Why Does Misinformation Propagate Faster? An Algorithmic Perspective on X

1. Introduction

• Prior empirical studies document differential misinformation propagation but cannot determine whether it arises from human attention or a specific recommender design choice.Observed records are shaped by algorithmic allocation, and the propagation step from engagement to further audience is inside the ranking system.Source: §1 · p. 2
• Engagement fungibility lets a strong predicted fast-class signal, such as a like or retweet, compensate for absent slow-class responses, such as replies or quotes, in X’s additive score.Low-credibility content draws a larger share of engagement from instant reactions, allowing it to gain recommendation exposure despite fewer thoughtful responses.Source: §1 · p. 3
• To study counterfactual recommender changes without production logs, the authors re-implement X’s algorithm and calibrate a simulation against the USC X 2024 election corpus.The simulator enables comparison of architectural variants under modified recommendation components.Source: §1 · pp. 3–4
• Across 46 robustness checks, variants that privilege slow signals over fast signals reduce the propagation gap, while changing signal weights alone does not alleviate it and can widen it.The checks span eight groups.Source: §1 · pp. 3–4
• A reflective-threshold gate that requires predicted slow-class engagement before full amplification reallocates exposure away from low-credibility content without reducing overall engagement.The gate’s benefit and cost vary with its platform-tunable threshold.Source: §1 · p. 4

2.1. Misinformation Research in Information Systems

• Misinformation research has examined user susceptibility, content judgments, and platform interventions, but largely asks how users evaluate content after it is shown rather than how it gets shown.Economic models represent the algorithm as an abstract policy and do not identify the responsible recommender component.Source: §2.1 · pp. 4–5

2.2. Algorithmic Amplification on Social Media

• Audits and experiments document algorithmic amplification across platforms, while X studies leave its internal recommendation components unexplored.Prior X analyses rely on empirical evidence rather than component-level intervention.Source: §2.2 · p. 5

2.3. Recommender Systems and Simulation-Based Research

• Recommender research shows that design shapes downstream outcomes, but typically treats the recommender as a whole; simulation studies likewise compare whole algorithms rather than components.These approaches do not probe a deployed architecture to identify which internal component drives misinformation outcomes.Source: §2.3 · p. 6

2.4. Research Gap

• The paper addresses three gaps by reconstructing open-sourced X, intervening on score aggregation in a calibrated simulation, and locating a misinformation-propagation mechanism inside the algorithm.The authors describe this as the first study, to their knowledge, to locate that mechanism inside X’s algorithm.Source: §2.4 · pp. 6–7

3.1. The X Recommendation Algorithm

• X’s pipeline sources roughly 1,500 candidate tweets, narrows them with the Light Ranker, predicts engagement probabilities with a Heavy Ranker, and ranks candidates by a weighted sum.The Heavy Ranker predicts replies, retweets, likes, profile clicks, video views, and negative feedback; reply receives weight 13.5, retweet 1.0, and like 0.5.Source: §3.1 · p. 7, Figure 1 · p. 8
• The authors focus on the Heavy Ranker and aggregation layer because the ranker’s scores primarily determine exposure and aggregation weights are a platform design choice.The upstream stages mainly control candidate-pool composition, while per-objective predictions forecast reactions.Source: §3.1 · pp. 7–8

3.2. Dual-Process Theory and the Cognitive Partition of Engagement

• Dual-process theory distinguishes automatic, fast System 1 responses from effortful, slow System 2 responses, which the paper uses to classify engagement actions by cognitive effort.The cognitive-effort partition is summarized in Table 2.Source: §3.2 · p. 8, Table 2 · p. 8
• Among tweets with the same total engagement, replies account for 33.0% for low-credibility tweets and 46.4% for high-credibility tweets, supporting a fast/slow partition rather than reliance on any particular weights.The additive score makes predicted actions exchangeable at a fixed rate, so high one-click reactions can compensate for low predicted replies.Source: §3.2 · pp. 8–9

4.1. Definition of the Mechanism

• Fast and slow engagement signals are fungible in the additive score, so a strong fast-class signal can drive exposure without slow-class engagement.The paper describes the signals as perfect substitutes, with their exchange rate set by the weights.Source: §4.1 · pp. 9–10
• Every fungible rule is ordinally equivalent to one normalized additive rule, so analyzing additive weights covers the full fungible class.The theorem's implications motivate requiring slow-class validation before fast-class predictions amplify content.Source: §4.1 · p. 10

4.2. Potential Solution and Hypotheses

• Reflective-gated aggregation is nonfungible and bounds a tweet’s score when predicted slow-class engagement is zero.The proposed design lets predicted fast-class engagement amplify only what slow-class engagement supports.Source: §4.2 · pp. 10–11
• Three proposed reflective-gated forms modify the additive score through slow-engagement validation, a fast-to-slow ratio correction, or a threshold gate.The ratio correction's penalty grows with fast-class predictions; the threshold form scales the score down smoothly below the slow-engagement threshold.Source: §4.2 · p. 11
• The hypotheses predict that reflective gating narrows the gap while re-tuning additive weights does not, because low-credibility seeds tilt further toward fast engagement.The stated account attributes demotion to the predicted engagement mix at a given slow-engagement level, not overall predicted engagement.Source: §4.2 · p. 12

5.1. USC X 2024 Election Corpus

• The USC X 2024 U.S. Election Dataset contains keyword-filtered tweets collected May 1–November 30, 2024, during the open-sourced 2023 algorithm era.The authors selected it for its election-period low-credibility content and matching algorithmic era.Source: §5.1 · p. 12
• 17 impressions is the corpus median, while tweets in the top percentile draw at least 20 replies and 206 likes, illustrating concentrated engagement.Replies account for around two-thirds of tweets and original posts for nearly one-quarter.Source: §5.1 · p. 12, Table 3 · p. 13

5.2. Labeling of Credibility

• Credibility labels use embedded publishing domains: all matched domains must fall on the low- or high-credibility list, while mixed or unmatched tweets receive different labels.The labels cover about 2.2% of records and measure source credibility, not whether a tweet’s claim is true.Source: §5.2 · p. 13

5.3. Simulation Framework Overview

• The simulation samples seed tweets and users from the corpus and uses a ranker plus score-aggregation layer to assign algorithmic exposure.The design holds the trained ranker, populations, exposure mapping, and calibrated parameters fixed while varying the scoring rule.Source: §5.3 · p. 14

5.4. Ranker, Score Aggregation, and Exposure

• The ranker predicts replies, retweets, likes, and quotes, then aggregates those predictions into a score that determines exposure.The score is normalized against the seed-population median and mapped to a bounded per-cascade exposure allocation.Source: §5.4 · pp. 14–15

5.5. Engagement Event Model

• The event model first samples whether a cascade attracts any engagement, then generates engagement counts from its exposure and predicted action probabilities.Calibration uses root-reply statistics under the additive rule and remains fixed in subsequent scoring-rule analyses.Source: §5.5 · p. 16

6. Empirical Analysis and Results

• The empirical analysis initializes the simulator on the USC X corpus and compares five scoring rules in a two-stage process.Stage 1 validates simulated cascade patterns; Stage 2 tests how changing aggregation affects low- and high-credibility tweet propagation.Source: §6 · p. 17, Table 5 · p. 17

6.1. Stage 1: Validation of Our Simulation Framework

• None of the four engagement-metric differences between observed and simulated means is statistically significant (all p > 0.10), and simulated curves track observed distributions closely.For exposure allocation, the simulator’s credibility gap is −55.9 points versus an observed −58.8; its exposure ratio is 0.372 versus 0.334, with overlapping bootstrap intervals.Source: §6.1 · pp. 17–19, Figure 2 · p. 18, Table 6 · p. 18

6.2. Testing Hypothesis 1: The Aggregation Form Narrows the Gap

• Reflective-gated aggregation significantly narrows the exposure gap under all three forms, and its cascade-size contrast is also significantly negative; the reflective threshold produces the largest reduction.Changing only the additive weights moves the exposure gap in the opposite direction.Source: §6.2 · p. 19, Table 8 · p. 19

6.3. Testing Hypothesis 2: Re-Tuning the Weights Does Not Help

• 74.3% of 2,083 additive-weight configurations widen the credibility gap, and only 1.0% achieve even half the reflective threshold’s reduction; none reach 80%.Setting retweet and like weights to zero achieves only 6.9% of the threshold form’s reduction; configurations that reduce the gap most conflict with preserving the published ordering.Source: §6.3 · p. 20, Table 9 · p. 21

6.4.1. Grounding the Cognitive Partition

• Features such as author-list presence and tweet length accompany reply-heavy engagement, while low-credibility links, capitalization, and angry or disgusted language accompany retweet-heavy engagement.The gate narrows the credibility gap under all four tested divisions of engagement.Source: §6.4.1 · p. 21, Table 10 · p. 22
• At matched total engagement, replies account for 33.0% of low-credibility tweets’ replies, retweets, and quotes, versus 46.4% for high-credibility tweets.Low-credibility tweets receive significantly more retweets and fewer replies in this comparison.Source: §6.4.1 · p. 22, Table 11 · p. 22

6.4.4. How the Gate Acts

• 63.4% of low-credibility tweets receive a gate value below 0.5, versus 47.3% of high-credibility tweets; the placebo fast-class gate narrows the exposure gap 12.1 less.Demotion rises with the predicted reactive-to-reflective ratio, not predicted slow-class engagement alone, and the tests attribute the gap reduction to engagement composition.Source: §6.4.4 · pp. 23–24

6.5. Robustness Checks

• The exposure pattern holds across all 46 robustness variations and 23 gate settings, with the reflective threshold producing 2.3× the exposure effect of the multiplicative gate.Every exposure contrast is significant in the predicted direction, while the parameter-only control always moves oppositely.Source: §6.5 · p. 25, Table 13 · p. 25

7.1. The Solution

• At θ = 2, the reflective-threshold gate reduces low-credibility tweets’ exposure share by about 8% without reducing high-credibility exposure.The reduction grows steadily with θ; high-credibility tweets receive the exposure low-credibility tweets lose.Source: §7.1 · p. 26, Figure 4 · p. 27
• At θ = 1, expected engagement per exposure rises about 17%, while the gate leaves the top 10% of tweets unchanged.At least 98% of the top 20% remain the same as under the additive rule; Table 15 reports these measures at different thresholds.Source: §7.1 · pp. 26–27, Table 15 · p. 27
• About 97% of corpus tweets lack credibility labels, so their response to the gate is not established by the labeled-tweet results.In a mixed set, unlabeled tweets move slightly up on average and account for 18% of the most-demoted tenth, despite making up half the set.Source: §7.1 · p. 27

7.2. Evaluation and Validation

• Validation tests whether demotion tracks instant-reaction reliance, narrows cascade-size gaps, and extends to tweets with misleading individual claims.The authors use these checks to address the concern that a thoughtful-engagement gate would hold back low-engagement content by design.Source: §7.2 · pp. 27–28

7.2.1. Targeting and Propagation

• Low-credibility tweets have higher substitutability shares and make up 61% of the most-demoted quarter, versus 44% of the least-demoted quarter.Within each credibility class, the gate moves tweets down according to substitutability share; the cascade-size gap also narrows significantly.Source: §7.2.1 · p. 28

7.2.2. The Claim-Level Limit

• The reflective threshold is not a detector of claim truth: misleading-claim ratings do not predict greater demotion, and noted tweets rely less on instant reactions.The Community Notes comparison, matched on post type and engagement, confirms this limitation with an independent set of labels.Source: §7.2.2 · pp. 28, 30
• At η = 0.15, adding a content gate retains 97% of the original demotion of high-substitutability tweets while moving misleading tweets down more.Targeting retention is 84%; the combined score changes the exposure gap by −38.2 (s.e. 5.41), not statistically different from the reflective threshold alone.Source: §7.2.2 · pp. 30–31
• A small content-gate weight can add claim-level demotion while retaining reflective-threshold targeting, whereas stronger tested rules do not solve the problem through engagement aggregation alone.The authors conclude that passing the claim-level check requires additional ranker inputs carrying information about claims themselves.Source: §7.2.2 · pp. 29–31

8.1. Implications for Researchers

• Engagement fungibility locates differential exposure in score aggregation, complementing content- and user-level explanations of reactive engagement.The algorithm turns reactive engagement into differential exposure; cited prior work addresses why content provokes such engagement (Vosoughi et al. 2018, Brady et al. 2017) and individual susceptibility (Pennycook and Rand 2019).Source: §8.1 · p. 32
• Requiring predicted thoughtful engagement before instant reactions can raise a tweet’s score narrows the credibility gap without abandoning engagement-based ranking.The authors locate the bias in a replaceable recommender layer, rather than treating it as intrinsic to engagement-based ranking.Source: §8.1 · p. 32
• In this corpus, high-credibility content receives more total engagement, but low-credibility content gets a larger share from instant reactions.The authors therefore suggest examining engagement composition as well as its amount.Source: §8.1 · p. 32

8.2. Implications for Practitioners

• Platforms can change the score-aggregation layer without retraining models or changing their published weights.The intervention uses predicted engagement patterns rather than truth labels, and can work alongside content signals through separate channels.Source: §8.2 · p. 33
• The authors report no engagement loss in simulation and suggest platforms can verify the trade-off on their own traffic.They also describe post-scoring deployment as adding little serving cost and predicted thoughtful engagement as harder to manipulate than one-click reactions.Source: §8.2 · p. 33

8.3. Limitations and Future Research

• Effects are immediate simulated changes, not a new long-run equilibrium, and exposure measures algorithmic distribution rather than what people ultimately see.The study reconstructs rather than replicates the production recommender; its corpus is keyword-filtered, from one platform and period, and labels source credibility rather than claim truth.Source: §8.3 · p. 34
• A powered claim-level test would require labels for a few thousand tweets and a production-grade credibility model; a cooperating-platform field test could measure real-world effects.The authors also propose cross-platform studies and audits of past score-aggregation changes.Source: §8.3 · p. 34

A. Stage 2 Analysis Plan

• The preregistered Stage 2 plan defined a scoring-rule gap as the low- minus high-credibility mean, and a contrast as that gap minus the additive-rule gap.Cascade size was primary, exposure and time to peak secondary; each rule was simulated 20 times, with uncertainty measured across replicates.Source: A. Stage 2 Analysis Plan · p. 39
• Before confirmatory results, an amendment added a 1,000-iteration bootstrap, raised simulations from 20 to 100, and made exposure a primary metric.The amendment noted that replicate spread is zero for exposure because exposure is fully determined by scores; the plan otherwise remained unchanged.Source: A. Stage 2 Analysis Plan · p. 39

• All three reflective-gated forms significantly narrow exposure and cascade-size gaps; the retuned control moves both gaps in the opposite direction.The reflective threshold has the largest effects on both metrics, and ratio correction the smallest.Source: B. Robustness Check Details · p. 39
• Across four engagement partitions, five ranker seeds, an adjacent week, and 20 partition-seed combinations, exposure and cascade-size contrasts remain negative.The adjacent-week exposure contrast is −16.3 (s.e. 4.82), close to the baseline −16.14 (s.e. 4.34).Source: B. Robustness Check Details · p. 40
• Across user pools of 25K, 50K, and 100K, exposure contrast is −16.14 (s.e. 4.34) and cascade-size contrast is −3.79 (s.e. 0.94).The reported contrasts do not depend on the number of simulated users.Source: B. Robustness Check Details · p. 40
• As training data grows from 1M to 5M tweets, F1 exposure contrasts widen from −16.1 to −19.3, while the retuned control remains opposite in direction.F1 cascade-size contrasts grow from −3.8 to −8.1; Table 16 reports results at 1M, 2M, and 5M.Source: B. Robustness Check Details · p. 40, Table 16 · p. 41
• Across four feature subsets and three prediction-model types, the multiplicative form narrows score gaps while the retuned control widens them.The transformer yields contrasts of −0.427 (s.e. 0.0954) for F1 and +0.301 (s.e. 0.0339) for the retuned control; Table 17 reports feature-subset results.Source: B. Robustness Check Details · p. 41, Table 17 · p. 41
• Exposure contrast is significantly negative at all 23 gate settings, with larger effects as gain or threshold increases.For θ ≥ 1, effects also grow as the gate becomes steeper (smaller scale); very large contrasts at θ ≥ 1.5 reflect score reductions for most tweets. Figure 7 presents sensitivity to gate hyperparameters.Source: Figure 7 · p. 42, B. Robustness Check Details · p. 42, Table 18 · p. 43

C. Cascade-Simulation and Calibration Procedure

• The activity gate, reply intensity scale, and dispersion parameter are calibrated against observed no-reply share, mean replies among replied-to tweets, and trimmed reply-count moments.Calibration uses the additive scoring rule; other scoring rules reuse these parameters so differences come from score aggregation.Source: C. Cascade-Simulation and Calibration… · p. 43
• Trimming the largest 5% of nonzero reply counts avoids an implausibly small dispersion estimate and improves fit to typical cascades at the cost of accuracy on extreme cascades.The raw variance is dominated by a few very large counts, while the trimmed estimate keeps the model focused on typical cascades relevant to the mechanism.Source: C. Cascade-Simulation and Calibration… · p. 44
• The procedure samples 50,000 authors stratified by follower-count quintile and paid verification, samples seed tweets, and runs M replicates with different random seeds.It estimates the hourly posting profile from observed posting times, records cascade totals and hourly counts by engagement objective, and combines replicates for Stage 1 and Stage 2 analyses.Source: C. Cascade-Simulation and Calibration… · p. 44
• Each replicate records cascade totals and hourly counts, which are then combined for the Stage 1 and Stage 2 analyses.Source: C. Cascade-Simulation and Calibration… · p. 44

D. Calibration Iteration History

• Calibration progressed from a mean-matched Poisson model to a zero-inflated Poisson, then a zero-inflated Negative-Binomial model with dispersion estimated from nonzero reply counts.The fourth iteration retained the Negative-Binomial model but estimated dispersion after trimming the largest 5% of nonzero counts.Source: D. Calibration Iteration History · pp. 44–45
• The fit comparison uses KS statistics for simulated versus observed root reply counts and reflective-engagement shares, with smaller values indicating better fit.Source: D. Calibration Iteration History · p. 45
• Trimming the largest 5% of nonzero counts yielded r = 0.318, close to the grid-search KS-minimizing r ≈ 0.5 (KS 0.046), and iteration (iv) fit best.The raw-variance estimate r ≈ 0.0098 made iteration (iii) fit worse than iteration (ii) on both distributions.Source: D. Calibration Iteration History · p. 45, Table 19 · p. 45

E. Data Processing Details

• Reply-to-tweet linking is approximate for long identifiers because conversation and replied-to tweet identifiers are stored in scientific notation with about 16 significant digits.Source: E. Data Processing Details · p. 45
• The corpus cleaning drops 453,307 malformed records and 2,096,410 duplicate records, leaving 40,302,075 tweets; retweets are excluded from engagement statistics because they carry the original tweet’s counts.Including retweets would raise the average retweet count from 8.3 to 20.5.Source: E. Data Processing Details · p. 46
• The ranker uses 18 features: ten account features and eight tweet-and-time features, including paid verification, interaction type, URLs, text length, and posting-time encodings.Source: E. Data Processing Details · p. 46