Political Fact-Checking Efforts are Constrained by Deficiencies in Coverage, Speed, and Reach

Damian Hodel; Kayla Duskin; Morgan Wack

REVIEW 4 major objections 6 minor 9 references

Political Fact-Checking Efforts are Constrained by Deficiencies in Coverage, Speed, and Reach

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Institutional fact-checking during the 2022 U.S. midterms covered fewer than half of prominent election rumors, lagged four days behind them, and reached barely one percent of the conversation.

desk verdict A genuinely new measurement of fact-checking limits built on an independent rumor corpus; the headline numbers are plausible but the fact-check identification needs a sensitivity check before they are treated as exact. read the letter →

arxiv 2412.13280 v1 pith:BIM4M672 submitted 2024-12-17 cs.SI cs.CY

Morgan Wack , Kayla Duskin , Damian Hodel This is my paper · ORCID

classification cs.SIcs.CY

keywords fact-checkingmisinformationelectionrumorscoveragespeedreachpartisanshipX/Twitter

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that institutional fact-checking, as practiced during the 2022 U.S. midterm elections on X/Twitter, was too selective, too slow, and too partisan-bounded to function as the primary defense against political misinformation. Drawing on a real-time dataset of 135 prominent false or misleading election narratives, the authors find that fewer than half were ever fact-checked; for those that were, the median fact-check appeared four days after the narrative first surfaced and after 79% of the narrative's misinformation posts had already been published; and fact-check posts made up under 1.2% of narrative-related conversation, mostly circulating within a single partisan community. Once claim type and timing are held constant, the data show little evidence of right-leaning bias in coverage; the apparent partisan asymmetry tracks the kinds of claims each side produced. If these findings hold, fact-checking as currently organized cannot carry the burden of correcting election misinformation, and faster, platform-embedded or better-coordinated corrections are needed.

What carries the argument

The load-bearing comparison is between two corpora built independently: ElectionRumors2022 (a real-time, manually curated dataset of 135 false or misleading election narratives on X/Twitter during the 2022 midterms) and a manually collected set of 164 fact-checks found by Google searches for each narrative. To that is added a partisanship assignment for users and narratives built from co-engagement networks of reposts, plus a logistic regression of which narratives get fact-checked. Together these allow the paper to measure coverage (share of narratives with at least one fact-check), speed (delay from a narrative's first post to its first fact-check), and reach (share of posts containing fact-check links and the partisanship of users sharing them).

What would settle it

A fully independent census of 2022 midterm election misinformation—assembled without real-time observer judgment, for example by mining a complete archive of election-related posts with a different narrative-clustering method or by using post-election retrospective media reviews—that found more than half of prominent narratives were fact-checked, or that the median lag was under 24 hours, would contradict the core coverage and speed claims. A narrower test: locate a widely shared voter-suppression narrative from the period that was quickly fact-checked by multiple organizations; the paper's finding that suppression narratives are checked at 19% and largely ignored would need revision.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the three practical constraints—coverage, speed, and reach—compound to make institutional fact-checking a marginal force in a live U.S. election. Only 63 of 135 prominent election rumors (47%) received any fact-check; the median first fact-check came four days after the rumor's first appearance and three days after peak activity, with 79% of related posts already published; and among posts about even fact-checked narratives, only 1.2% contained or linked to a fact-check. Sharing of fact-checks was heavily partisan: left-leaning users posted 85% of fact-checks for left-leaning narratives and 80% for right-leaning narratives, while right-leaning users posted about 10%; only 0.30% of right-leaning users and 4.15% of left-leaning users shared any fact-check. The paper also shows that selection into fact-checking tracks claim type and timing more than virality or influencer involvement: all improbable-result narratives were checked, while only 19% of voter-suppression narratives were, and post-election narratives were checked more often. It reads the near-zero partisan coefficient after controls as evidence against the claim that fact-checking is systematically biased against the political right.

Load-bearing premise

The results depend on the 135 narratives in ElectionRumors2022 being a complete and unbiased record of prominent election misinformation; if the real-time observation missed major narratives or systematically excluded certain communities, every headline rate—47% coverage, four-day delay, and the partisan comparisons—would shift.

Editorial extensions

If this is right

If the 47% coverage figure is correct, reliance on institutional fact-checks alone leaves more than half of prominent election rumors publicly unrebutted in the critical window.
A median four-day delay with 79% of posts already out implies the modal fact-check functions as a record of a rumor rather than a timely correction.
Fact-check posts being 1.2% of the conversation and mostly within one partisan community means the audiences most exposed to a rumor rarely encounter the correction.
Using fact-check archives as a sampling frame for misinformation research inherits a bias toward easily debunkable, post-election narratives and overstates partisan asymmetry.
Coverage decisions driven by claim type rather than virality or influencer engagement suggest that important rumors, such as voter-suppression claims, are systematically under-served.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the same coverage-speed-reach triad is a ready-made template for evaluating platform-embedded corrections, such as community notes, in the same 2022 data or in later elections.
Editorial inference: because suppression narratives are the least-covered type, a concrete reform—dedicating a share of fact-checking capacity to first-person, hard-to-verify claims—would directly target the largest coverage gap.
Editorial inference: the reach figure was measured on X/Twitter only; corrections attached to posts rather than circulated as links might cross partisan lines differently, so transferring the 1.2% number to other platforms is an extrapolation, not a finding.
Editorial inference: the archive-bias result implies that prior misinformation studies sampling from fact-check databases may have overestimated effect sizes conditional on claim type; re-running such studies on fact-check-independent corpora would show whether their conclusions survive.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The thing to know: this paper measures fact-check coverage, speed, and reach starting from an independent rumor corpus rather than from fact-check archives. That design choice is the real contribution, and it makes the headline numbers worth engaging with.

The good parts are real. ElectionRumors2022 gives the authors 135 narratives and 1.8M posts collected in real time, and they manually validated the earliest post for each narrative so the timing analysis has a solid anchor. The results — 47% of narratives ever fact-checked, a four-day median delay between first post and first fact-check, 1.2% fact-check share of conversation — are plausible and, as far as I know, new as a package. The partisan finding is more nuanced than the raw percentages suggest: right-leaning narratives are fact-checked more often, but once you control for claim type, the gap mostly disappears. That is a useful corrective to the 'moderation is biased against the right' narrative in public debate.

The soft spots are exactly where your reader flags them. Fact-check identification relies on Google search, restricted to MBFC high-credibility sources and the literal term 'fact-check'. That will miss corrections, which means the 47% is a lower bound on coverage and the four-day delay is an upper bound on speed. The 1.2% reach figure only counts posts containing a link to an identified fact-check, so it too is a lower bound. The denominator — the 135 narratives — was manually curated in real time, and the paper admits very low-spread cases were excluded; the abstract says 'prominent', but the precision of the percentages should not be oversold. The logistic regression also shows signs of perfect separation (the neutral and improbable-results coefficients are around ±16 with standard errors under 1), so the OLS re-estimation is a sensible check and should perhaps be the primary specification for those predictors. Finally, the arXiv version contains a data-availability promise but no code or data. All of these are addressable rather than fatal.

Who is this for: misinformation researchers, fact-checking practitioners, and platform policy staff. The qualitative conclusion — institutional fact-checking is partial, slow, and partisan-bounded — survives the identification worries. The exact magnitudes need a sensitivity check against a comprehensive fact-check archive and a public code/data deposit.

Recommendation: send it to review. It deserves a serious referee. I would ask for the sensitivity analysis, confidence intervals, and data/code before accepting.

Referee Report

4 major / 6 minor

Summary. The paper evaluates real-world constraints on institutional fact-checking during the 2022 U.S. midterm elections by combining three data sources: ElectionTweets2022 (446 million election-related posts), ElectionRumors2022 (135 manually curated false/misleading narratives), and a new corpus of 164 fact-checks identified by Google search with Media Bias Fact Check (MBFC) source filtering. The central empirical claims are that only 47% of prominent narratives were fact-checked, that the median first fact-check appeared four days after a narrative's first post and after 79% of its misinformation posts, and that fact-check-linked posts constitute only 1.2% of the narrative conversation and circulate almost exclusively within partisan communities. The paper also argues that partisan asymmetry in coverage attenuates once narrative type is controlled, using logistic and OLS regressions with robustness checks on virality and influencer thresholds.

Significance. If the headline estimates hold, the paper makes an important contribution: it measures fact-checking effectiveness in a naturalistic setting using a rumor corpus that is not built from fact-check archives, thereby avoiding the selection/circularity problem that affects much of this literature. The study is also transparent in reporting manual validation of narrative timelines, robustness checks on variable cut-offs, and a stated intention to release code and data, all of which strengthen confidence in its descriptive findings. The coverage, speed, and reach magnitudes, should they survive sensitivity analysis, would be directly relevant to debates about whether institutional fact-checking can serve as a primary correction mechanism.

major comments (4)

[Data & Methods, Fact-check Identification (p. 9)] The headline coverage figure of 47% is sensitive to the completeness of fact-check identification, which relies on the first ten Google results, a requirement that pages use the term "fact-check," and MBFC ratings of unbiased/high credibility. A debunk that does not surface in that search, is published by an organization MBFC rates as biased, or is formatted as a news correction or platform label would be missed, and if any such debunk corresponds to a narrative coded as "not fact-checked," the 47% figure is inflated and the four-day median delay is lengthened. The manuscript needs a sensitivity analysis that varies these inclusion rules, for example by (a) matching narratives against IFCN/International Fact-Checking Network signatory databases or the Duke Reporter's Lab list, (b) relaxing the MBFC filter to include all fact-checking organizations regardless of MBFC bias rating, and (c) reporting how many "not fact-checked" narratives have at least one candidate correction from these broader sources. This is load-bearing for the paper's central claim.
[Data & Methods, False and Misleading Narratives (p. 8)] The 47% coverage rate and the four-day median delay are conditional on the completeness of the ElectionRumors2022 denominator, which was constructed in real time by manual observation and which the authors state eliminated "very low-spread" cases. If the curation missed prominent narratives or systematically underrepresented particular communities or types of claims, the coverage, speed, and partisan-comparison results all shift, and the direction of the bias is not obvious a priori. The paper should state this limitation explicitly in the main text and provide an external validation check, such as comparing the 135-narrative corpus against an independently constructed list of prominent election-claims (e.g., from the Election Integrity Partnership or media casebooks), including how many narratives overlap and whether the coverage rate changes on the union of the two lists.
[Results, Coverage; SI Table 3 (logistic regression)] The logistic regression in Table 3 shows signs of quasi-complete separation: the coefficients for "Partisanship: Neutral" (-16.036, SE 1.239) and "Classification: Improbable" (16.484, SE 0.832) are implausibly large with small standard errors, which is a classic separation artifact rather than a meaningful estimate. Consequently, the claim in the Results that partisanship has "limited evidence" of influencing coverage rests on unstable model estimates, especially given only 3 neutral narratives and 9 improbable narratives. The authors should re-estimate with Firth's penalized likelihood or a Bayesian model, or present the OLS specification in Table 6 as the primary model, and should note in the text that the "Improbable" and "Neutral" coefficients are not interpretable as finite odds ratios.
[Results, Reach (pp. 16-18)] The reach analysis defines a fact-check post only as a post containing a link to a URL of an identified fact-check. This measure excludes quotes, replies, and text-based corrections that name or describe the fact-check without a URL, and the Discussion later acknowledges that "textual corrections which do not link to an external fact-check" are outside the analysis. As a result, the 1.2% conversation share and the 4.15%/0.30%/0.38% sharing rates are lower bounds, not point estimates. The main text should state this directional bias explicitly, and ideally the authors should hand-code a sample of posts that discuss fact-checks without linking to quantify how much the reach estimate would change. The Discussion's statement that "fewer than 2% of users ... also shared a fact-check" should also be reconciled with the Table 2 percentages, which use different denominators.

minor comments (6)

[p. 3, Introduction] The citation "election period Schafer et al., 2024" is missing an opening parenthesis; it should read "election period (Schafer et al., 2024)."
[p. 12, Results, Coverage] The text says "at least 100,00 followers" and should read "100,000."
[p. 15, Results, Speed] The sentence beginning "While this outcome likely comes of no surprise to fact-checkers, who prioritize ... based on their as their human and technological resources" contains a duplicated "as their" and should be rephrased.
[Figure 2 caption] The caption refers to the "modal fact-check (red)" in one place and to the "aggregated (median) fact-check response time (vertical line)" in another, which is confusing about whether the vertical line marks the mean, median, or mode; the text and legend should use one consistent statistic.
[Abstract and Discussion] The abstract's phrase "most comprehensive assessment to date" is an overclaim for a single-platform, single-election study with a manually curated corpus; I suggest tempering it to something like "a comprehensive assessment" or "one of the first campaign-wide assessments."
[References] Several author names appear with a spacing artifact (e.g., "V ogels" for Vogels, "V osoughi" for Vosoughi, "V ." for V.). These should be corrected for a publication version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: fact-check coverage, speed, and reach are computed by matching an independently curated rumour corpus against a new fact-check collection, with no fitted parameter renamed as a prediction.

full rationale

The derivation chain is self-contained. Coverage (63/135 narratives), speed (median four-day delay), and reach (1.2% of posts; 0.30-4.15% of users sharing) are all computed by matching the ElectionRumors2022 corpus against a fact-check set newly collected for this paper via Google queries and Media Bias Fact Check filtering. The decisive anti-circularity design choice is explicit: "Unlike prior studies on online misinformation, we start from a baseline set of false and misleading election claims that are not determined by their prior fact-check status." The outcome variables - whether a fact-check exists, when it was published, and how often it was shared - were not inputs to the corpus construction, which was real-time observation with duplicate, over-broad, and very low-spread cases removed. The one self-citation overlap (Duskin and Wack also appear on the ElectionRumors2022 paper, and the Acknowledgements thank Joseph Schafer for data help) supplies the denominator but qualifies as independent support rather than circularity: the dataset is a codified, externally auditable artifact whose construction did not include the target results, and the present conclusions do not reduce to the citation. The logistic regression (SI Equation 1) is descriptive and in-sample; "predicted probability" is standard marginal-effects language, and no parameter is fitted to a subset and then reported as a prediction of a closely related quantity. The reach operationalization, counting as fact-check posts only those linking to identified fact-check URLs, is a construct definition, and the paper candidly bounds it in the Discussion: the analyses "do not address the full set of corrective interventions available to online platforms, including most notably textual corrections which do not link to an external fact-check as well as crowd-sourced alternatives." That is a measurement-coverage caveat (a possible undercount), which belongs under correctness risk rather than circularity, as do the Google/MBFC filtering choices and reliance on the completeness of the 135-narrative corpus. No uniqueness theorem is imported, no ansatz is smuggled in via citation, and no known result is renamed; the Table 1 versus SI inconsistency in the count of "improbable" narratives (9 versus 10) is a data-consistency issue, not a circular step.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The headline estimates are direct counts rather than outputs of a fitted model, so the free-parameter ledger contains only the design thresholds that shape partisanship and engagement measures. The axioms are the data-construction assumptions without which the coverage, speed, and reach percentages would not be interpretable. No new physical or theoretical entities are introduced; the new artifacts are datasets, not explanatory constructs.

free parameters (3)

Partisan label threshold = 80% of reposts to a single partisan influencer cluster
Used to label users as left-leaning, right-leaning, or neutral in the reach analysis; the 4.15%/0.30%/0.38% fact-check sharing statistics and the partisan composition in Table 2 depend on this choice. No robustness check for this threshold is reported.
Influencer follower threshold = 100,000 followers
Used to define high influencer involvement in the coverage regression (with a 250K robustness check); it is a design choice, not a fitted constant.
Virality window = 24 hours (with a 10-hour robustness check)
Used to define high virality narratives; the conclusion that virality does not predict fact-checking persists across the 10h/24h cutoffs in the SI.

assumptions (5)

domain assumption ElectionRumors2022 is a complete and accurate enumeration of prominent election misinformation narratives on X/Twitter in the 2022 US midterm period.
The coverage denominator (135 narratives) and all timeline measurements come from this dataset, whose construction involved real-time observation and manual curation by a multi-stakeholder group (Schafer et al. 2024). If narratives were missed, the 47% coverage figure is wrong. Section: Data & Methods, False and Misleading Narratives.
domain assumption Google search with a 'fact check' query on the first ten results is a valid proxy for the accessibility of every relevant institutional fact-check.
Fact-checks were identified by manual Google queries and filtered by MBFC ratings; fact-checks not discoverable that way are excluded, which directly affects the coverage and speed estimates. Section: Data & Methods, Fact-check Identification.
domain assumption MBFC bias and credibility ratings are a valid filter for 'unbiased' fact-check sources.
The inclusion criteria require sources rated neutral and high credibility by Media Bias Fact Check. If MBFC is unreliable, some fact-checks are misclassified. Section: Data & Methods, Fact-check Identification.
domain assumption Influencer partisanship labels derived from co-engagement networks and the 80% repost threshold are accurate enough for the reach analyses.
User partisanship is inferred from reposts of manually validated influencers; the 80% threshold and the manual validation are not accompanied by inter-coder reliability or external validation. Section: Data & Methods, User and Narrative Partisanship Labeling.
domain assumption Manual validation that the earliest narrative-linked post is really related to the narrative correctly establishes first-appearance times.
The four-day median delay depends on these manually validated timestamps. Section: Data & Methods, False and Misleading Narratives.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Political Fact-Checking Efforts are Constrained by Deficiencies in Coverage, Speed, and Reach." pith.science (2026). https://pith.science/paper/BIM4M672

@misc{pith2026241213280,
  author       = {Pith},
  title        = {Pith review of: Political Fact-Checking Efforts are Constrained by Deficiencies in Coverage, Speed, and Reach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIM4M672}},
  note         = {Machine review of arXiv:2412.13280}
}

read the original abstract

Fact-checking has been promoted as a key method for combating political misinformation. Comparing the spread of election-related misinformation narratives along with their relevant political fact-checks, this study provides the most comprehensive assessment to date of the real-world limitations faced by political fact-checking efforts. To examine barriers to impact, this study extends recent work from laboratory and experimental settings to the wider online information ecosystem present during the 2022 U.S. midterm elections. From analyses conducted within this context, we find that fact-checks as currently developed and distributed are severely inhibited in election contexts by constraints on their i. coverage, ii. speed, and, iii. reach. Specifically, we provide evidence that fewer than half of all prominent election-related misinformation narratives were fact-checked. Within the subset of fact-checked claims, we find that the median fact-check was released a full four days after the initial appearance of a narrative. Using network analysis to estimate user partisanship and dynamics of information spread, we additionally find evidence that fact-checks make up less than 1.2\% of narrative conversations and that even when shared, fact-checks are nearly always shared within,rather than between, partisan communities. Furthermore, we provide empirical evidence which runs contrary to the assumption that misinformation moderation is politically biased against the political right. In full, through this assessment of the real-world influence of political fact-checking efforts, our findings underscore how limitations in coverage, speed, and reach necessitate further examination of the potential use of fact-checks as the primary method for combating the spread of political misinformation.

Figures

Figures reproduced from arXiv: 2412.13280 by the authors.

**Figure 1.** Predicted probability of being fact-checked in the 2022 Midterm Election. Additional classifications are excluded for brevity but included in the SI. Points represent regression coefficients. Lines represent the 95% confidence interval. The dashed line represents the exact null effect. Black lines are statistically significant at α = .05. work by Mosleh et al. (2024) on Twitter moderation, once we control for addit… view at source ↗

**Figure 2.** Misinformation posts published before (black) and after (grey) the modal narrative was fact-checked (grouped by 12 hour totals). Fact-checks most often occur only after the typical rumor has already lost momentum, as indicated by aggregated (median) fact-check response time (vertical line). When aggregated across all 63 fact-checked narratives, we find that the median first factcheck was published a full four days… view at source ↗

**Figure 3.** Network visualization of online discussion of false and misleading narratives and their corresponding fact-checks. Nodes represent X/Twitter users while edges represent reposts between users. In Figure A, nodes are all colored gray while edges are colored based on the content of the repost. Red edges are fact-checks of right-leaning narratives while blue edges are fact-checks of left-leaning narratives. Gray edges a… view at source ↗

Figures from the paper (11 more)

**Figure 4.** Figure 4: Mechanisms of spread between users with different party leanings for posts discussing the false and misleading narratives and posts sharing fact-check links. Y-axis denotes the party (L - left-leaning, N - neutral/apolitical, R - right-leaning) of the posting user (e.g…

**Figure 5.** Figure 5: The plot illustrates the expected influence of influence involvement (total users with 100K+ followers) on Fact-Check probability [PITH_FULL_IMAGE:figures/full_fig_p033_5.png]

**Figure 6.** Figure 6: The plot illustrates the expected influence of narrative virality (¡24 hrs) on Fact-Check probability [PITH_FULL_IMAGE:figures/full_fig_p034_6.png]

**Figure 7.** Figure 7: The plot illustrates the expected influence of partisan alignment on fact-check probability. Numeric variables are held at their mean while factor variables are assigned their modal value from the dataset [PITH_FULL_IMAGE:figures/full_fig_p035_7.png]

**Figure 8.** Figure 8: The plot illustrates the expected influence of the timing of a narrative’s appearance on fact-check probability. Numeric variables are held at their mean while factor variables are assigned their modal value from the dataset [PITH_FULL_IMAGE:figures/full_fig_p036_8.png]

**Figure 9.** Figure 9: Classification of Narratives In viewing the results, the primary takeaways can best be viewed in the contrast between ”Improbable” narratives and ”Suppression” narratives. All ten narratives coded as ”Improbable” were fact-checked in our dataset, while only six of thi…

**Figure 10.** Figure 10: Cumulative posts, normalized between 1 and 0, before (in blue) and after (in green) the publication of the first fact-check. The bold line represents the mean across all fact-checked narratives [PITH_FULL_IMAGE:figures/full_fig_p040_10.png]

**Figure 11.** Figure 11: Number of fact-checks categorized into 3-day intervals, divided by partisanship. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_11.png]

**Figure 12.** Figure 12: Number of narratives categorized into 3-day intervals, divided by partisanship [PITH_FULL_IMAGE:figures/full_fig_p042_12.png]

**Figure 13.** Figure 13: Total Fact-checks for Each Narrative in the Dataset Duplicated efforts were common in the dataset. As seen in [PITH_FULL_IMAGE:figures/full_fig_p043_13.png]

**Figure 14.** Figure 14: OLS Recreation of Main Logit Analysis 48 [PITH_FULL_IMAGE:figures/full_fig_p048_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 6 canonical work pages

[1]

Allcott, H., & Gentzkow, M. (2017). Social media and fake news in the 2016 election. Journal of economic perspectives, 31(2), 211–236. Amazeen, M. A. (2016). Checking the fact-checkers in 2008: Predicting political ad scrutiny and assessing consistency. Journal of Political Marketing, 15(4), 433–464. Aslett, K., Godel, W., Sanderson, Z., Nagler, J., Bonne...

work page 2017
[2]

Kim, J. (2009). Describing and predicting information-seeking behavior on the web.Journal of the american society for information science and technology, 60(4), 679–693. Koch, T. K., Frischlich, L., & Lermer, E. (2023). Effects of fact-checking warning labels and social endorsement cues on climate change fake news credibility and engagement on social medi...

arXiv 2009
[7]

M., Pennycook, G., Berinsky, A

Brashier, N. M., Pennycook, G., Berinsky, A. J., & Rand, D. G. (2021). Timing matters when correcting fake news.Proceedings of the National Academy of Sciences, 118(5), e2020043118. Bruns, A. (2021). After the ‘APIcalypse’: Social media platforms and their fight against critical scholarly research. Disinformation and Data Lockdown on Social Platforms, 14–...

work page 2021
[9]

The odds ratios corresponding to this equation and partially displayed in Figure 1 in the main paper are reflected in the table below

=β0 + β1 × Partisanship + β2 × log(Virality) + β3 × log(Influencers) + β4 × Classification + β5 × Timing (1) where the output represents the binary outcome of whether a fact-check was produced or the specified narrative remained un-fact-checked (0/1), β0 is the intercept, β1 is a categorical vari- able with three levels denoting variations partisan leanin...

work page 2022
[19]

V osoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. science, 359(6380), 1146–1151. Vraga, E. K., & Bode, L. (2020). Correction as a solution for health misinformation on social media. American Journal of Public Health, 110(S3), S278–S280. Walter, N., Cohen, J., Holbert, R. L., & Morag, Y . (2020). Fact-checking: A meta-ana...

work page 2018
[115]

E., & Wilson, C

Jiang, S., Robertson, R. E., & Wilson, C. (2019). Bias misperceived: The role of partisanship and misinformation in youtube comment moderation. Proceedings of the International AAAI Conference on Web and Social Media, 13, 278–289. Kennedy, I., Wack, M., Beers, A., Schafer, J. S., Garcia-Camargo, I., Spiro, E. S., & Starbird, K. (2022). Repeat spreaders an...

work page 2019
[150]

Friggeri, A., Adamic, L., Eckles, D., & Cheng, J. (2014). Rumor cascades.Eighth International AAAI Conference on Weblogs and Social Media. Garrett, R. K., & Weeks, B. E. (2013). The promise and peril of real-time corrections to political misperceptions, 1047–1058. Godel, W., Sanderson, Z., Aslett, K., Nagler, J., Bonneau, R., Persily, N., & Tucker, J. A. ...

arXiv 2014
[237]

Rosenfeld, N., Szanto, A., & Parkes, D. C. (2020). A kernel of truth: Determining rumor ve- racity on twitter by diffusion pattern alone. Proceedings of The Web Conference 2020, 1018–1028. Saeed, M., Traub, N., Nicolas, M., Demartini, G., & Papotti, P. (2022). Crowdsourced Fact- Checking at twitter: How does the crowd compare with experts?Proceedings of t...

arXiv 2020

Show all 9 references

[2021]

https://www.pewresearch.org/internet/2021/04/07/social-media-use-in-2021/ Azevedo, L

Pew Research Center, 1, 1–4. https://www.pewresearch.org/internet/2021/04/07/social-media-use-in-2021/ Azevedo, L. (2018). Truth or lie: Automatically fact checking news. Companion Proceedings of the The Web Conference 2018, 807–811. Bardon, A. (2019). The truth about denial: ...

2018

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Allcott, H., & Gentzkow, M. (2017). Social media and fake news in the 2016 election. Journal of economic perspectives, 31(2), 211–236. Amazeen, M. A. (2016). Checking the fact-checkers in 2008: Predicting political ad scrutiny and assessing consistency. Journal of Political Marketing, 15(4), 433–464. Aslett, K., Godel, W., Sanderson, Z., Nagler, J., Bonne...

work page 2017

[2] [2]

Kim, J. (2009). Describing and predicting information-seeking behavior on the web.Journal of the american society for information science and technology, 60(4), 679–693. Koch, T. K., Frischlich, L., & Lermer, E. (2023). Effects of fact-checking warning labels and social endorsement cues on climate change fake news credibility and engagement on social medi...

arXiv 2009

[3] [7]

M., Pennycook, G., Berinsky, A

Brashier, N. M., Pennycook, G., Berinsky, A. J., & Rand, D. G. (2021). Timing matters when correcting fake news.Proceedings of the National Academy of Sciences, 118(5), e2020043118. Bruns, A. (2021). After the ‘APIcalypse’: Social media platforms and their fight against critical scholarly research. Disinformation and Data Lockdown on Social Platforms, 14–...

work page 2021

[4] [9]

The odds ratios corresponding to this equation and partially displayed in Figure 1 in the main paper are reflected in the table below

=β0 + β1 × Partisanship + β2 × log(Virality) + β3 × log(Influencers) + β4 × Classification + β5 × Timing (1) where the output represents the binary outcome of whether a fact-check was produced or the specified narrative remained un-fact-checked (0/1), β0 is the intercept, β1 is a categorical vari- able with three levels denoting variations partisan leanin...

work page 2022

[5] [19]

V osoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. science, 359(6380), 1146–1151. Vraga, E. K., & Bode, L. (2020). Correction as a solution for health misinformation on social media. American Journal of Public Health, 110(S3), S278–S280. Walter, N., Cohen, J., Holbert, R. L., & Morag, Y . (2020). Fact-checking: A meta-ana...

work page 2018

[6] [115]

E., & Wilson, C

Jiang, S., Robertson, R. E., & Wilson, C. (2019). Bias misperceived: The role of partisanship and misinformation in youtube comment moderation. Proceedings of the International AAAI Conference on Web and Social Media, 13, 278–289. Kennedy, I., Wack, M., Beers, A., Schafer, J. S., Garcia-Camargo, I., Spiro, E. S., & Starbird, K. (2022). Repeat spreaders an...

work page 2019

[7] [150]

Friggeri, A., Adamic, L., Eckles, D., & Cheng, J. (2014). Rumor cascades.Eighth International AAAI Conference on Weblogs and Social Media. Garrett, R. K., & Weeks, B. E. (2013). The promise and peril of real-time corrections to political misperceptions, 1047–1058. Godel, W., Sanderson, Z., Aslett, K., Nagler, J., Bonneau, R., Persily, N., & Tucker, J. A. ...

arXiv 2014

[8] [237]

Rosenfeld, N., Szanto, A., & Parkes, D. C. (2020). A kernel of truth: Determining rumor ve- racity on twitter by diffusion pattern alone. Proceedings of The Web Conference 2020, 1018–1028. Saeed, M., Traub, N., Nicolas, M., Demartini, G., & Papotti, P. (2022). Crowdsourced Fact- Checking at twitter: How does the crowd compare with experts?Proceedings of t...

arXiv 2020