Conversational presentation mode increases credibility judgements during information search with ChatGPT

Source: Christine Anderl, Stefanie H. Klein, Büsra Sarigül, Frank M. Schneider, Junyi Han, Paul L. Fiedler, Sonja Utz; Scientific Reports 14, journal article (open access) · 2024-07-25

People increasingly use large language model (LLM)-based conversational agents to obtain information. However, the information these models provide is not always factually accurate. Thus, it is critical to understand what helps users adequately assess the credibility of the provided information. Here, we report the results of two preregistered experiments in which participants rated the credibility of accurate versus partially inaccurate information ostensibly provided by a dynamic text-based LLM-powered agent, a voice-based agent, or a static text-based online encyclopedia. We found that people were better at detecting inaccuracies when identical information was provided as static text compared to both types of conversational agents, regardless of whether information search applications were branded (ChatGPT, Alexa, and Wikipedia) or unbranded. Mediation analysis overall corroborated the interpretation that a conversational nature poses a threat to adequate credibility judgments. Our research highlights the importance of presentation mode when dealing with misinformation.

Introduction. When OpenAI released their large language model (LLM)-powered conversational agent ChatGPT in November 2022, it was adopted faster than any prior medium or technology— within two months, ChatGPT had 100 million active users. Since then, LLMs have become ubiquitous in both public and academic discourse, and multiple companies have launched their own LLM-based conversational agents. Because LLMs interpret human language in a way that allows them to take into account user intent and the contextual meaning of terms, enabling unprecedentedly human-like conversations, they are increasingly used for information enquiries. Indeed, their potentially disruptive implications for search engines are widely discussed. However, LLMs are known to “hallucinate” at times, that is, to make up information—often convincingly so. Therefore, it is crucial to understand what helps accurately assess the credibility of the provided information. We address this question in two preregistered experiments. While prior research descriptively explored the accuracy of responses provided by ChatGPT or the percentage of made-up information in narrowly defined areas, the present work focuses on how the presentation mode of information influences credibility judgments. We build on prior theorizing and empirical work demonstrating that the modality in which information is presented affects credibility judgments and extend it by providing robust evidence indicating that conversational nature plays an important additional role. To do so, we compare information allegedly stemming from ChatGPT with information allegedly stemming from Alexa or from Wikipedia. These applications differ on the one hand in the modality in which information is presented—ChatGPT and Wikipedia are text-based, so information is read, Alexa is voice-based, so information is listened to. On the other hand, they also differ in whether they present information in a conversation-like, dynamic style (Alexa and ChatGPT) or not (Wikipedia). We furthermore assessed several potential mediators to get first hints on why these differences might occur, that is, on the likely underlying mechanisms. By also using unbranded versions in Experiment 2, we demonstrate the generalizability of our findings to dynamic text-based agents, voice-based agents, and static texts beyond ChatGPT, Alexa, and Wikipedia. Auditorily presented information is perceived as more credible than textually presented information.

Method. Experiment 1 Design and participants We employed a 3 (presentation mode: static text-based encyclopedia vs. dynamic text-based agent vs. voice-based agent; between-subjects factor) × 2 (information accuracy: low vs. high; within-subjects factor) × 2 (set: A vs. B) mixed factorial experimental design. After providing written informed consent, participants were randomly assigned to one of the three presentation modes. In each condition, we presented participants with six questions and answers that varied in accuracy. We counterbalanced the low and high accuracy answers by randomly assigning participants to one of two sets: low accuracy answers in set A (low: Appendicitis, Bones, Wolf) were presented with a high accuracy answer in set B (low: Slovenia, Hookah, Titanic) and vice versa. Detailed absolute cell sizes for both Experiments 1 and 2 are reported in Table S8 of the online supplemental materials. The study received ethics approval from the IRB of the first author’s institution (Ethics Committee of the Leibniz-Institut für Wissensmedien). We preregistered the study on the Open Science Framework (OSF) platform: https://osf.io/puyv7. We based our sample size considerations on prior work which analyzed data from N = 300 participants. Because our experiment included the additional between-subjects factor set and because we wanted to test more and higher-level interaction effects, we decided to collect more observations. We recruited N = 600 UK-based participants via Prolific in June 2023. Participants had to be at least 18 years old, and fluent in English. They received 2.25 GBP for taking part in the experiment. Following our preregistration, we excluded participants who searched for the presented information online while completing the experiment (n = 44); fact-checking did not differ between presentation modes (χ2 = 1.52, p = 0.47). The final sample comprised N = 556 participants (nfemale = 281, Mage = 43.33, SDage = 13.10). Most participants (n = 348) had studied or were currently studying at a higher education institution. One hundred and eighty-nine participants were assigned to the static text-based encyclopedia condition. The voice- and dynamic text-based agent conditions comprised 183 and 184 participants, respectively. Two-hundred and eighty-one participants saw Set A and 275 participants saw Set B.

Procedure Once they had started the questionnaire, participants were first asked about their information search behavior. Then, they were presented with six information snippets, each consisting of a question and an answer. Participants were asked to rate the credibility of the information in each snippet. After participants had read or listened to all information snippets, we assessed the potential mediators. In addition, we asked how involved participants were in the topics, how trustworthy they perceived different information sources to be, and whether they had searched online for the information presented. Participants were then redirected to Prolific and paid for their participation.

Stimuli and manipulation of independent variables To prepare the information snippets, we selected six questions that related to general knowledge and covered diverse topics: What do I do when I encounter a wolf? What are the risks of hookah smoking? How many people died when the Titanic sank? What is appendicitis? How many bones are in the human body? Tell me something about the country Slovenia! For each of them, we first created the accurate answers (high accuracy condition) by shortening the respective Wikipedia articles and informational texts from additional, trustworthy sources. To create the snippets for the low accuracy condition, we used the same, short answers, but built in false information. To make sure that detecting inaccuracies did not entirely depend on prior knowledge, we included logical inconsistencies in addition to factual errors. For instance, the answer to the question of how to behave when encountering a wolf included the correct recommendation to make loud noises, combined with the inconsistent reason to do so in order not to be noticed.

Discussion. With recent advances in LLMs, people increasingly turn to conversational agents like ChatGPT, Bing AI, or Google Bard when searching for information and LLMs are increasingly implemented into voice-based agents like Alexa as well. However, concerns have been raised about their oftentimes factually inaccurate responses. Our findings here suggest that this may be particularly problematic because the conversational nature in which information is delivered by these agents decreases people’s readiness to discern between accurate and inaccurate information, even when text-based presentation would allow them to reread the information more than once. The present findings provide robust evidence that information received by dynamic text-based agents like ChatGPT is perceived as more credible than identical information presented as static text, but as less (Experiment 1) or equally (Experiment 2) credible compared to the same information presented by voice-based agents. These effects were strongly driven by differences in discernment between accurate and inaccurate information, which is, arguably, highly relevant, and importantly complement previous research showing similar findings for actual information search that were however observed in a much less controlled setting (because the information content itself differed). The present research also makes important theoretical contributions by demonstrating that the observed effects were independent of specific branded applications and even showed opposite pattern to people’s self-reported general assessments of the applications’ trustworthiness. The most plausible interpretation for the observed pattern of results appears to be that both a modality effect (i.e., reading vs. listening) and an effect of conversational nature (i.e., conversational vs. non-conversational) work in parallel and in partially opposing ways: discernment between accurate and inaccurate information benefitted from reading (vs. listening) and from being presented in a non-conversational (vs. conversational) way. Because dynamic text-based agents combine both, higher discernment through reading and reduced discernment through the conversational nature, they score between voice-based agents (lower discernment through listening and conversational nature) and static text (higher discernment through reading and non-conversational nature). Mediation analyses overall corroborated this interpretation in suggesting that when comparing the two conversational agents with the static text-based encyclopedia, the positive and negative effects of the different cognitive and social-affective processes seemed to cancel each other out. Interestingly, when comparing the conversational agents with each other, processing fluency and perceived enjoyment seemed to partially (Experiment 1) or fully (Experiment 2) suppress the effect of presentation mode on credibility because the effect got stronger when accounting for the indirect effects. Notably, all reported significant effects were small to medium in size and thereby exceeded our a priori defined threshold for a minimal effect size of interest (SESOI). These mediation results should nevertheless be interpreted with caution because the effect of enjoyment was opposite to what we had predicted and because non-experimental mediation does not allow for causal interpretation. As such, the identified putative mechanisms should be seen as setting a direction for future experimental work rather than a final conclusion. For instance, the effects of processing fluency could be experimentally investigated by adapting the fonts of the text-based conditions to be less easily readable.

Limitations. However, the observed effect of the conversational nature (i.e., agency and interactivity) on credibility should, arguably, be stronger rather than weaker in a more interactive setting, especially because experiences of social presence, which mediated the effect, may be assumed to get stronger with increasing familiarity with the agent. Moreover, we would like to stress that because our proposed mediators were not independently manipulated, the present work does not allow for the conclusion that they are indeed the causal drivers of the observed effects of presentation mode on credibility. Much like for other correlational work, it is impossible to rule out that the observed correlational effects are driven by another, potentially unmeasured factor. Combined with our experimental findings addressing the how, findings from our mediation analysis may, however, be helpful in identifying fruitful directions for future research addressing the why. To establish causality, future research will need to experimentally manipulate mediators to test their causal effects on credibility. Finally, we kept information constant between the three applications.