The World Health Organization and the United Nations have issued warnings about an “infodemic,” highlighting the spread of misinformation alongside the COVID-19 pandemic on social media [1]. Misinformation is characterized as “factually incorrect information not backed up by evidence” [2]. This misleading information frequently encompasses harmful health advice, misinterpretations of government control measures and emerging sciences, and conspiracy theories [3]. This phenomenon has inflicted detrimental impacts on public health, carrying “severe consequences with regard to people’s quality of life and even their risk of mortality” [4].

Automatic algorithms are increasingly recognized as valuable tools in mitigating the harm caused by misinformation. These techniques can rapidly identify misinformation, predict its spread, and have demonstrated commendable performance. The state-of-the-art detection techniques exhibit accuracy ranging from 65% to 90% [5,6], while spread-prediction techniques achieve performance levels between 62.5% and 77.21% [7,8]. The high accuracy of these techniques can be largely attributed to the incorporation of handcrafted or deep-learned linguistic and social features associated with misinformation [9-11]. Scholars have consistently invested efforts in integrating theoretically relevant features into algorithmic frameworks to enhance accuracy further.

Scholars have introduced diverse frameworks for misinformation detection and spread-prediction algorithms. Nevertheless, existing frameworks have predominantly concentrated on the intricate post-level signals of misinformation, emphasizing linguistic and social features (such as user relationships, replies, and knowledge sources) associated with misinformation. Notably, these frameworks have often overlooked the characteristics of the information environment in which misinformation originates and proliferates [12]. This neglect could potentially result in diminished performance for misinformation detectors when applied in various real-world misinformation contexts. This is due to the fact that different misinformation contexts possess unique characteristics within their information environment, influencing the types of misinformation that can emerge and thrive [13]. An indispensable characteristic of the information environment concerning misinformation is uncertainty. Uncertainty arises when the details of situations are ambiguous, complex, unpredictable, or probabilistic, and when information is either unavailable or inconsistent [14]. In uncertain situations, individuals tend to generate and disseminate misinformation as a means of resisting uncertainty and seeking understanding amid chaotic circumstances [15,16]. The COVID-19 pandemic serves as a notable example, marked by a lack of understanding of emerging science [17], uncertainties surrounding official guidelines and news reports [18], and unknown impacts on individuals and society [19]. Hence, in this study, we recognize uncertainty as the pivotal feature in the information environment of misinformation. Our objective is to formulate a novel framework for perceiving environmental uncertainty, specifically tailored for the detection and spread prediction of misinformation during the COVID-19 pandemic.

Our contributions can be outlined as follows. Theoretically, we provide a comprehensive exploration of uncertainty across 4 distinct scales of the information environment, namely, the physical environment, macro-media environment, micro-communicative environment, and message framing. These scales collectively contribute to the emergence and dissemination of misinformation. Furthermore, we hold the distinction of being the pioneers in integrating Environmental Uncertainty Perception (EUP) into the realms of misinformation detection and spread prediction. In terms of methodology, we introduce the EUP framework, designed to capture uncertainty signals from the information environment of a given post for both misinformation detection and spread prediction. Our experiments conducted on real-life data underscore the effectiveness of the EUP framework.

This paper unfolds as follows: In the “Related Work” section, we provide a concise review of the related work. The “Proposed Theoretical Framework” section elucidates uncertainty features within the information environment, which are pertinent to misinformation detection and spread prediction. Moving on to the “Research Objectives” section, we outline our study objectives. The “Methods” section details our methodology for testing the proposed framework. In the “Data Set and Experiment” section, we present our data set, experiments, and comprehensive analyses. The “Discussion” section delves into discussions on our findings, unraveling the theoretical and practical implications of our work. Finally, the “Conclusions” section concludes with a summary and outlines directions for future research.

Detecting misinformation on social media represents a burgeoning research field that has garnered considerable academic attention. Multiple frameworks have been put forth for this task, primarily falling into 2 approaches: the post-only approach and the “zoom-in” approach [12]. In the former, frameworks focus on studying post features to differentiate misinformation from general information. Linguistic features, including novelty, complexity, emotions, and content topics, are frequently explored [6,11]. Additionally, researchers have delved into multimodal features, particularly those based on visuals [20,21]. Deep learning models in natural language processing have also proven beneficial for the misinformation detection task [5,22].

The “zoom-in” approach places emphasis on socio-contextual signals, centering on users’ networking aspects (eg, user relationships, number of replies, number of created threads; [23,24]) and network characteristics (eg, degree centrality [25]). Another line of research underscores the significance of relevant knowledge sources, including fact-checking websites [26] and knowledge graphs [27], which can be used to validate specific claims of interest.

Recently, Sheng et al [12] introduced a “zoom-out” approach, concentrating on the information environments of misinformation that can offer signals for detection. In their approach, they incorporated the news environment into fake news detection. Their hypothesis posited that fake news should not only be relevant but also novel and distinct from recent popular news, enabling them to capture audience attention and achieve widespread dissemination. Their findings revealed that signals of popularity and novelty can enhance the performance of state-of-the-art misinformation detectors.

In the realm of misinformation detection, misinformation spread prediction represents another challenging task, albeit one that has received limited attention. This task involves predicting whether a piece of misinformation is likely to be disseminated to a broader audience through actions such as likes, comments, and shares. Within this context, our specific focus is on predicting whether misinformation is likely to be retweeted. This can be viewed as a binary classification task, akin to misinformation detection. Frameworks for this task typically incorporate linguistic and social features, which may overlap with or differ from those used in misinformation detection. Linguistic features such as persuasive styles, emotional expressions, and message coherence prove valuable in predicting the spread of misinformation [28,29]. Additionally, social features, including user metadata (eg, number of friends, verification) and tweet metadata (eg, presence of images and URLs), are identified as relevant factors for predicting misinformation spread [25].

Our research objective is to explore whether uncertainty features within the information environment can enhance the effectiveness of misinformation detection and spread prediction. To achieve this, we introduce a novel EUP framework specifically designed for both tasks. We seek to assess the standalone effectiveness of the EUP and anticipate that it can augment the capabilities of existing state-of-the-art misinformation detectors and predictors. Therefore, we conducted experiments to answer the following research questions:

Figure 1 offers an overview of the EUP pipeline. The model consists of 4 uncertainty extraction components. Upon receiving a post (denoted as p), the initial step involves constructing its macro-media environment and micro-communicative environment. This is accomplished by extracting recent news and social media data, respectively. Subsequently, we use a probabilistic model and a similarity calculation method to derive the uncertainty information for the 2 environments mentioned above, denoted as I_M and I_C. Likewise, we utilized the probabilistic model to capture the uncertainty of the post p itself, resulting in the representation of message framing denoted as I_F. Simultaneously, the operationalization of uncertainty in the physical environment entails using the number of COVID-19 cases and the volume of news as key indicators, denoted as I_P. Lastly, the 4 vectors are integrated using a gate guided by the extracted post feature o (which may not necessarily equal p) from the misinformation detector, such as bidirectional encoder representations from transformers (BERT) [52]. The fused vectors I and o are then input into the final classifier, typically a multilayer perceptron (MLP), to predict whether p is fake or real in task 1 and low or high in task 2.

For detecting uncertainty in natural language [53], we used a probabilistic model that considers the local n-gram features of sentences. Each n-gram is assigned a weight that reflects its tendency to convey uncertainty. The definition of each feature involves a quadruplet (type, size, context, and aggregation). “Type” signifies the type of n-gram considered, such as lemma or morphosyntactic pattern. “Size” indicates the size of the n-gram. “Context” serves as an indicator, specifying whether the weight is based on the occurrence frequency of the n-gram in an uncertain sentence or on the occurrence frequency of the n-gram as an uncertainty marker. “Aggregation” refers to the method used to consolidate different scores of the n-grams within a sentence. Multimedia Appendix 1 [49,54-57] furnishes a summary of the diverse features, denoted as F_i, that are scrutinized in the uncertainty detection model.

Next, we exemplify the calculation of uncertainty using 1 of these features, F₁, as an illustration. F₁ is defined by the quadruplet (Lemma, 1, uncertainty marker, and sum). For each lemma w, we can compute the number of occurrences in the corpus, the number of occurrences in uncertain sentences, and the number of occurrences as an uncertainty marker, denoted as F_s, F_u, and F_m, respectively. The conditional probability of a lemma w becoming an uncertainty marker is calculated using the following equation:

where c represents the class of context uncertainty under analysis, specifically whether it pertains to being an uncertainty marker. Additionally, we introduce a confidence score linked to the probability of mitigating the impact of instances where certain lemmas occur infrequently in the corpus yet yield a high probability:

F₁ takes into account both the conditional probability of each lemma w and the corresponding confidence score in the sentence s, and the formula is calculated as follows:

Similarly, other features F_i can be derived using the above method. We generated the uncertainty of the whole sentence by mean pooling to represent the average uncertainty signals of F_i:

where Norm(·) denotes the normalization.

We collect news reports from mainstream media outlets released within T days before the post p is published to construct a macro-media environment according to the following definition:

where E denotes the set of all collected news items, M denotes the set of news items in the macro-media environment of the post p, and t_p and t_e represent the release time of post p and news e, respectively. For post p or each news item e, the initial representations are the output of a pretrained language model (eg, BERT [52]), denoted as p and e, respectively.

The macro-media environment is expected to reflect the impact of a trending topic with high uncertainty on the veracity of a post. That is, if a post is related to a trending topic with (vs without) high uncertainty, it is then expected to be more likely misinformation and disseminated. To this end, the representation of the macro-media environment should consider both the correlation between the post and the environment and the uncertainty of the environment. We first calculate cosine similarity between p and each news item e in E:

We combine the similarity and environment representations to represent the similarity representation of a post p to the environment:

where e^M_i represents each news item in M and is the Hadamard product operator.

We then measure the uncertainty of the macro-media environment using the model described in the “Uncertainty Detection Model” section. The uncertainty representation of the macro-media environment, denoted as U_M, can be expressed by the following equation:

Finally, the macro-media environment of a post p is represented as an aggregation of the similarity representation of p to the environment (S_M) and the uncertainty representation of the environment (U_M) using an MLP, denoted as I_M:

where is the concatenation operator. The integration of an MLP is instrumental in the dual objective of retaining crucial information while concurrently achieving data dimensionality reduction. All MLPs are individually parameterized. We omit their index numbers in the above equations for brevity.

We collected tweets from Twitter (X; X Corp.) published within T days before the post p was published to construct the micro-communicative environment. We calculated the similarity of all tweets to the post p and selected the top k of them, using them as a micro-communicative environment (C), which is defined as follows:

where V denotes the set of all collected tweet items and t_v represents the release time of the tweet v.

where Topk(·) represents the operation of selecting the k tweets that have the highest similarity to p, k = r·|C′|, and r ∈ (0,1) represents the percentage of extraction.

Using the same approach as in the previous 2 sections, we derive the similarity representation of the post p to the micro-communicative environment and the uncertainty representation of the environment:

Finally, the micro-communicative environment of a post p is represented as an aggregation of the similarity representation of a post p to the environment (S_C) and the uncertainty representation of the environment (U_C) using an MLP, denoted as I_C:

To perceive the uncertainty in the message framing of post p, we used the same approach as described in the “Uncertainty Detection Model” section to construct the uncertainty representation of the post p:

To measure uncertainty in the physical environment, we collected the daily number of new cases from the start of the COVID-19 outbreak and counted the number of daily news items related to the outbreak, denoted as N^Cases and N^News, respectively. Intuitively, the higher the number of new cases and news items for a day, the more sensitive the public is to the social environment and the more uncertain the environment is on that day. Thus, the uncertainty factor in the physical environment is defined as follows:

where f^ph_i denotes the uncertainty factor at day i and abs is the absolute value operation. For each post, we can obtain the uncertainty factor for its corresponding date f^ph(p).

We added the uncertainty factor of the physical environment to the representations of macro-media environment (I_M), micro-communicative environment (I_C), and post message framing (I_F) to get the representation of the physical environment, denoted as I_P:

The study is exempt from ethical review for human subject research for the following reasons. First, the study uses data from 2 publicly available Twitter data sets collected through the official application programming interface (API) of the Twitter platform for gathering tweets. The news data set was obtained from the official websites of news media. Second, the data used in this study are anonymized and do not contain any personally identifiable information. It is also impossible to reidentify individuals from the data set. The data set is stored on a dedicated secure data server, and the analysis is conducted on the platform’s designated site. This process is undertaken for research purposes and adheres to Chinese data privacy laws and regulations. Third, this study does not involve any experimental manipulation of human individuals or other ethical concerns. For instance, it does not include data on children under 18 years of age, which require legally mandated parental or guardian supervision. It also does not encompass sensitive aspects of participants’ behavior or pose any physical, psychological, or economic harm or risk to the research participants.

Tables 3 and 4 showcase the performances of the EUP without baseline models and those of various baseline models, with and without EUP, for the misinformation detection and spread prediction tasks, respectively. The results indicate that the performances of EUP are comparable to those of state-of-the-art baseline models in both tasks. Moreover, it is noteworthy that all baseline models exhibit performance improvements when incorporating EUP for both tasks. These observations suggest the effectiveness of our proposed EUP.

We systematically eliminated individual components, namely, macro-media environment, micro-communicative environment, message framing, and physical environment, and assessed the modeling performances on the data set. Tables 5 and 6 illustrate that, under all experimental conditions, performance degrades when any of these components are removed. These results underscore the effectiveness of all 4 uncertainty features of the information environment for both misinformation detection and spread prediction.

To explore the impact of the day parameter (T) on the results during the construction of the macro-media environment and the micro-communicative environment, we experimented with different values of T. Specifically, we sequentially set T=1, 3, 5, 7, and 9 for the BERT + EUP model, and the experimental results are depicted in Figure 2. Despite the fact that increasing T results in larger macro-media and micro-communicative environments, the optimal performance was achieved when T=1.

We maintained the setting T=3 and systematically varied r, using values of 0.05, 0.1, 0.15, 0.2, 0.25, and 0.3 on the BERT + EUP model to examine the impact of r on the experimental results, as illustrated in Figure 3. The accuracy performance exhibited fluctuations with varying values of r. Notably, the highest accuracy for both tasks was observed when r=0.1.

In real-world scenarios, the distribution of real and fake information often exhibits significant imbalance. To evaluate the efficacy of our proposed EUP framework on unbalanced data sets, we conducted tests on data sets with varying ratios of real to fake data, ranging from 10:1 to 100:1. We measured and reported macro-F₁-scores and standardized partial area under the curve (AUC) with a false-positive rate of at most 0.1 (ie, spAUCFPR≤0.1 [66]) to assess the effectiveness of our EUP framework in handling nonbalanced data sets. As depicted in Figure 4, EUP yields relative improvements of 21.5% and 5.7% in macro-F₁-score and spAUCFPR≤0.1, demonstrating its effectiveness on unbalanced data sets.

First, this study enhances scholars’ comprehension of the misinformation detection and spread prediction problem by highlighting the significance of uncertainty in information environments. Notably, this research contributes to the literature by recognizing uncertainty features in the information environments of misinformation as a pivotal factor for improving detection and prediction algorithms during a pandemic. Our findings underscore that the EUP alone is sufficient for both tasks and has the potential to enhance the capabilities of state-of-the-art algorithms. In contrast to prior misinformation research that primarily concentrates on post content (such as post theme, sentiments, and linguistic characteristics, as seen in [6,11,29]) and network connections (eg, number of followers [25]) on social media, this study advances scholars’ understanding of the misinformation problem by emphasizing the importance of uncertainty in information environments. Recognizing and incorporating uncertainty as a fundamental concept in misinformation detection and spread prediction during crises hold theoretical significance. This is particularly relevant as a crisis is characterized by its unpredictable, unexpected, and nonroutine nature, inherently giving rise to uncertainty [38,67]. This uncertainty has been theorized to compel individuals to seek information as a coping mechanism for dealing with the anxiety and pressure generated by uncertainty. This process allows people to diminish uncertainty, restore a sense of normalcy, and alleviate anxiety [14,68]. Regrettably, this coping mechanism can inadvertently fuel the proliferation and dissemination of misinformation, particularly when there is a lack of timely and accurate information, contributing to the concurrent occurrence of an infodemic [6,11,50]. The current research seeks to advance the literature by establishing the legitimacy of uncertainty in the information environments of misinformation as a central indicator for the detection and prediction of misinformation during public health crises.

Second, this study delves into the intricacies of uncertain information environments for misinformation across 4 distinct scales, namely, the physical environment, macro-media environment, micro-communicative environment, and message framing. Our findings demonstrate the effectiveness of all 4 uncertainty features in misinformation detection and spread prediction. In contrast to prior misinformation literature during the COVID-19 pandemic, which often overlooked the role of the information environment in increasing the likelihood of misinformation dissemination, our research emphasizes the importance of considering uncertainty beyond the content of misinformation itself, such as ambiguous wording [29,50]. Our study broadens the concept of linguistic uncertainty in misinformation message framing to encompass a more comprehensive uncertainty across various information environments. We define uncertainty in information environments using a multiscale approach that highlights the significance of the interaction between the physical environment and macro-/micro-media environments. This approach diverges from focusing on a single dimension, such as ambiguities about official guidelines and news reports [18], or the misinformation framing strategy on social media [29].

Third, our findings indicate that uncertainties in information environments play a crucial role as motivators for the emergence and spread of misinformation. While previous studies have provided preliminary evidence suggesting that uncertainty stemming from government policies and news media could coincide with the occurrence of related misinformation during the COVID-19 pandemic, often relying on descriptive big data analyses [3,32], our study contributes stronger empirical evidence. We leverage machine learning techniques to demonstrate that uncertainty arising from the crisis and crisis communication through media can indeed incentivize individuals to generate and disseminate misinformation. Significantly, our findings revealed that the algorithm achieved its best performance for both detection and spread prediction tasks when incorporating items from the information environments published 1 day before the post (T=1). This discovery emphasizes the acute impact of uncertainty in the information environment on the emergence and spread of misinformation, underscoring the importance of timely uncertainty reduction in crisis communication. Furthermore, the algorithm attained the highest accuracies when it included items highly relevant to the post but with an appropriate size (r=0.1). This rationale is reasonable, as a too-small r may fail to encompass enough misinformation-related items, while a larger r might include a significant amount of irrelevant information. The evidence theoretically establishes a connection between crisis communication research and misinformation research, reinforcing the notion that crisis communication and misinformation containment are 2 intertwined aspects of crisis management [3].

This study offers significant practical implications for misinformation detection and spread prediction. First, unlike previous studies that separately investigated computational frameworks for these tasks [24,29], this study introduces a unified uncertainty–based framework capable of addressing both tasks simultaneously. Second, our framework operates instantaneously, as it only requires easily accessible data such as posts, mainstream news, and relevant social media discussions published a few days prior. Moreover, the uncertainty detection algorithm has been trained using external data, rendering our algorithm easy to implement and capable of providing timely detection and prediction for streaming textual data. Third, this study affirms the effectiveness of uncertainty in various information environments for detecting and predicting misinformation on social media. Hence, the 4 proposed uncertainty components in information environments could be leveraged by social media platforms to improve the accuracy of misinformation detection and spread prediction, thereby safeguarding individuals from harm caused by infodemic. The benefits offered by our algorithm may serve as an impetus for integrating uncertainty components into practical systems.

This study is the first to incorporate the uncertainty present in the information environment of a post for both misinformation detection and spread prediction. However, it has some limitations. First, our framework concentrated solely on text-only detection and prediction. Future work should extend the framework to incorporate multimodal and social graph–based detection. Second, we used an uncertainty detection algorithm developed from a generic corpus sourced from Wikipedia. Nevertheless, past research has indicated that expressions of uncertainty may vary slightly across domains [53]. In other words, uncertainty expressions in the context of the COVID-19 pandemic may differ from those in general situations. Therefore, future work should aim to enhance our uncertainty measure by utilizing a corpus specifically designed for uncertainty detection in the discourse related to COVID-19.

We introduced an EUP framework for both misinformation detection and spread prediction. Our framework delves into uncertainty within information environments across 4 scales: the physical environment, macro-media environment, micro-communicative environment, and message framing. The experiments demonstrated the effectiveness of our proposed uncertainty components in enhancing the performance of existing models. There are several directions for further investigation and extension of this work. First, we can explore the impact of different news and social media environments (eg, biased vs neutral; left wing vs right wing) on the emergence and spread of misinformation. Second, extending our algorithms to include multimodal misinformation detection could be beneficial, as misinformation increasingly incorporates images and videos. Third, investigating the interaction between misinformation detection and spread prediction using a multitask, transfer-learning model is a promising avenue, given the shared uncertainty framework identified in this study for both tasks.