Why Does Misinformation Propagate Faster? An Algorithmic Perspective on X

Pan Li; Shuang Gao

cs.SISep 24, 2026

Why Does Misinformation Propagate Faster? An Algorithmic Perspective on X

Authors: Pan Li, Shuang Gao

Organizations: Arizona State University

Abstract

Misinformation is widely reported to propagate faster on engagement-based platforms, yet prior work largely focused on empirical analysis, without identifying a specific algorithmic mechanism that results in this phenomenon. Thanks to the open-sourcing of X's recommendation algorithms, we conduct what is, to our knowledge, the first component-level study of the recommendation algorithm deployed by a social media platform, which examines how each of its components affects misinformation propagation. Specifically, we identify the engagement fungibility mechanism in the algorithm, where the final recommendation score is constructed as a weighted sum of all predicted user activities. As a result, a tweet can be repeatedly recommended simply because it is predicted to draw many instant reactions (e.g., likes and retweets), even when it is not expected to draw thoughtful responses (e.g., replies and quotes). Since misinformation typically draws a larger share of its engagement from instant reactions, this mechanism enables it to receive more recommendation exposure and to propagate faster. To empirically validate this mechanism, we re-implement X's recommendation algorithm on the USC X 2024 election corpus, and build a calibrated simulation study to analyze the impact of different scoring rules. We find that re-tuning the metric weights has little or even a negative impact on reducing the credibility exposure gap, while those scoring rules that set a precondition of thoughtful engagement for amplification would be able to alleviate the gap significantly, across 46 robustness checks. Our diagnosis, therefore, yields a simple and deployable fix, a reflective-threshold gate that withholds amplification until a tweet is predicted to draw thoughtful engagement, which we find to reallocate exposure away from low-credibility content at no cost to mainstream exposure and with no loss of engagement.

Figures & tables

Research stream	Level of analysis	Method	Inside algorithm
Misinformation research (§ 2.1 )	User and interface	Lab and survey experiments, analytical models	No
Algorithmic amplification (§ 2.2 )	Platform output	Audits, field experiments	No
Recommender and simulation (§ 2.3 )	Whole algorithm	Field experiments, simulation	No
This paper	Algorithmic component	Calibrated simulation of the open-sourced algorithm	Yes

Table 1: Positioning of this study relative to the reviewed literature.

Action	Cognitive effort	Dual-process class
Like	one click, no reading	fast (System 1)
Retweet	one click, often no reading, in-group signaling	fast (System 1)
Reply	read $+$ compose text $+$ defend a position	slow (System 2)
Quote-tweet	read $+$ compose $+$ frame for own audience	slow (System 2)

Table 2 : Cognitive-effort partition of engagement actions.

			Mean	Median
Coverage		Engagement per tweet
Tweets	$40{,}302{,}075$	Replies	$3.81$	$0$
Unique authors	$4{,}292{,}568$	Retweets	$8.31$	$0$
Unique conversations	$17{,}433{,}960$	Likes	$39.35$	$0$
		Quote-tweets	$0.63$	$0$
Composition (% of tweets)		Impressions	$2{,}108$	$17$

Table 3: Descriptive statistics of the corpus (May 1 to November 30, 2024).

Component	Population-level modeling	Per-agent state (static)
User pool	50,000 users sampled from observed USC authors, stratified by follower-count quintile $\times$ paid verification.	Follower/friend/status/favorite/list counts and the paid-verification flag
Seed population	Sampled from observed tweets, stratified by credibility label for the hypothesis test, each seed treated as a fresh cascade root.	Posting time, text, embedded URLs, and engagement counts (training targets only)
Ranker $+$ aggregation layer	Parallel MaskNet ( Wang et al. 2021 ) with four engagement objectives, published weights for replies, retweets, and likes, and the scoring rule as the only component varied.	Per-objective probabilities, aggregate and relative scores, exposure allocation, and activity indicator

Table 4: Overview of the three simulated-system components.

Scoring rule	Formula	Role
Additive (baseline)	$S=\sum_{k}w_{k}\,p_{k}$	Production-style baseline
Slow-gates-fast multiplicative (F1)	$S=S_{\mathrm{slow}}\,(1+\alpha\,S_{\mathrm{fast}})$	Reflective-gated form
Ratio-correction (F2)	$S=S_{\mathrm{additive}}\,/\,\bigl(1+\alpha\,S_{\mathrm{fast}}/(S_{\mathrm{slow}}+\epsilon)\bigr)$	Reflective-gated form
Reflective-threshold sigmoid (F3)	$S=S_{\mathrm{additive}}\cdot\sigma\bigl((S_{\mathrm{slow}}-\theta)/\mathrm{scale}\bigr)$	Reflective-gated form
Retuned additive (control)	Additive with $w_{\mathrm{reply}}$ halved	Parameter-only control

Table 5: Scoring rules compared in the analysis.

Metric	Observed mean	Simulated mean	Welch $t$	$p$
Root reply count	$8.20$	$8.88$	$-0.43$	$0.67$
Aggregate engagement	$17.89$	$15.98$	$0.85$	$0.39$
Reflective engagement share	$0.097$	$0.099$	$-0.52$	$0.60$
Time to peak (hours)	$1.66$	$1.51$	$1.53$	$0.13$

Table 6 : Stage 1 validation statistics.

Statistic	Observed impressions	Simulated exposure
Low-credibility share of the total	$20.6\%$	$22.1\%$
Credibility gap in shares (low $-$ high)	$-58.8$ pts $[-70.5,-43.4]$	$-55.9$ pts $[-59.4,-52.4]$
Mean ratio, low to high	$0.334$ $[0.223,0.508]$	$0.372$ $[0.329,0.415]$
Gini across tweets	$0.972$	$0.906$
Top 10% share of the total	$98.1\%$	$82.5\%$
$p90/p50$	$19.8$	$14.3$

Table 7: Impressions received by tweets on the platform and exposure allocated by the simulator.

Scoring rule	Exposure contrast	Cascade size contrast
Slow-gates-fast multiplicative (F1)	$-16.14$ ∗∗∗ (4.34)	$-3.79$ ∗∗∗ (0.94)
Ratio-correction (F2)	$-2.63$ ∗∗∗ (0.28)	$-0.76$ ∗∗ (0.38)
Reflective-threshold sigmoid (F3)	$-37.83$ ∗∗∗ (5.05)	$-8.64$ ∗∗∗ (2.48)
Retuned additive (control)	$+1.52$ ∗∗∗ (0.11)	$+0.54$ ∗∗ (0.26)

Table 8 : Exposure and cascade size contrasts against the additive baseline.

Fidelity to the published ordering	Configurations	Best gap reduction attained
Ordering closely preserved ( $\rho\geq 0.95$ )	1,489	$24.3\%$
Ordering loosely preserved ( $\rho\geq 0.90$ )	1,946	$64.0\%$
Unconstrained	2,083	$74.6\%$

Table 9 : Best attainable gap reduction by re-weighting, at each level of fidelity to the published ordering.

Feature	Reflective-to-reactive gap	Interpretation
Listed count (log)	$+0.286$	prestige $\rightarrow$ reply-heavy
Followers (log)	$+0.258$	prestige $\rightarrow$ reply-heavy
Text length	$+0.219$	substance $\rightarrow$ reply-heavy
Paid verification	$+0.214$	verification $\rightarrow$ reply-heavy
Follower-to-friend ratio (log)	$+0.191$	influence $\rightarrow$ reply-heavy
Mention count	$+0.134$	conversation $\rightarrow$ reply-heavy

Table 10 : Engagement mix by account and content feature.

Metric (mean per tweet)	low-credibility ( $n=3{,}116$ )	high-credibility ( $n=6{,}853$ )
Replies	$1.6$	$2.1$
Retweets	$3.7$	$3.2$
Likes	$8.1$	$8.0$
Reply share of engagement	$33.0\%$	$46.4\%$
Cascade-size proxy	$5.5$	$5.6$

Table 11 : Observed per-tweet engagement by credibility class, for tweets matched on total engagement.

Quantity (mean per tweet)	low-credibility	high-credibility
$P(\text{reply})$	$0.238$	$0.282$
$P(\text{retweet})$	$0.323$	$0.303$
$P(\text{like})$	$0.485$	$0.466$
$P(\text{quote})$	$0.103$	$0.116$
Predicted reactive/reflective	$\mathbf{2.973}$	$2.620$
$S_{\mathrm{additive}}$	$3.935$	$4.417$

Table 12 : Ranker-predicted engagement by credibility class, for the matched tweets.

Check	Variation	Finding
1	Aggregation form (F1–F3 and the retuned control)	All three reflective-gated forms narrow the gap significantly, the control moves it the opposite way
2	Slow/fast partition (four assignments)	Significantly negative under all four
3	Ranker training seed (five retrainings)	Contrasts cluster in $[-21.3,-15.5]$ , every one significant
4	Corpus period (adjacent week-long slice)	$-16.3$ on the adjacent slice against $-16.1$ at baseline
5	Partition $\times$ training seed ( $4\times 5$ combinations)	Significantly negative in all 20 combinations
6	User-pool size (25K, 50K, 100K users)	Identical contrasts at all three sizes

Table 13: Summary of robustness checks.

Hypothesis	Sections	Result
H1: The aggregation form narrows the gap	6.2 , 6.5	Supported at both layers, on exposure under all three reflective-gated forms and in all 46 robustness variations, and on cascade size under the baseline ranker.
H2: Re-tuning the weights does not help	6.2 , 6.3	Supported, as the retuned control moves the gap in the opposite direction and no weight configuration reaches the threshold form’s gap reduction.
H3: The gate acts on the engagement mix	6.4	Supported, as demotion tracks the predicted reactive-to-reflective ratio, and the placebo gate and the correlation analysis show that neither alternative explanation accounts for the targeting.

Table 14: Summary of hypothesis-test results.

$\theta$	Low-credibility exposure share	High-credibility exposure share	Expected engagement per exposure	Top-20% overlap
$0.5$	$-2.9\%$	$+2.3\%$	$+9.4\%$	$99.7\%$
$1.0$	$-5.0\%$	$+3.9\%$	$+17.4\%$	$99.3\%$
$1.5$	$-7.0\%$	$+5.5\%$	$+27.2\%$	$99.0\%$
$2.0$	$-8.1\%$	$+6.4\%$	$+37.4\%$	$98.6\%$
$2.2$	$-8.2\%$	$+6.5\%$	$+41.4\%$	$98.4\%$

Table 15 : Benefit and engagement cost of the reflective-threshold gate at different values of

\theta

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Training corpus	$n_{\mathrm{train}}$	Metric	Scoring rule	Estimate
Baseline	1.0M	Exposure	F1 (multiplicative)	$-16.14$ ∗∗∗ (4.34)
Baseline	1.0M	Cascade size	F1 (multiplicative)	$-3.79$ ∗∗∗ (0.94)
Enlarged	2.0M	Exposure	F1 (multiplicative)	$-17.22$ ∗∗∗ (3.93)
Enlarged	2.0M	Cascade size	F1 (multiplicative)	$-7.04$ ∗∗∗ (0.82)
Enlarged	5.0M	Exposure	F1 (multiplicative)	$-19.28$ ∗∗∗ (4.25)
Enlarged	5.0M	Cascade size	F1 (multiplicative)	$-8.07$ ∗∗∗ (0.73)

Appendix

Table 16 : Contrasts when the ranker is trained on larger corpora (Check 7).

Subset	$n_{\mathrm{features}}$	F1 score contrast	Control score contrast
Full	18	$-0.354$ ∗∗∗ $(0.081)$	$+0.261$ ∗∗∗ $(0.031)$
Tweet and time only	8	$-0.142$ ∗∗∗ $(0.012)$	$+0.167$ ∗∗∗ $(0.011)$
Account only	10	$-0.216$ ∗∗ $(0.088)$	$+0.096$ ∗∗∗ $(0.031)$
Minimal	3	$-0.199$ ∗∗∗ $(0.072)$	$+0.141$ ∗∗∗ $(0.030)$

Appendix

Table 17 : Score contrasts across ranker feature subsets (Check 8).

Form	Setting	Estimate
F1 (multiplicative)	$\alpha=0.25$	$-6.38$ ∗∗∗ (1.54)
	$\alpha=0.5$	$-9.81$ ∗∗∗ (2.70)
	$\alpha=1.0$	$-16.14$ ∗∗∗ (4.34)
	$\alpha=2.0$	$-26.09$ ∗∗∗ (8.24)
	$\alpha=4.0$	$-40.23$ ∗∗∗ (13.0)
	$\alpha=8.0$	$-56.01$ ∗∗∗ (18.4)

Appendix

Table 18 : Exposure contrast at every hyperparameter setting (Check 9).

Iteration	Event model	Estimate of the dispersion $r$	KS (root reply)	KS (reflective share)
(i)	Poisson with the mean matched	—	0.60	0.62
(ii)	Zero-inflated Poisson	—	0.090	0.097
(iii)	Zero-inflated Negative-Binomial	Mean and variance of the non-zero counts	0.165	0.150
(iv)	Zero-inflated Negative-Binomial	Mean and variance after dropping the largest 5% of the non-zero counts	0.057	0.061

Appendix

Table 19 : Calibration iterations and their Kolmogorov–Smirnov fit.

Explore similar work

Jul 2, 2026cs.LG

Gaming Consensus: Coordinated Manipulation in Crowdsourced Fact-Checking

Crowdsourced fact-checking systems have been adopted by major social media companies such as X, Meta, TikTok and Google with the aim of combating misleading information at scale without relying on centralized editorial control. These systems have been developed around a common underlying concept: a bridging mechanism that identifies notes flagging misleading information when they receive support from people with different perspectives rather than simple majority support. To our knowledge the only publicly disclosed bridging algorithms deployed for fact-checking are based on matrix factorization, as deployed by both X and Meta, augmented with additional components addressing abuse, targeted manipulation, and contributor brigades. This work examines the core matrix factorization portion of these systems, presenting theoretical and empirical evaluations of the degree to which coordinated users could vote strategically by leveraging the latent representations to fabricate the appearance of synthetic consensus within the bridging mechanism. Using historic production data, we find that up to 10.7% of lower quality notes could be manipulated above consensus thresholds using less than 10 ratings. We complement these findings with a theoretical analysis, revealing counterintuitively that rating a note as "Not Helpful" can increase its helpfulness score, as well as a cost model quantifying manipulation effort. We have developed and deployed mitigations within X's Community Notes algorithm to address synthetic consensus.

Nikil Roashan Selvam, Jay Baxter, Sophie Hilgard +4

Stanford University, Stanford, CA, USA · X Community Notes, xAI, Palo Alto, USA

Fact-Checking Crowdsourced Framework Social Media

Ask PDF

Aug 31, 2026cs.LG

A Multi-Branch Feature Fusion Approach for Health Misinformation Detection and Propagation

This paper presents a multi-branch fusion framework for detecting and characterising the propagation of health misinformation in online social networks (OSNs). Grounded in the Elaboration Likelihood Model (ELM) and the Theory of Planned Behaviour (TPB), the model fuses transformer-based semantics with rhetorical cues, stance representations, and psychologically motivated proxies in a unified multi-task architecture. In addition to binary classification, we introduce the Cognitive Propagation Score (CPS), an interpretable post-hoc auxiliary score computed from psychologically motivated, text-derived cues capturing argument complexity, emotional intensity, and content-derived virality potential, to support diffusion-risk reasoning when engagement ground truth is incomplete or unavailable. Experiments on three benchmark datasets, Constraint, COVID--19_FNIR, and Monkeypox, show strong classification performance, achieving ROC--AUC up to 0.9999 on COVID--19_FNIR, while propagation-oriented ranking achieves near-perfect agreement when engagement-derived supervision is available (Monkeypox, Spearman's

ρ= 0.9952

) and similarly high ranking alignment under proxy-based supervision on COVID--19_FNIR (

ρ= 0.9954

). Compared with representative literature baselines, the fusion model improves detection on Constraint and COVID--19_FNIR, while Monkeypox remains more challenging, reflecting domain- and signal-specific differences. Ablation analysis further indicates that psychological and rhetorical branches provide complementary gains beyond semantic embeddings. Overall, the framework bridges cognitive theory and neural modelling to improve transparency and to support scalable misinformation monitoring, with future work required to validate CPS against human-centred diffusion judgements.

Mkululi Sikosana, Sean Maudsley-Barton, Oluwaseun Ajao

Department Of Computing And Mathematics, Manchester Metropolitan University, Manchester, UK

Misinformation Detection Misinformation Social Media False-Negative Rate Propagation Rhetoric+1

Ask PDF

Aug 31, 2026cs.HC

One AI Signal, Many Human Judgments: A Bayesian Cascade Analysis of AI-based Credibility Indicators in Online Information Spread

Social media platforms increasingly use AI-based credibility indicators to help users judge misinformation. Unlike individual human-AI decision-making, these indicators are embedded in information spread: users see both an AI prediction and earlier judgments shaped by the same AI, and their own judgments may then enter the public history. Yet how to analytically characterize this process remains under-explored. We therefore introduce a social-learning lens for this setting by extending the classical Bayesian cascade model with the AI indicator as a shared public signal. The resulting Gateway condition compares the evidence from the AI prediction with users' private impressions. Through this view, we show that AI changes what public history means. Crowd agreement may reflect accumulated independent human evidence, or repeated dependence on the same AI prediction. This creates a preservation-correction trade-off: stronger reliance on AI can preserve correct predictions, but can also lock in incorrect ones by blocking corrective private impressions. We calibrate the model using human-subject data on news veracity judgments. Although the AI outperforms human users, the average user weights it below her own impression but above several peer judgments, while individual users vary from discounting the AI to relying on it enough to cascade. Simulations show that over-reliance on a weak AI is especially harmful, and that diversifying AI signals across users can better keep the crowd informative. We conclude with implications for understanding human-AI interaction in information spread and designing misinformation interventions.

Zhuoran Lu, Weilong Wang, Yangyang Yu +4

Purdue University West Lafayette, Indiana, USA · St John’s University New York, New York, USA · Stevens Institute of Technology Hoboken, New Jersey, USA +2

Credibility Misinformation Artificial Intelligence Systems Human Judgment

Ask PDF