How the crisis of trust in experts occurs on social media in China? Multiple-case analysis based on data mining

The paper addresses why and how a crisis of public trust in experts unfolds on Chinese social media during the COVID-19 policy transition period. With social media amplifying both scientific information and misinformation, the authors argue that effective science communication requires public trust in experts, yet such trust is fragile, particularly amid uncertainty and conflicting messages. They note that most prior research focuses on democracies, while China’s government-led communication context may shape trust differently. The study focuses on Weibo discourse from Nov 11, 2022 to Jan 22, 2023—a period of major policy relaxation—asking three research questions: (RQ1) What are public positions and main aspects of questioning experts across nine cases? (RQ2) Are these questions related to knowledge content and reporting characteristics, and are they justified? (RQ3) What roles do experts, media, and the public play in the trust crisis? The study aims to clarify mechanisms of public behavior after receiving expert-communicated knowledge and to improve social media science communication effectiveness.

The review synthesizes strands on trust in expert systems and science communication. Trust is framed as confidence under uncertainty (Giddens), with public trust in experts and expert systems interlinked; vulnerabilities at access points (e.g., individual experts, communicated knowledge) can erode trust in systems (Smith). Expert trustworthiness depends on competence, ethics, and public-interest orientation (O’Doherty; Duijf; Goldenberg). Information characteristics—scientific standards, transparency, understandability, and consistency—shape trust (Intemann; Gottschling; Nagler). On social media, the volume/velocity of information complicates authenticity judgments; enhancing expert credibility alone is insufficient. Some accounts attribute the trust crisis to public irrationality and postmodern/post-truth dynamics, elevating personal experience and emotive judgments over evidence, which can favor conspiracy thinking when expert messages conflict with lay experience (Levy; Ruser; Kien; Rughiniş & Flaherty). Media can also drive distrust by emotive packaging, fake/rumor content, or decontextualization to gain influence (Tandoc; Zhou & Zafarani; Figenschou & Ihlebæk; Irwin). During COVID-19, expert advice was pivotal but politicization and conflicting information reduced trust across contexts (Bhatia; Green; Latkin; Van der Linden). In China’s supervised social media environment, shifts in policy and messaging (e.g., downplaying severity post-relaxation) affected perceptions of expert credibility; Weibo enables visible public reactions and serves as data for topic and sentiment analysis.

Design: Multiple-case analysis of nine Weibo “Hot Search List” (HSL) cases involving expert-communicated COVID-19 knowledge between 11/11/2022 and 01/22/2023, a period spanning major policy relaxation. Case selection: Queried the Yunhe Data Platform HSL history using keyword “expert,” yielding 121 hashtags. Inclusion criteria—(1) High relevance (expert COVID-19 knowledge), (2) High discussion (≥6 h on HSL, top-3 peak, >5,000 historical comments, >0.5M influence), (3) High visibility (presenter’s original post and comments accessible; excluded deleted or comment-blocked posts). Nine cases were retained: C1 Omicron virulence ~ seasonal cold; C2 conditions maturing for Class B management; C3 massive epidemic impact in 1–2 months; C4 tangerines boiled in water as better remedy; C5 repeat infections double risk of death; C6 2–3 weeks recovery after turning negative; C7 Omicron mutation essentially saturated; C8 expert responds to perception of more elderly deaths; C9 do not visit elderly not yet positive during Spring Festival. Data collection: Wrote Python crawler to collect public comments under the presenter’s key post for each case (focus on the initial source to avoid drift introduced by re-analyses). After deduplication and removing non-text/meaningless items, final comment counts per case: 8507, 3146, 7455, 1945, 7561, 3210, 2623, 3886, 4302. Also recorded usernames and timestamps. Text mining pipeline: - Preprocessing: Custom dictionary, Jieba segmentation, stop-word removal, harmonize similar words, build comment words database. - Sentiment classification: Built a KNN-based classifier (K-Nearest Neighbor). Labeled training data were comments from posts either reproducing the presenter’s content or screenshots thereof. Trained the model to classify comments as positive/negative and applied it to presenter-post comments. - Topic modeling: LDA (Latent Dirichlet Allocation) on comment corpora per case. Determined optimal number of topics via perplexity, coherence, and visualization. Constructed topic meanings using top feature words and representative comments. Combined each comment’s topic probability with sentiment to assign a sentiment tendency to topics. Case feature coding: For each presenter post, coded basic characteristics: BC1 expert identity disclosed; BC2 audiovisual interview materials included; BC3 textual interview extracts included; BC4 disparity between post content and expert viewpoints. All presenters were media; only Case 3 lacked disclosed identity; interview evidence was provided across cases; exact alignment of post content with expert statements occurred in Cases 2, 6, 8, 9.

Overview: Across nine Weibo cases, public distrust emerged from intersecting roles of experts, media, and the public, shaped by the policy context. Media were the primary presenters of expert knowledge; their framing strongly influenced public reactions. Case-specific salient negative themes and sample proportions: - Case 1 (Omicron ~ cold): Negative focus on lack of rigor and data (Topic 3: 27%) and conflicting expert-system knowledge across stages (Topic 8: 13%). - Case 2 (Return to Class B): Calls to disclose expert identity/real-name reporting (Topic 3: 24%). - Case 3 (Massive impact 1–2 months): “Experts should not advise”/dismissals (Topic 4: 19%). - Case 4 (Tangerines remedy): Broad negativity; perceived experts as product promoters (Topic 1: 29%), price influence (Topic 2: 20%), adverse effects of tangerines (Topic 3: 28%), and vitamin C destroyed by heat (Topic 4: 23%). - Case 5 (Repeat infections double death risk): Conflicts across stages (Topics 3: 16%; 6: 9%) and demands for unified expert consensus (Topic 4: 23%); some accused media of creating panic/out-of-context reporting (Topic 5: 9%). - Case 6 (2–3 weeks recovery): Strong negativity on conflicting stage messages (Topic 3: 37%); many described lingering symptoms. - Case 7 (Mutation saturated): Doubts about applying foreign data to China (Topic 1: 15%), requests for data support (Topic 2: 26%), broad expert-system denial (Topic 3: 29%), and skepticism that mutations can be “saturated” (Topic 4: 16%). - Case 8 (More elderly deaths): “Not feelings but facts” (Topic 1: 22%), demand for detailed data (Topic 2: 26%), concerns about vaccine efficacy, perception that experts lack human touch (Topic 5: 15%). - Case 9 (Don’t visit uninfected elderly): Supportive comments coexist with negativity about common-sense advice (Topic 4: 11%), conflict with family expectations (Topic 5: 14%), and calls for experts to be silent (Topic 6: 36%). Cross-case patterns: - Knowledge conflicts: Persistent references to contradictory messages across stages (e.g., downplaying severity post-relaxation vs earlier emphasis on risks) reduced trust. - Common-sense perception: Advice viewed as obvious or trivial (e.g., not visiting uninfected elders) prompted dismissals of expert value. - Demands for rigor/data: Recurrent calls for transparent data and cautious, evidence-backed claims. - Media misalignment: In 5 of 9 cases, presenter posts departed from expert content through subject substitution, conservative word deletion/substitution, knowledge grafting, ignoring antecedents, or inverting causality; only in Case 5 did commenters widely identify media decontextualization. - Identity/authority signals: Requests for disclosed identity/real-name attribution (notably Case 2). - Political/context coupling: Perceived synchronization between expert messaging and policy changes fueled skepticism of “policy experts.” - Irrational/populist negativity: Basis-free denials and “experts shut up” refrains appeared (e.g., Cases 3, 7, 9).

The trust crisis is co-constructed by experts, media, and the public, and is sensitive to policy context. Conflicts within expert systems—natural in evolving science—were interpreted by the public as unreliability, especially amid rapid policy shifts, fostering polarization and anger that undermine adherence. In China, high alignment between expert messaging trends and government policy orientations prompted some to view experts as policy advocates, distinct from partisan polarization seen in democracies. Perceived common-sense advice and experts’ instrumental rationality (data-first framing without affective sensitivity) clashed with public value rationality, generating resentment, especially where advice conflicted with family norms (e.g., visiting elders). Blurred boundaries between experts and capital/administration encouraged suspicion of ulterior motives. Media acted as powerful intermediaries; misunderstandings or commercialized framing led to misrepresentation (e.g., Case 4’s tangerine narrative), yet blame defaulted to “experts say,” rarely to media. Public irrationality, emotional catharsis, and status-based mistrust amplified distrust in post-truth dynamics. Implications include: educating about the scientific process and normalcy of uncertainty/iteration; careful communication of uncertainty (balancing candor with clarity to avoid eroding authority); value transparency and “transparent subjectivity” about positions and interests; strengthening expert–media coordination and bidirectional feedback channels; and cultivating an expert image centered on public interest to mitigate status-based distrust.

Analyzing public comments under nine Weibo cases (11/11/2022–01/22/2023) showed that crises of trust in experts on social media arise from the combined roles of experts, media, and the public. Conflicting knowledge—though consistent with scientific development—fueled dissatisfaction. The public increasingly demands rigorous, data-supported, non-trivial advice and rejects knowledge perceived as common sense. Distrust grew when experts appeared instrumentally rational, or boundaries with administration/capital seemed blurred. Media’s unconscious processing (subject substitution, conservative word substitution, knowledge grafting, ignoring antecedents, inverting cause/effect) misled audiences and triggered misinterpretations. In the post-truth environment, some users’ irrational emotions amplified dissatisfaction with experts/systems. Improving trust requires coordinated efforts in how experts produce/present knowledge, how media frame it, and how publics engage with it.

- Scope of comments: Only presenter-post comments were analyzed; discussions under other users’ posts were excluded, limiting insight into how attitudes evolve over time and across the wider conversation. Future work will track all posts/comments under case hashtags to model topic evolution and trust dynamics. - Individual differences: Social media mining overlooks heterogeneity in experiences, culture, and political positions. Planned experimental designs will simulate social media knowledge communication and collect participant dialogues to link individual traits with expert-trust attitudes. - Context specificity: Findings reflect China’s supervised social media environment and a sudden policy relaxation context, which may heighten emotions/confusion. Normal-context cases should be studied to expand applicability. - Focus on extrinsic factors: Emphasis on content features and communication modes during a politicized public health crisis limits understanding of intrinsic factors (e.g., perceived political trust). Future experiments should test how political trust affects expert trust.