From alternative conceptions of honesty to alternative facts in communications by US politicians

The paper examines how conceptions of honesty in political communication—specifically belief speaking (appeals to personal belief, intuition, and feelings) versus fact speaking (appeals to evidence, accuracy, and truth-seeking)—have evolved in US political discourse and how they relate to the spread of misinformation. Against a backdrop of democratic backsliding and increasing misinformation online, the authors note that some politicians are perceived as honest for ‘speaking their mind’ despite factual inaccuracy, as exemplified by persistent perceptions of honesty regarding President Trump despite extensive fact-check records of falsehoods. The authors disentangle honesty (a social and moral quality) from truth (a property of information), arguing that these map onto belief speaking and fact speaking, respectively. They focus on US Congress members’ Twitter communications (2011–2022) to address research questions: whether belief and fact speaking can be detected in political statements; how they evolve over time; partisan differences; and whether these honesty components relate to the trustworthiness of information shared. The United States is chosen due to its partisan polarization and Twitter’s prominence in political agenda-setting.

The authors situate their work in literature on misinformation’s societal impacts, including behavioral change, voting for populist parties, and links to hate crimes. They discuss psychological attributes of misinformation, such as persistence despite corrections and the potential valorization of norm-violating dishonesty as ‘authenticity.’ They review conceptual work on ontologies of political truth and honesty distinguishing belief speaking and fact speaking, noting historical antecedents (e.g., 1930s fascism’s intuition-based ‘organic truth’) and contemporary parallels in postmodern theory and populism. They emphasize democracy’s need for shared factual knowledge and leadership committed to fact speaking. Prior studies show conservative publics are more exposed to untrustworthy information and that leadership cues can shape public beliefs (e.g., climate change polarization). The study also connects to computational text analysis, dictionary-based measures, and sentiment/linguistic feature comparisons (LIWC, VADER).

Data: The team scraped tweets from members of both houses of the US Congress (official, staff, and campaign accounts) from 6 Nov 2010 to 31 Dec 2022 using Twitter API v2 via twarc. They compiled 1,278 unique Twitter handles across the 114th–118th Congresses; 220 accounts were inaccessible (deleted/suspended/private). The raw corpus comprised 5,914,107 tweets (3,463,409 original, 531,289 quote tweets, 575,044 replies, 1,351,346 retweets). Retweets were removed; exact duplicates were removed; only tweets with more than 10 words were retained, yielding a final corpus of 3,897,032 tweets. Party affiliation and account metadata were linked.

Honesty dictionaries: Using computational grounded theory, the authors curated two dictionaries capturing belief speaking and fact speaking. Seed words (e.g., belief speaking: believe, opinion, feel; fact speaking: evidence, examine, truth) were expanded using fastText embeddings and colexification networks, then pruned to remove duplicates/overlap/inflections. Human validation via Prolific (N=50–51 across tasks) rated each keyword’s representativeness on Likert scales; paired t-tests determined category assignment. Final dictionaries included 37 keywords each (Extended Data Table 1).

Semantic similarity scoring: Following distributed dictionary representation, the authors embedded dictionary words with GloVe (840B Common Crawl) and computed a dictionary vector for each honesty component. Tweets (and later, article texts) were embedded by averaging token vectors; cosine similarities to each dictionary yielded belief-speaking (Db) and fact-speaking (Df) similarities in [-1,1]. To correct for similarity’s correlation with text length (r=0.37 for Db; r=0.42 for Df) and the 2017 increase in Twitter’s character limit, Db and Df were centered and length-corrected by regressing on length and subtracting predictions. Robustness checks used word2vec and fastText embeddings, and perturbation tests removed 20% of dictionary words at random (100 iterations); effects remained significant and directionally stable.

Validation: Document-level validation sampled tweets from extreme quartiles for each component and neutral tweets; Prolific raters (N=50 after attention check) provided ground-truth labels. Classifier performance achieved AUC=0.824 (belief speaking) and AUC=0.772 (fact speaking). External validation applied dictionaries to 809,271 NYT archive abstracts categorized as opinion (240,567), politics (518,123), and science (50,581): science aligned most with fact speaking; opinion aligned most with belief speaking. Comparisons with LIWC-22 and VADER showed modest correlations, indicating conceptual distinctness: both components negatively correlated with LIWC ‘analytic’ (r=-0.27 belief; r=-0.16 fact), positively with ‘authentic’, ‘moral’, and negative sentiment; positive sentiment slightly positive for belief (r=0.06) and slightly negative for fact (r=-0.01) (all P<0.001).

Topic/word keyness: Scattertext identified words characteristic of party and honesty dimensions, visualized via scaled F-scores mapping to a 2D space.

Link analysis and NewsGuard: For misinformation linkage, they analyzed tweets containing at least one link (2,700,539 tweets), expanding to 2,844,901 link-level entries. Domains were matched to NewsGuard’s trustworthiness database (6,860 English-language domains, 9-criteria scoring to 100; <60 deemed ‘not trustworthy’). After excluding social media/search domains, NewsGuard covered 20–60% of links over time with no partisan coverage differences. An independently compiled outlet-reliability list was also used for robustness.

Article scraping: Using Newspaper3k, they scraped text from linked articles; after excluding <100-word texts and duplicates, and removing articles shared by both parties (2,462; 0.91%), 261,765 unique article texts remained (about 65% scraping coverage overall; 65% for trustworthy, 82% for untrustworthy links).

Statistical modeling: For tweet-level link quality, a linear mixed-effects model predicted rescaled NewsGuard score SNG ~ 1 + Db × Df + Db × Df × Party + (1 + Db × Df | userID). For article-level text, OLS modeled SNG ~ 1 + Db × Df + Db × Df × Party. Additional analyses checked effects of home-state 2020 voting patterns (no effect) and an independent reliability list (consistent results).

• Detection/validation of honesty components: Human ratings confirmed the dictionaries differentiate belief speaking and fact speaking at both keyword and document levels; classifier AUC=0.824 (belief) and 0.772 (fact). NYT validation showed science abstracts more similar to fact speaking (Δ≈+0.033 vs overall) and opinion more similar to belief speaking (Δ≈+0.013 vs overall), consistent with expectations.
• Partisan and temporal trends: From 2011–2013 to 2019–2022, mean belief-speaking similarity increased for Democrats (from -0.031 to 0.017; t(558)=-11.317, P<0.001, d=0.850, 95% CI [-0.06,-0.04]) and Republicans (from -0.040 to 0.012; t(493)=-10.819, P<0.001, d=0.854, 95% CI [-0.06,-0.04]). Fact-speaking similarity also increased for Democrats (from -0.027 to 0.009; t(516)=-9.753, P<0.001, d=0.748, 95% CI [-0.04,-0.03]) and Republicans (from -0.038 to -0.003; t(483)=-8.442, P<0.001, d=0.671, 95% CI [-0.04,-0.03]). Increases were pronounced post-2016.
• Association with information quality (tweet text predicting linked domain quality): Mixed-effects model (504,809 observations) showed significant fixed effects: Db (coef 0.022 for Democrats baseline; t=3.6, P<0.001), Party=Republican (coef -0.069; t=-29.9, P<0.001), Party×Db (coef -0.128; t=-14.4, P<0.001), Party×Df (coef 0.085; t=9.6, P<0.001), and the three-way interaction Db×Df×Party (coef -0.085; t=-5.6, P<0.001). Interpreting effects: a 10% increase in belief-speaking similarity predicted a 12.8-point decrease in NewsGuard score for Republicans; for Democrats there was no significant negative relationship. A 10% increase in fact-speaking similarity predicted an increase of 2.1 points for Democrats and 10.6 points for Republicans.
• Association with information quality (article text predicting domain quality): OLS on 261,765 articles confirmed a significant negative association for Republicans and belief speaking (Party×Db coef -0.540, t=-33.8, P<0.001) and a positive association for fact speaking (Df coef 0.026, t=2.3, P=0.003; Party×Df coef 0.110, t=6.0, P<0.001). A small negative Df effect for Democrats was observed (coef -0.065, t=-6.6, P<0.001). Three-way interaction Party×Db×Df was significant (coef -0.594, t=-14.8, P<0.001).
• Partisan baseline differences: Republicans shared lower-quality sources overall than Democrats (tweet-level Party coef -0.069; article-level Party coef -0.099; both P<0.001).
• Topics: Controversial topics often featured heightened belief or fact speaking relative to average; vaccine discourse showed less belief speaking than topics like climate change or the opioid crisis.
• Robustness and external factors: Results replicated with an independent outlet-reliability database; no effect of 2020 home-state voting patterns on link quality. Dictionary perturbations and alternative embeddings yielded consistent effect directions and significance.
• Public consequences: No electoral penalty detected for belief-speaking-based low-quality sharing (no association with Trump’s 2020 vote share at state level).

The study demonstrates that belief speaking and fact speaking are measurable, distinct components of political communication, both increasing over the past decade, especially after 2016. Crucially, the components relate differently to the quality of information shared: among Republicans, higher belief speaking strongly predicts sharing lower-quality sources, whereas fact speaking predicts higher-quality sharing for both parties. This helps explain asymmetric exposure to untrustworthy information found in prior public-level studies and suggests elite cues from political leaders may shape partisans’ information diets. The findings support the hypothesis that an alternative conception of honesty—authenticity via belief speaking—can facilitate dissemination of low-quality or false information without electoral penalty. Mediation analyses reported in the Supplementary suggest that negative emotionality and outgroup derogation, which correlate with belief speaking, may drive the association between belief speaking and low-quality sharing among Republicans. At the same time, Republican members engaging in fact speaking approach Democrats’ accuracy levels, indicating that rhetorical style, not ideology per se, is pivotal. Overall, the work links rhetorical honesty ontologies to misinformation spread, highlighting the democratic importance of fact-speaking leadership to sustain shared evidence-based reality.

The paper introduces and validates computational measures of two honesty ontologies—belief speaking and fact speaking—and applies them to a comprehensive corpus of US Congress tweets and linked news content. Both honesty modes have increased over time, but they diverge in their relationship to information quality: belief speaking among Republicans is associated with lower-quality sources, while fact speaking correlates with higher-quality sources across parties. These results provide a framework for understanding how elite rhetoric can shape misinformation dissemination and public perceptions of honesty. Future work should: (1) examine generalizability beyond the United States by analyzing other countries and party systems; (2) assess temporal stability amid evolving platforms (e.g., post-2022 Twitter changes) and political dynamics; (3) directly measure public perceptions of honesty to link rhetorical styles with perceived integrity; and (4) further disentangle the role of social media versus mainstream media in facilitating belief- versus fact-speaking discourse.

• Observational, correlational design precludes causal inference about the effects of belief speaking on misinformation sharing.
• Focus on US members of Congress and their staff/campaign accounts limits generalizability to other countries, political systems, or non-elite discourse.
• Platform-specific constraints and changes (e.g., Twitter’s character limit change in 2017; ownership changes in 2022) may influence linguistic measures and content sharing dynamics.
• Trustworthiness estimation relies on domain-level scores (NewsGuard and an independent list) rather than article-level fact-checks; although robust across sources, coverage was 20–60% of shared links and article scraping achieved ~65% coverage.
• The study cannot assess perceived honesty directly; electoral analyses found no penalty but do not measure voter perceptions.
• Potential confounds (topic selection, audience targeting, media ecosystem shifts) cannot be fully controlled; data distributions were assumed normal in regressions without formal tests.