 31 , 287–321 (1995). [ Google Scholar ] 43. Ghafouri, V. et al. ACM,. AI in the Gray: Exploring Moderation Policies in Dialogic Large Language Models vs. Human Answers in Controversial Topics. in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management 556–565 (2023). 10.1145/3583780.3614777 44. Bardol, F. ChatGPT Reads Your Tone and Responds Accordingly – Until It Does Not – Emotional Framing Induces Bias in LLM Outputs (2025). 45. Hautzenberger, C. & Muellen, S. The Hostility of LLMs Towards Humans on Borderline Controversial Topics When Induced with Maliciously Crafted Prompts (2024). 10.36227/techrxiv.171777771.17020327/v1. 46. Chaves, A. P., Doerry, E., Egbert, J. & Gerosa, M. It’s How you say it: Identifying appropriate register for chatbot language design. in Proceedings of the 7th International Conference on Human-Agent Interaction 102–109 (ACM, 2019). 10.1145/3349537.3351901 47. Liao, W., Zhang, J., Oh, Y. J. & Palomares, N. A. Linguistic Accommodation Enhances Compliance to Charity Donation: The Role of Interpersonal Communication Processes in Mediated Compliance-Gaining Conversations. J. Comput. -Mediat Commun. 26 , 167–185 (2021). [ Google Scholar ] 48. Miehle, J., Minker, W. & Ultes, S. Exploring the Impact of Elaborateness and Indirectness on User Satisfaction in a Spoken Dialogue System. in Adjunct Publication of the 26th Conference on User Modeling, Adaptation and Personalization 165–172 (Association for Computing Machinery, 2018). 10.1145/3213586.3226213 49. Chevi, R., Inui, K., Solorio, T. & Aji, A. F. How Individual Traits and Language Styles Shape Preferences In Open-ended User-LLM Interaction: A Preliminary Study (2025). 50. Štajner, S., Yenikent, S. & Franco-Salvador, M. Five Psycholinguistic Characteristics for Better Interaction with Users. in 8th International Conference on Behavioral and Social Computing (BESC) 1–7 (2021). 10.1109/BESC53957.2021.9635267 51. Pennebaker, J. W., Chung, C. K., Frazee, J., Lavergne, G. M. & Beaver, D. I. When Small Words Foretell Academic Success: The Case of College Admissions Essays. PLOS ONE . 9 , e115844 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Boyd, R. L., Ashokkumar, A., Seraj, S. & Pennebaker, J. W. The development and psychometric properties of LIWC-22. Austin TX Univ. Tex. Austin 10 (2022). 53. Ding, Y. et al. Citations and Trust in LLM Generated Responses. Proc. AAAI Conf. Artif. Intell. 39 , 23787–23795 (2025). 54. Kools, M., Ruiter, R. A. C., van de Wiel, M. W. J. & Kok, G. The effects of headings in information mapping on search speed and evaluation of a brief health education text. J. Inf. Sci. 34 , 833–844 (2008). [ Google Scholar ] 55. Wojcieszak, M. et al. Can AI Enhance People’s Support for Online Moderation and Their Openness to Dissimilar Political Views? J. Comput. -Mediat Commun. 10.1093/JCMC/ZMAB006 (2021). [ Google Scholar ] 56. Kim, Y., Lee, J., Kim, S., Park, J. & Kim, J. Understanding users’ dissatisfaction with chatgpt responses: Types, resolving tactics, and the effect of knowledge level. in Proceedings of the 29th International Conference on Intelligent User Interfaces 385–404 (2024). 57. Brave, S., Nass, C. & Hutchinson, K. Computers that care: investigating the effects of orientation of emotion exhibited by an embodied computer agent. Int. J. Hum. -Comput Stud. 62 , 161–178 (2005). [ Google Scholar ] 58. Bagchi, K. Factors Contributing to Global Digital Divide: Some Empirical Results. J. Glob Inf. Technol. Manag . 8 , 47–65 (2005). [ Google Scholar ] 59. Vogels, E., Auxier, B. & Anderson, M. Partisan differences in social media use show up for some platforms, but not Facebook. Pew Res. Center (2021). https://www.pewresearch.org/short-reads/2021/04/07/partisan-differences-in-social-media-use-show-up-for-some-platforms-but-not-facebook/ 60. Gur, T., Hameiri, B. & Maaravi, Y. Political ideology shapes support for the use of AI in policy-making. Front Artif. Intell (2024). [ DOI ] [ PMC free article ] [ PubMed ] 61. Mao, K. et al. Association for Computational Linguistics, Singapore,. Large Language Models Know Your Contextual Search Intent: A Prompting Framework for Conversational Search. in Findings of the Association for Computational Linguistics: EMNLP 2023 (eds Bouamor, H., Pino, J. & Bali, K.) 1211–1225 (2023). 10.18653/v1/2023.findings-emnlp.86 62. Mathur, L., Liang, P. P. & Morency, L. P. Advancing Social Intelligence in AI Agents: Technical Challenges and Open Questions (2024). https://arxiv.org/abs/2404.11023 63. Addawood, A. et al. Telling apart tweets associated with controversial versus non-controversial topics. in Proceedings of the Second Workshop on NLP and Computational Social Science (eds Hovy, D.) 32–41 (Association for Computational Linguistics, 2017). 10.18653/v1/W17-2905 64. Alter, A. L. The Benefits of Cognitive Disfluency. Curr. Dir. Psychol. Sci. 22 , 437–442 (2013). [ Google Scholar ] 65. Labroo, A. A. & Pocheptsova, A. Metacognition and consumer judgment: fluency is pleasant but disfluency ignites interest. Consum. Behav. 10 , 154–159 (2016). [ Google Scholar ] 66. See, Y. H. M., Petty, R. E. & Evans, L. M. The impact of perceived message complexity and need for cognition on information processing and attitudes. J. Res. Personal . 43 , 880–889 (2009). [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (183.8KB, pdf) Data Availability Statement The study data and replication code are available at: https://osf.io/9fbw7 . Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.1 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top