**==> picture [81 x 21] intentionally omitted <==**

# **AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X** 

Haiwen Li Michiel A. Bakker 

Massachusetts Institute of Technology 

## **Abstract.** 

Large language models (LLMs) show promising capabilities for contextual facton social media: can contested claims re- checking they verify through deep search, synthesize evidence from multiple sources, and draft explanations at scale. However, prior work evaluates LLM fact-checking only in controlled settings usbenchmarks or crowdworker how these ing judgments, leaving open systems perform in authentic platform environments. We present the first field evaluation of LLM-based fact-checking deployed on a live social media platform, testing performance directly through X Community Notes’ “AI writer” feature over a threemonth period. Our LLM writer, a multi-step pipeline that handles multimodal content (text, images, and videos), conducts web and platform-native search, and writes contextual notes, was deployed to write 1,614 notes on 1,597 tweets and compared against 1,332 human-written notes on the same tweets using 108,169 ratings from 42,521 raters. Direct comparison of note-level platform outcomes is complicated by differences in submission timing and rating exposure between LLM and human notes; we therefore pursue two complementary strategies: a rating-level analysis modeling individual rater evaluations, and a note-level analysis that equalizes rater exposure across note types. Rating-level analysis shows that LLM notes receive more positive ratings than human notes across raters with different political viewpoints, suggesting the potential for LLM-written notes to achieve the cross-partisan consensus. Note-level analysis confirms this advantage: among raters who evaluated all notes on the same post, LLM notes achieve significantly higher helpfulness scores. Our findings demonstrate that LLMs can contribute high-quality, broadly helpful fact-checking at scale, while highlighting that real-world evaluation requires careful attention to platform dynamics absent from controlled settings. 

## **1 Introduction** 

Social media on mechanisms to add con- platforms increasingly rely community-based text to potentially misleading content. X Community Notes exemplifies this approach 

1 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

(Wojcik et al., 2022): users propose short, source-cited notes to add context to posts, and community ratings from a diverse group of raters determine whether a note is “broadly helpful” and should be shown publicly. This design creates a stringent alignment target that rewards neutral, evidence-grounded, cross-perspective explanations. Communitybased fact-checking has also been examined to increase trust in fact-checking on social media and to reduce misinformation diffusion and engagement with misleading content (Drolsbach et al., 2024; Slaughter et al., 2025). 

The core tasks of note authorship—researching relevant information, analyzing competing evidence, and writing clear factual explanations—align closely with capabilities where large language models have shown promise. Prior work finds that LLMs can generate fact-checking notes comparable to, and under certain conditions better than, human-written notes while requiring far less time and effort (Zhou et al., 2024; De et al., 2025; Singh et al., 2026). However, existing evaluations of LLM-based fact-checking remain offline: performance is measured against automated benchmarks, expert annotations, or crowdworker judgments in controlled settings (Zhou et al., 2024; De et al., 2025; Costabile et al., 2025; Singh et al., 2026). Crowdworker evaluations in particular are susceptible to demand effects where people evaluate content in ways they believe the study expects rather than reflecting genuine judgement (Mummolo and Peterson, 2019; Iarygina et al., 2025). Hence, these evaluations fail to capture the on-platform dynamics such as timing, interface presentation, limited user attention, and subjective judging criteria that shape real-world impact. A critical gap therefore remains: we lack evidence on how LLM-generated fact-checks perform in situ, where their ultimate efficacy is determined not by gold-standard labels but on the dynamics of a live platform environment. 

We present the first field evaluation of LLM fact-check writing deployed on-platform. X’s recently released Community Notes “AI writer” API enables users to build AI writers that propose notes on potentially misleading posts flagged by human users.[1] Notes generated by the system are clearly labeled as AI-created and are evaluated under the same crowd-sourced rating criteria as human-written notes. We developed and deployed our own LLM-based Community Notes writer on X through the “AI writer” API. This LLM writer conducts web search and platform-native search to collect information, assesses whether sufficient evidence exists to warrant a note, generates contextual notes, applies quality control filters, and publishes notes directly to the platform. The writer also handles multimodal posts that include images and videos. Over a three-month deployment period (November 1, 2025 to January 31, 2026), our LLM writer published 1,614 notes on 1,597 unique tweets, which we compared against 1,332 human-written notes on the same tweets, drawing on 108,169 ratings from 42,521 Community Notes raters. While there are 

> 1https://communitynotes.x.com/guide/en/api/overview 

2 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

multiple AI writers operating on the platform, we focus on evaluating our own writer because we can document and its our results to be fully open-source pipeline, enabling interpreted in light of specific, reproducible design decisions. 

Our analysis reveals several key findings. First, direct comparison of note-level platform outcomes between LLM and human notes is complicated by systematic differences in submission timing due to platform design. Because notes submitted later receive fewer ratings and the platform’s algorithm penalizes notes with fewer ratings, these timing differences can bias comparisons. We therefore pursue two analytical strategies. In our primary, rating-level analysis, we show that LLM notes receive uniformly more positive evaluations from raters across the political spectrum, suggesting LLM-generated content may be likely to achieve the cross-partisan consensus that Community Notes seeks to foster. In a complementary note-level analysis that equalizes rater exposure across note types, LLM notes achieve significantly higher helpfulness scores than human notes. We additionally find that the LLM advantage varies by topic: it is largest for health and medicine content and conspiracy theories/pseudoscience, and smallest for posts about AI-generated content. 

Our contributions are threefold. Methodologically, we provide the first evaluation of LLM fact-checking in a real-world platform setting with organic user feedback, moving beyond laboratory studies, synthetic benchmarks, and proxy metrics to assess performance under authentic conditions, providing ecological validity. Practically, we open-source our writer and which were implementation, including prompts multi-step writing pipeline, developed and refined over a month to improve evidence retrieval, prompt quality, and pipeline structure.[2] Empirically, our findings demonstrate that LLMs can contribute highquality fact-checking annotations at scale, achieving broader cross-ideological acceptance than human-written notes, suggesting that automated systems can meet or exceed human performance on the demanding standard of broad helpfulness. We also discuss factors that complicate note helpfulness measures on a platform: note creation timing and differential exposures shape platform-level outcomes beyond intrinsic note helpfulness, suggesting that the true effects of AI writing in the wild are more complicated than those measured in lab settings. 

> 2 Code for the writing pipeline and analysis is available at https://github.com/haiwen-li/cn-ai-writer. 

3 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **2 Results** 

## **2.1 Platform Dynamics Complicate Note-Level Comparisons** 

X Community Notes uses a matrix factorization–based bridging algorithm to aggregate ratings from individual users and evaluate note quality.[3] Rather than relying on simthe scores to notes that re- ple majority voting, algorithm assigns higher helpfulness ceive “helpful” ratings from raters who have historically disagreed in their past ratings. This design surfaces notes that are broadly acceptable across different viewpoints, and notes whose helpfulness scores exceed a platform-defined threshold receive the Currently Rated Helpful (CRH) status and become publicly visible on the platform. 

note scores and CRH status intuitive mea- Although helpfulness provide performance sures, comparing LLM and human-written notes using these metrics is complicated by platform dynamics that are unrelated to intrinsic note helpfulness. The Community Notes algorithm implicitly penalizes notes with fewer ratings and deflates their helpfulness scores,[4] yet the number of ratings a note receives depends heavily on when it is submitted. Earlier notes gain more exposure through longer visibility and higher UI ranking, which in turn increases rating accumulation. At present, LLM notes face a systematic exposure disadvantage. While human writers can propose notes whenever they notice misleading posts, LLM writers can only retrieve and write notes on posts after enough users have flagged them as potentially misleading (this is a current platform policy constraint rather than a limitation of LLM fact-checking itself). Among the 814 tweets receiving both note types in our data, 66.0% of human notes were created before the LLM notes. As a result, human notes accumulated substantially more ratings (mean = 109.94 vs. 59.50; median = 51 vs. 22; Mann–Whitney _U_ = 740,798.0, _p <_ 0.001). This temporal asymmetry implies that note-level metrics differences in driven submission may partly reflect exposure by timing, rather than differences in note helpfulness alone. 

To address these confounds, we pursue two analytical strategies. First, we conduct analyses at the rating level, where each observation corresponds to an individual user’s explicit evaluation of a note (helpful, somewhat helpful, or unhelpful). These ratings provide the ground truth of user judgment. We model individual ratings while accounting for raters’ political ideology, aligning with the bridging principle underlying Community Notes. We treat this rating-level analysis as our primary analysis because it uses the full set of available ratings and, by construction, isolates note helpfulness from platform-level ex- 

> 3https://communitynotes.x.com/guide/en/under-the-hood/ranking-notes#matrix-factorization 

> 4Community Notes computes note helpfulness scores via matrix factorization with regularization that penalizes the note intercept term with a fixed shrinkage toward zero. This penalty is constant regardless of how many ratings a note receives, but the data signal that pushes the intercept away from zero scales with rating count. As a result, notes with fewer ratings have insufficient data to overcome the regularization pressure, leading to systematically deflated helpfulness scores. 

4 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

posure effects. Second, we perform a note-level analysis designed to equalize exposure across LLM and human notes, by re-calculating helpfulness scores using only ratings from raters who saw and rated all notes on a given tweet, equalizing exposure across LLM and human notes before applying the Community Notes matrix factorization-based bridging algorithm. 

## **2.2 LLM Notes Receive More Positive Ratings Across the Ideological Spectrum** 

Community Notes is designed around the principle of “bridging”: a note is considered helpful if it is rated as such by raters with different viewpoints, and operationalizes it by estimating a numeric factor for each rater that captures systematic differences in rating behavior. In practice, this rater factor aligns closely with political ideology, with negative values corresponding to left-leaning rating patterns and positive values to right-leaning ones (Wojcik et al., 2022). Motivated by this design, our rating-level analysis examines how individual raters evaluate LLM notes versus human-written notes across the ideological spectrum. By leveraging the rater factor, we test whether LLM notes receive more positive ratings that hold across raters with different inferred political viewpoints—the core criterion by which Community Notes defines helpfulness. 

**==> picture [422 x 251] intentionally omitted <==**

**Figure 1:** Mean % helpful and % unhelpful ratings per note for LLM and human notes, stratified by rater ideology group (left, neutral, right). Error bars show 95% confidence intervals across notes. 

5 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

To examine whether rating patterns varied across the political spectrum, we stratified raters into three ideology groups based on their rater factor: left (factor _< −_ 0.15), neutral _−_ ( 0.15 _≤_ factor _≤_ 0.15), and right (factor _>_ 0.15). For each note and ideology group, we computed the percentage of ratings marked helpful and not helpful, then aggregated across notes to obtain mean percentages with 95% confidence intervals. Results are pre1. sented in Figure LLM notes have higher avg. % helpful ratings and lower avg. % unhelpful ratings than human notes across all rater ideology groups. The largest difference appears among neutral raters, followed by left-leaning raters, and right-leaning raters. 

To quantify differences in individual ratings between LLM and human notes, we estimated a linear mixed effects model predicting individual rating scores from LLM writer authorship, rater political leaning, and their interactions. This analysis mirrors the “bridging” principle of Community Notes algorithm by modeling how the LLM advantage varies as a function of rater ideology, and the extremity of their ideology. Community Notes allows users to rate each note as Helpful, Somewhat Helpful, or Not Helpful; following the platform convention, we coded these responses numerically as 1.0, 0.5, and 0.0, respectively. Specifically, we use the following primary specification: 

**==> picture [408 x 35] intentionally omitted <==**

where AI = 1 for LLM notes and 0 for human notes, and coreRaterFactor1 is each rater’s ideology factor, with negative values indicating left-leaning rating patterns and positive values indicating right-leaning patterns. The model includes both the linear and quadratic forms of the ideology factor, each interacted with LLM authorship, to allow the LLM effect to vary asymmetrically across the political spectrum, consistent with patterns 1. observed in Figure The linear ideology term and its interaction with LLM capture directional differences in whether the LLM advantage is larger among left- or right-leaning raters. The quadratic term and its interaction capture extremity effects in whether ratings decline and the LLM advantage narrows as raters become more ideologically extreme in either direction. Note and rater random intercepts account for within-note correlation of ratings and between-rater heterogeneity in baseline rating tendencies. Table 1 shows that LLM notes receive significantly more positive ratings at the center of the ideology spectrum (AI coefficient = 0.104, _p <_ 0.001), given that human notes’ average helpful ratings achieve around 78.5% from centrist raters, this corresponds to approximately a 10-percentage-point increase. The negative quadratic interaction indicates that this advantage is largest among moderate raters and diminishes toward both ideological extremes, while the linear interaction indicates that the advantage decreases more steeply among right-leaning raters. 

6 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

**Table 1: Rating-Level linear mixed effects model results.** To assess robustness, we estimate three additional specifications. Model 2 replaces note random intercepts with tweet random intercepts that absorbs tweet-level confounders while still identifying the AI advantage partly from within-tweet variation. Model 3 is an OLS regression with standard errors clustered by note. Model 4 replaces the continuous coreRaterFactor1 with categorical rater group (left-leaning, neutral, right-leaning; reference: neutral). The estimates of AI advantage stay consistent across all specifications. 

||Note + Rater RE|Tweet + Rater RE|OLS Clustered|Note + Rater RE<br>(Group)|
|---|---|---|---|---|
|(Intercept)|0.785***|0.842***|0.826***|0.785***|
||(0.006)|(0.005)|(0.003)|(0.007)|
|AI|0.104***|0.066***|0.082***|0.098***|
||(0.009)|(0.004)|(0.004)|(0.010)|
|coreRaterFactor1|-0.015***|-0.007|-0.005||
||(0.004)|(0.004)|(0.003)||
|coreRaterFactor12|-0.175***|-0.195***|-0.291***||
||(0.010)|(0.010)|(0.009)||
|AI:coreRaterFactor1|-0.087***|-0.112***|-0.120***||
||(0.005)|(0.005)|(0.005)||
|AI:coreRaterFactor12|-0.190***|-0.169***|-0.181***||
||(0.013)|(0.013)|(0.014)||
|left-leaning rater||||-0.031***|
|||||(0.005)|
|right-leaning rater||||-0.059***|
|||||(0.005)|
|AI:left-leaning rater||||-0.003|
|||||(0.007)|
|AI:right-leaning rater||||-0.086***|
|||||(0.007)|
|SD (Intercept|0.160|0.160||0.164|
|raterParticipantId)|||||
|SD (Observations)|0.304|0.335||0.304|
|SD (Intercept noteId)|0.196|||0.197|
|SD (Intercept tweetId)||0.142|||
|Num.Obs.|108169|108169|108169|108169|



* _p <_ 0.05, ** _p <_ 0.01, *** _p <_ 0.001. Standard errors in parentheses. Models 1–2: model-based SEs from mixed model. Model 3: CR2 cluster-robust SEs by note. Model 4: mixed model with categorical rater group. Reference category is neutral. 

7 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

**==> picture [422 x 242] intentionally omitted <==**

**Figure 2:** LLM vs. human note rating advantage (AI main-effect coefficient with 95% CI) by tweet modality (text-only, image, video). 

**==> picture [422 x 241] intentionally omitted <==**

**Figure 3:** LLM vs. human note rating advantage (AI main-effect coefficient with 95% CI) by tweet topic category. 

8 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

To examine how the LLM advantage in helpfulness ratings varies across tweet types, we conducted exploratory subgroup analyses by tweet modality and topic. We used an LLM to assign a topic to each tweet based on its content (see Appendix E.4 for prompt). We subset notes by each category and re-estimated our main rating-level specification within each subset. Figures 2 and 3 present the coefficient on the LLM main effect with 95% confidence intervals. The LLM advantage is consistently positive across all three modality types but varies in magnitude: the point estimate is largest for text-only posts, followed by posts with images and videos. Topic-level heterogeneity is more pronounced. LLM notes receive better ratings on posts about health and medicine, and conspiracy theories and pseudoscience claims. Notably, LLM notes show minimal or non-significant advantages on posts about AI-generated content. These patterns suggest that the LLM writer is more effective at fact-checking claims in domains with well-established authoritative sources and clear factual grounding. 

## **2.3 Equal-Exposure Note-Level Analysis** 

The rating-level analysis demonstrates that LLM notes receive more positive individual evaluations across the ideological spectrum. We complement this with a note-level analysis to investigate whether this advantage translates into better note-level outcomes. In the full sample, the comparison shows mixed signals. LLM notes accumulate fewer ratings on average (59.50 vs. 109.94), and have a lower CRH rate (13.07% vs. 18.02%, see Appendix B), while having similar note helpfulness scores (mean: 0.25 vs. 0.24) and a lower rate of reaching Currently Rated Not Helpful status (1.12% vs. 4.13%). However, as discussed in the first section of Results, direct comparison of platform note-level outcomes is confounded by different exposures, and LLM notes are systematically submitted later and consequently accumulate fewer ratings. To address this, we construct a subset of ratings that equalizes exposure by design. For each tweet, we retain only ratings from raters who saw and rated every note on that tweet, ensuring that each note within a tweet received the same set of evaluations. We then recompute note helpfulness scores by applying the Community Notes scoring algorithm to these equal-exposure ratings alone. This yields a subset of 13,721 ratings across 1,674 notes on 663 tweets. A potential concern with this approach is that raters who evaluated all notes on a given tweet may represent a selective subset of the overall rater population. To assess this, we compare the distributions and find these raters do not differ substantially from the broader population in terms of their overall tendency to rate notes as helpful or their political leaning (see Appendix D for details). 

We then estimated the differences in re-computed note-level outcomes with a linear mixed effects model: 

outcome _∼_ AI + (1 _|_ tweetId) (2) 

9 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

where AI = 1 for LLM-written notes and 0 for human-written notes, and tweet random intercepts account for within-tweet dependence and absorb tweet-level confounders such as topic difficulty and post virality. We use mixed models rather than standard parametric tests because notes targeting the same tweet are not independent, violating the independence assumption. We use this model to analyze three note-level outcomes: note helpfulness score, CRH status, and CRNH status, and report Benjamini–Hochberg adjusted p-values. 

Under LLM notes achieve scores than equal exposure, significantly higher helpfulness human notes (mean: 0.21 vs. 0.18; AI coefficient = 0.019, _z_ = 2.944, _adj_ . _p_ = 0.010), confirming that the LLM quality advantage identified in our rating-level analysis carries through to the note-level metric once the exposure confound is removed. We also derive final rating status from the re-computed helpfulness scores: a higher proportion of LLM notes reach CRH status (2.40% compared to 1.89% for human notes) and a lower proportion reach CRNH status (0.90% compared to 1.39% for human notes), though these differences are not statistically significant. Because restricting to equal-exposure raters substantially reduces the number of ratings per note, both the note helpfulness scores and the proportion of notes reaching CRH status are lower than those observed in the full sample (see Appendix B). We report further note-level robustness checks in the Appendix C, including restricting the sample to notes with at least 30 ratings to mitigate the algorithm’s penalty on low-rating-count notes, and creation-time–matched analyses to control for differences in exposure. 

## **2.4 LLM and Human Notes Differ in Writing and Sources** 

LLM and human notes differed in both writing and source citation behaviors. LLM notes were much longer than human notes (mean 35.8 words vs. 26.9 words, _t_ = 22.190, _p <_ 0.001). The LLM writer also cited more URLs on average (mean: 1.51 vs. 1.23 URLs; _t_ = 8.67, _p <_ 0.001). Analysis of cited domains reveals distinct sourcing patterns (Tables 2a and 2b). LLM notes most frequently referenced mainstream news outlets and social media platforms, including reuters.com (7.7% of LLM notes), en.wikipedia.org (7.2%), instagram.com (5.8%), youtube.com (5.3%), x.com (4.2%), bbc.com (4.2%), snopes.com (3.9%), cnn.com (2.4%), facebook.com (2.0%), and yahoo.com (2.0%). In contrast, human notes most commonly cited x.com (18.7%), nearly four times the rate observed in LLM notes (4.2%). Wikipedia appeared at similar rates in both groups (7.6% human vs. 7.2% LLM), but other mainstream news sources appeared less frequently in human notes: reuters.com (1.6% vs. 7.7%), bbc.com (1.1% vs. 4.2%), and snopes.com (0.7% vs. 3.9%). This suggests that the LLM writer relies more heavily on traditional authoritative sources, whereas human writers more frequently reference platform-native content and social media posts. We note that this pattern characterizes our specific implementation. 

10 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **(a) Top 10 domains in LLM notes.** 

|Rank|Domain|% LLM notes citing|% human notes citing|
|---|---|---|---|
|1|reuters.com|7.7|1.6|
|2|en.wikipedia.org|7.2|7.6|
|3|instagram.com|5.8|2.0|
|4|youtube.com|5.3|3.9|
|5|x.com|4.2|18.7|
|6|bbc.com|4.2|1.1|
|7|snopes.com|3.9|0.7|
|8|cnn.com|2.4|0.8|
|9|facebook.com|2.0|0.7|
|10|yahoo.com|2.0|0.5|



## **(b) Top 10 domains in human notes.** 

|Rank|Domain|% LLM notes citing|% human notes citing|
|---|---|---|---|
|1|x.com|4.2|18.7|
|2|en.wikipedia.org|7.2|7.6|
|3|youtube.com|5.3|3.9|
|4|x.com/grok|0.4|2.6|
|5|instagram.com|5.8|2.0|
|6|reuters.com|7.7|1.6|
|7|theguardian.com|1.2|1.6|
|8|t.co|0.0|1.1|
|9|bbc.com|4.2|1.1|
|10|share.google|0.0|1.0|



11 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **3 Discussion** 

Our results demonstrate that LLM-generated Community Notes can achieve broader cross-ideological acceptance than human-written notes, receiving more positive ratings from raters across the political spectrum. This finding is particularly significant given that Community Notes deliberately prioritizes cross-perspective consensus over simple accuracy or neutrality. The fact that LLMs perform better than human notes on average on this standard suggests they are capable of being broadly helpful rather than merely presenting factual information, a capability with important implications for scaling content moderation while maintaining or even improving the quality of contextual information provided to users. The topic-level heterogeneity we observe offers insight into where LLM fact-checking is most and least effective. The LLM writer’s strongest advantages appear on health and medicine claims and conspiracy theories/pseudoscience, domains where authoritative sources are abundant and factual grounding is relatively clear-cut. By contrast, the LLM shows minimal advantage on posts about AI-generated content, where detection requires specialized capabilities which our current LLM does not have access to. However, tools like reverse image search or Google’s SynthID could be easily integrated into the LLM pipeline to further improve our agent’s capability to detect AI-generated content. More broadly, our results reflect the performance of a single LLM writer pipeline; other AI writer implementations on the platform may differ in relative strengths across topics and overall performance. 

These findings should be interpreted alongside important platform-specific confounds. LLM notes currently face a structural timing disadvantage: they can only be written after sufficient users flag a post, meaning they are systematically submitted later and accumulate fewer ratings. Because the Community Notes algorithm penalizes notes with fewer ratings, this timing disadvantage could deflate LLM notes’ helpfulness scores. Our notelevel analysis that equalizes exposures and robustness checks address this issue in several ways, suggesting that LLM notes receive higher note helpfulness scores when differential exposures are controlled. Together, the pattern indicates that the core quality signal favors LLM notes, while timing-induced exposure effects obscure this signal in aggregated platform metrics. 

Our findings have broader implications for the role of LLMs in public information ecosystems (Li et al., 2025). To our knowledge, this study is the first and the largest field evaluation of AI-generated fact-checking on any social media platform, made possible by the transparent data access provided by Community Notes. This open data enables a truly ecological and valid assessment: rather than relying on recruited evaluators or proxy metrics, we examine how a diverse population of real platform users respond to AI-generated content under natural conditions. 

12 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

Notes a distinctive for AI notes are Community provides testing ground fact-checking: evaluated by a diverse, self-selected rater population on a live platform. The fact that LLMs can meet this standard at scale suggests that AI-augmented fact-checking is viable not merely in controlled offline settings, but under the conditions of real social media platforms. This is particularly consequential given the scale mismatch between misinformation production and human fact-checking capacity. Human contributors, however motivated, are limited in the volume of content they can monitor and annotate. LLM writers face no such constraint, and our results suggest that scaling up AI contributions need not come at the cost of quality or cross-ideological acceptability. At the same time, LLM and human contributors have different strengths that are not easily substituted (Li et al., 2025). LLMs are good at rapidly synthesizing widely available information and generating contextual notes at scale, while human contributors are better equipped to handle novel situations, niche topics, and rapidly evolving events, drawing on deep domain knowledge and an understanding of how posts may be interpreted by diverse audiences. By having LLM- and human-written notes compete within the same rating system, the most helpful notes, regardless of origin, rise to the top. Platforms and researchers designing AI-augmented fact-checking systems should consider how to harness this complementarity that enhances overall note quality rather than treating AI deployment as a replacement for human community participation. 

## **4 Data and Methods** 

## **4.1 Dataset** 

Our analysis dataset comprised 2,946 Community Notes written between November 1, 2025 and January 31, 2026, targeting 1,597 unique tweets, with 108,169 ratings from 42,521 unique raters. The sample included 1,614 notes written by our LLM writer and 1,332 notes written by human users on the same set of tweets. We applied two quality filters: excluding media notes, and filtering ratings to include only those from raters with a valid rater factor. 

## **4.2 Analysis and multiple testing strategy** 

Our analyses operate at two levels. The rating-level analysis estimates a linear mixed effects model predicting individual rating scores, where p-values on regression coefficients are interpreted without multiple-testing correction, as they are jointly estimated within a single model. Subgroup analyses by tweet modality and topic re-estimate this specification within subsets and are treated as exploratory. At the note level, our primary analysis uses the equal-exposure design, in which we test three outcomes, note helpfulness score, CRH status, and CRNH status, using a linear mixed effects model with tweet random 

13 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

intercepts to account for within-tweet dependence among notes targeting the same post. We apply Benjamini-Hochberg correction across these three tests to control the false discovery rate. 

## **4.3 Note Writing Pipeline** 

Upon retrieving a potentially misleading post via the AI writer API, we compile its complete context by concatenating the post’s creation timestamp, its primary text, and any text and image descriptions from quoted or replied-to posts (if applicable). To enable multimodal note writing, the LLM writer is directly provided with images or video thumbnails associated with the target post. Next, we use Grok-4-fast with its web search and X search capabilities to conduct information research and gather relevant evidence. These research outputs are used to inform the note-writing process (Grok was chosen for its ability to surface X posts and understand video content from X, in addition to performing web search.) After evidence collection, the LLM writer (GPT-5-mini) decides whether a note should be written for the post. A note may not be written if the post is unlikely to be perceived as misleading or if there is insufficient evidence. This decision step acts as a safeguard to prevent generation of spam notes on non-misleading posts. Once the decision to write a note is made, the LLM (GPT-5-mini) writes the community note given the post content and Grok research outputs, following Community Notes guidelines (prompts are provided in Appendix E). The note next goes through URL validity check, length check, and quality check with the Community Notes ClaimOpinion model[5] prior to submission. Our LLM writing pipeline code, analysis code, and prompts are publicly available at https://github.com/haiwen-li/cn-ai-writer. 

## **Acknowledgements** 

We thank Dean Eckles, Martin Saveski, Soham De for insightful conversations and feedback on this work. We thank Brad Miller, Jay Baxter, Keith Coleman for valuable discussions on Community Notes and AI fact-checking. 

We used AI to assist with writing and debugging analysis code and to provide feedback on drafts of this paper. All analyses, interpretations, and final writing decisions were made by the authors. 

> 5https://docs.x.com/x-api/community-notes/evaluate-a-community-note 

14 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **References** 

- Luigia Costabile et al. Assessing the potential of generative agents in crowdsourced factchecking. _Online Social Networks and Media_ , 48:100326, 2025. 

- Soham De, Michiel A Bakker, Jay Baxter, and Martin Saveski. Supernotes: Driving consensus in crowd-sourced fact-checking. In _Proceedings of the ACM on Web Conference 2025_ , pages 3751–3761, 2025. 

- Chiara Patricia Drolsbach, Kirill Solovev, and Nicolas Pröllochs. Community notes increase trust in fact-checking on social media. _PNAS Nexus_ , 3(7):pgae217, 2024. 

- Olga Iarygina, Kasper Hornbæk, and Aske Mottelson. Demand characteristics in human– computer experiments. _International Journal of Human-Computer Studies_ , 193:103379, 2025. 

- Haiwen Li, Soham De, Manon Revel, Andreas Haupt, Brad Miller, Keith Coleman, Jay Baxter, Martin Saveski, and Michiel Bakker. Scaling human judgment in community notes with LLMs. _Journal of Online Trust and Safety_ , 3(1), September 2025. ISSN 27703142. doi: 10.54501/jots.v3i1.255. 

- Jonathan Mummolo and Erik Peterson. Demand effects in survey experiments: An empirical assessment. _American Political Science Review_ , 113(2):517–529, 2019. 

- Sahajpreet Singh, Kokil Jaidka, and Min-Yen Kan. GitSearch: Enhancing community notes generation with gap-informed targeted search. _arXiv preprint arXiv:2602.08945_ , 2026. 

- Isaac Slaughter, Axel Peytavin, Johan Ugander, and Martin Saveski. Community notes reduce engagement with and diffusion of false information online. _Proceedings of the National Academy of Sciences_ , 122(38):e2503413122, 2025. doi: 10.1073/pnas.2503413122. 

- Stefan Wojcik, Sophie Hilgard, Nick Judd, Delia Mocanu, Stephen Ragain, MB Hunzaker, Keith Coleman, and Jay Baxter. Birdwatch: Crowd wisdom and bridging algorithms can inform understanding and reduce the spread of misinformation. _arXiv preprint arXiv:2210.15723_ , 2022. 

- Xinyi Zhou, Ashish Sharma, Amy X. Zhang, and Tim Althoff. Correcting misinformation on social media with a large language model. _arXiv preprint arXiv:2403.11169_ , 2024. 

15 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **Appendix** 

## **A Within-Rater Pairwise Comparison** 

As a to the we conducted a re- complement rating-level analysis, pairwise comparison stricted to raters who evaluated both an LLM and a human note on the same tweet. We constructed 1,186 unique LLM–human note pairs, yielding 21,978 rater–pair observations from 10,345 raters. For each observation, we encoded a win/loss/tie outcome based on whether the rater’s rating for the LLM note was better than their rating for the human note. 

We observed a high tie rate: 71% of all rater–pair observations resulted in equal scores for both note types, suggesting that these raters found the two notes equivalently helpful. Excluding ties and fitting a Bradley-Terry logistic model with standard errors clustered by rater, we find that the LLM note was preferred in 54.4% of non-tied head-to-head comparisons ( _β_ = 0.178, SE = 0.035, _z_ = 5.10, _p <_ 0.001; OR = 1.19), significantly above the 50% chance level. 

## **B Full-Sample Note-Level Outcomes** 

We present note-level analysis of note helpfulness scores, CRH status, and CRNH status in the full sample. A higher proportion of human notes achieved CRH status than LLM notes (18.02% vs. 13.07%; AI coef. = _−_ 0.058, _z_ = _−_ 4.556, _p <_ 0.001). And among CRH notes, human notes reached that status faster (median: 5.90 hours vs. 7.38 hours for LLM notes; Mann-Whitney _U_ = 27,979, _p_ = 0.054). However, LLM notes had a lower rate of Currently Rated Not Helpful (CRNH) status (1.12% vs. 4.13% for human notes; AI coef. = _−_ 0.027, _z_ = _−_ 4.714, _p <_ 0.001) and similar average note helpfulness scores (mean: 0.25 vs. 0.24; AI coef. = 0.007, _z_ = 1.143, _p_ = 0.253). These results should be interpreted in light of substantial platform-specific confounds: the Community Notes algorithm used to compute these outcomes penalizes notes with few ratings, and LLM writers face a submission timing disadvantage that results in accumulating fewer ratings. 

To benchmark LLM writer’s performance against the distribution of individual human writers who have written community notes, we compare the CRH rate and hit rate. The CRH rate is the fraction of a writer’s notes achieving CRH status, and the hit rate is de- _−_ fined as (CRH notes CRNH notes) / total notes to account for notes reaching unhelpful status. Our LLM writer achieves a CRH rate of 13.07%, outranking 84.7% of human writers on CRH rate (78.4% among human writers who have written at least 30 notes), and a hit rate of 11.96%, corresponding to the 85.4th percentile among human writers (79.0% 

16 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

among human writers who have written at least 30 notes). 

Tables A1 and A2 summarize the rating-level and note-level results for the full sample, respectively. 

**Table A1: Mean % helpful and % unhelpful ratings per note by rater ideology bucket (full sample).** _n_ denotes the number of notes receiving at least one rating from raters in that ideology bucket. 95% CIs are computed across notes. 

|Rater Ideology|% Helpful, mean [95% CI]<br>LLM<br>Human|% Unhelpful, mean [95% CI]<br>LLM<br>Human|
|---|---|---|
|Left (_n_: 1,300 / 1,183)<br>Neutral (_n_: 1,174 / 1,049)<br>Right (_n_: 1,300 / 1,196)|79.55 [77.83, 81.27]<br>68.10 [65.95, 70.24]<br>89.58 [88.32, 90.84]<br>76.83 [74.86, 78.79]<br>73.13 [71.32, 74.93]<br>67.46 [65.52, 69.41]|17.98 [16.33, 19.63]<br>28.72 [26.61, 30.83]<br>7.60 [6.52, 8.69]<br>18.40 [16.56, 20.25]<br>23.65 [21.92, 25.38]<br>28.58 [26.68, 30.48]|



**Table A2: Note-level comparisons in the full sample.** CRH, CRNH, and helpfulness score are analyzed using linear mixed-effects models (LMMs), and the reported coefficient is the estimated effect of the AI-note indicator. Number of ratings and time to CRH are compared using two-sided Mann–Whitney _U_ tests. 

|Metric|LLM|Human|Test|AI coef.|Statistic|_p_|
|---|---|---|---|---|---|---|
|_N_(total notes)|1,614|1,332|—|—|—|—|
|CRH rate|13.07%|18.02%|LMM|_−_0.058|_z_=_−_4.56|_<_0.001|
|CRNH rate|1.12%|4.13%|LMM|_−_0.027|_z_=_−_4.71|_<_0.001|
|Helpfulness score|0.25|0.24|LMM|0.007|_z_=1.14|0.253|
|(_N_with scores: 1,243 / 1,130)|||||||
|Num. ratings, median|22|51|Mann–Whitney_U_|—|_U_ =740,798|_<_0.001|
|Time to CRH, hrs, median|7.38|5.90|Mann–Whitney_U_|—|_U_ =27,979|0.054|
|(_N_CRH notes: 211 / 240)|||||||



## **C Robustness Checks for Note-Level Outcomes** 

In addition to the note-level analysis with equal-exposure raters, we conducted two robustness checks to address potential confounds in the note-level outcomes. To address the concern that notes with fewer ratings have over-penalized note helpfulness scores, we restrict the sample to notes with at least 30 ratings ( _n_ = 1,538 notes; LLM = 680, human = 858). After filtering, the percentage of CRH notes was no longer statistically different (LLM = 27.94%, human = 27.04%; AI coef. = _−_ 0.006, _z_ = _−_ 0.317, _p_ = 0.751), and LLM notes continued to show a significantly lower proportion of CRNH notes (LLM = 0.59%, human = 4.55%; AI coef. = _−_ 0.035, _z_ = _−_ 4.129, _p <_ 0.001). The difference in helpfulness scores increased (AI coef. = 0.019, _z_ = 2.410, _p_ = 0.016). Among CRH notes, human notes still reached CRH status more quickly (median = 5.78 hours vs. 7.16 hours; Mann-Whitney _U_ = 24,102, _p_ = 0.098). 

17 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

To account for differential note exposure due to submission timing, we conduct creation time-matched analyses. We retain human notes created within _±_ 30, _±_ 60, or _±_ 90 minutes of the LLM note creation on the same post, under the assumption that notes written close in time have similar Because LLM and human notes are often written hours exposure. apart, matching rates were low (6.3%, 11.8%, and 15.9%, respectively) and sample sizes were substantially reduced. Within the _±_ 60-minute window ( _n_ = 405 notes; LLM = 190, human = 215), the CRH gap was narrow and non-significant (LLM = 21.58% vs. human = 22.79%; AI coef. = _−_ 0.019, _z_ = _−_ 0.610, _p_ = 0.542). The difference in average note helpfulness scores was larger but not statistically significant (LLM = 0.28 vs. human = 0.25; AI coef. = 0.023, _z_ = 1.703, _p_ = 0.089). In this matched sample, LLM notes reached CRH status faster (median = 4.61 hours vs. 5.55 hours). 

Tables A3 and A4 present the full note-level statistics for the _≥_ 30-ratings subset and timing-matched subsets. 

**Table A3: Note-level comparisons in the subset of notes with at least 30 ratings.** CRH, CRNH, and helpfulness score are analyzed using linear mixed-effects models (LMMs); number of ratings and time to CRH are compared using two-sided Mann–Whitney _U_ tests. 

|Metric|LLM|Human|Test|AI coef.|Statistic|_p_|
|---|---|---|---|---|---|---|
|_N_(total notes)|680|858|—|—|—|—|
|CRH rate|27.94%|27.04%|LMM|_−_0.006|_z_=_−_0.32|0.751|
|CRNH rate|0.59%|4.55%|LMM|_−_0.035|_z_=_−_4.13|_<_0.001|
|Helpfulness score|0.31|0.28|LMM|0.019|_z_=2.41|0.016|
|(_N_with scores: 645 / 798)|||||||
|Num. ratings, median|67.00|91.50|Mann–Whitney_U_|—|_U_ =236,530.5|_<_0.001|
|Time to CRH, hrs, median|7.16|5.78|Mann–Whitney_U_|—|_U_ =24,102|0.098|
|(_N_CRH notes: 190 / 232)|||||||



18 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

**Table A4: Note-level comparisons in timing-matched subsets.** Each sub-table restricts analysis to LLM notes matched to human notes within the indicated submission-time window. CRH, CRNH, and helpfulness score are analyzed using linear mixed-effects models (LMMs); number of ratings and time to CRH are compared using two-sided Mann–Whitney _U_ tests. 

|Metric|LLM|Human|Test|AI coef.|Statistic|_p_|
|---|---|---|---|---|---|---|
|_N_(total notes)|102|116|—|—|—|—|
|CRH rate|21.57%|20.69%|LMM|0.004|_z_=0.09|0.929|
|CRNH rate|1.96%|6.90%|LMM|_−_0.049|_z_=_−_1.74|0.082|
|Helpfulness score|0.28|0.24|LMM|0.041|_z_=2.13|0.033|
|(_N_with scores: 96 / 109)|||||||
|Num. ratings, median|64.00|55.00|Mann–Whitney_U_|—|_U_ =6,491.5|0.216|
|Time to CRH, hrs, median|6.01|5.67|Mann–Whitney_U_|—|_U_ =288|0.605|
|(_N_CRH notes: 22 / 24)|||||||



## **(a)** _±_ 30-minute window. Matched 102/1,614 LLM notes (6.3%). 

|Metric|LLM|Human|Test|AI coef.|Statistic|_p_|
|---|---|---|---|---|---|---|
|_N_(total notes)|190|215|—|—|—|—|
|CRH rate|21.58%|22.79%|LMM|_−_0.019|_z_=_−_0.61|0.542|
|CRNH rate|2.11%|6.05%|LMM|_−_0.039|_z_=_−_1.98|0.048|
|Helpfulness score|0.28|0.25|LMM|0.023|_z_=1.70|0.089|
|(_N_with scores: 178 / 193)|||||||
|Num. ratings, median|54.00|57.00|Mann–Whitney_U_|—|_U_ =21,064.5|0.587|
|Time to CRH, hrs, median|4.61|5.55|Mann–Whitney_U_|—|_U_ =950|0.662|
|(_N_CRH notes: 41 / 49)|||||||



## **(b)** _±_ 60-minute window. Matched 190/1,614 LLM notes (11.8%). 

|Metric|LLM|Human|Test|AI coef.|Statistic|_p_|
|---|---|---|---|---|---|---|
|_N_(total notes)|257|313|—|—|—|—|
|CRH rate|19.46%|20.13%|LMM|_−_0.013|_z_=_−_0.47|0.638|
|CRNH rate|1.56%|5.43%|LMM|_−_0.037|_z_=_−_2.37|0.018|
|Helpfulness score|0.27|0.24|LMM|0.027|_z_=2.37|0.018|
|(_N_with scores: 240 / 285)|||||||
|Num. ratings, median|57.00|59.00|Mann–Whitney_U_|—|_U_ =40,772|0.778|
|Time to CRH, hrs, median|4.96|5.46|Mann–Whitney_U_|—|_U_ =1,562|0.942|
|(_N_CRH notes: 50 / 63)|||||||



**(c)** _±_ 90-minute window. Matched 257/1,614 LLM notes (15.9%). 

19 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

## **D Representativeness of Complete Raters** 

We compare the distribution of rater characteristics between the full rater population and the subset of “complete raters” who evaluated all notes on a given tweet. The left panel shows the distribution of coreRaterIntercept, which measures a rater’s baseline tendency to rate notes as helpful. The right panel shows the distribution of coreRaterFactor1, which captures political leaning inferred from historical rating patterns (negative = left-leaning, positive = right-leaning). Both distributions are highly similar across the two groups, suggesting that complete raters do not systematically differ from the overall population in helpfulness leniency and political leaning (Figure A1). Together, these comparisons indicate that restricting the analysis to raters who rated all notes on a tweet does not substantially alter the distribution of key rater characteristics. 

**==> picture [468 x 193] intentionally omitted <==**

**Figure A1:** Distribution of rater characteristics for the full rater population vs. complete raters who evaluated all notes on a given tweet. Left: coreRaterIntercept captures baseline helpfulness leniency. Right: coreRaterFactor1 captures political leaning (negative = leftleaning, positive = right-leaning). The close overlap indicates that complete raters are not systematically different from the overall rater population. 

## **E LLM Writing Pipeline Prompts** 

## **E.1 Evidence Retrieval Prompt** 

```
"""InvestigatewhethertheXpostbelowismisleading.Imagesorvideopreviews
fromthepostareprovidediftheyexist.
```

```
Step1:Understandthepostanditscontext
```

- `Find the post by post id {post_id} on X. If not found, search the post’s main text on X.` 

- `Identify the post author. Also note any signals that the account is` 

20 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

```
parody/satire(e.g.bio,handle,verification).
```

- `Fetch the thread and top comments. Understand how others interpret the post` 

   - `(e.g. joke/satire, potentially misleading, expressing an opinion).` 

```
-Summarizethepostcontextin1-2sentenceswithoutrewordingthepostitself.
IncludewhotheOPisandhowothersinterpretthepost,especiallyifcomments
suggestit’sajokeorsatire,orprovidepotentialfact-checkdirections.
Forexample:"TheOPisSen.X.ManycommentssayhisclaimaboutXis
unverified."or"TheOPisaparodyaccount,andcommentersarelaughingand
treatingitasajoke."
```

```
Step2:Searchforevidence
```

`1. Search both the web and X for factual sources that refute or confirm the post’s claims. Use the post context to guide your search if it could provide potential fact-check directions.` 

`2. Aim for {target_url_count} pieces of evidence / URLs if possible.` 

`3. For each source, include the URL and a brief note describing how it verifies or challenges the post. Include the publication date of the source if available.` 

`4. Cover outlets across the ideological spectrum (left, center, right). Overlapping reasoning is acceptable when it comes from different publishers.` 

`5. Prioritize evidence that is relevant, solid, and up to date.` 

```
Targetpost(ID:{post_id}):
{post}
```

```
YourresponseshouldbereturnedasaJSONobjectwiththefollowingstructure:
‘‘‘
```

```
{{
"post_context":"one/two-sentencesummaryofthepostcontext",
"research":[
{{"url":"url1","description":"howthecontentoftheURLfact-checksthepost
"}},
...
]
}}
‘‘‘
```

```
Ifyoucannotfindsufficientevidencetofact-checkthepost,returnanempty
researcharray."""
```

## **E.2 Note Triage Prompt** 

## **System prompt:** 

```
"""DecidewhetherapostneedsaCommunityNotebasedontheprovidedevidence.
```

```
Returns:
```

```
-"WRITENOTE"ifanoteshouldbewritten
```

21 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

- `"NO NOTE NEEDED" if the post doesn’t need a note` 

- `"NOT ENOUGH EVIDENCE" if there’s insufficient evidence` 

```
"""
```

## **User prompt:** 

```
"""YouareaCommunityNoteswriter.Yourjobistodecideifthetargetpost
couldbeperceivedasmisleadingandwhetheritneedsacommunitynotetoaddress
itsissues.Theoutputshouldbeoneof:"WRITENOTE","NONOTENEEDED",or
"NOTENOUGHEVIDENCE".
```

## `Task rules` 

- `Focus on the main claims of the post, not trivial errors.` 

- `If media is included, use any legible text, recognizable logos/landmarks, and clearly identifiable public figures as part of the claim.` 

- `If the target post is a quote/reply, evaluate the target post, using the quoted/replied post only as context.` 

```
-Usethecontextinformation(authorinfoandaudiencereactions)toinformyour
decision.Theyarehelpfulforunderstandingwhetherthepostislikelytobe
perceivedasmisleading.
```

- `If you are unsure, err on the side of NO NOTE NEEDED.` 

```
Decisionlogic
```

```
OutputNONOTENEEDEDif:
```

```
-TheOPisasatire/parodyaccount,orthepostisjoking/sarcastic/exaggerated
tobeironic,andcommentersareinterpretingitthatway(withnostrongsigns
ofmisunderstanding);or
```

```
-Thepostismostlyopinion,subjectivetakes,orpersonalexperience;or
```

- `The post contains no major factual claims, or the claims are not verifiable; or` 

- `The post contains factual claims, but the provided evidence indicates those claims are accurate or not meaningfully misleading.` 

```
Ifthepostcontainsmajorfactual,verifiableclaimsandtheevidenceis
relevant:
```

```
-OutputNOTENOUGHEVIDENCEiftheevidenceisweak,mixed,orinsufficientto
confidentlyverifyorrefutethemainclaims.Whenunsureaboutthe
strength/sufficiencyoftheevidence,erronNOTENOUGHEVIDENCE.
```

```
-OutputWRITENOTEonlyiftheevidenceclearlyshowsthatthepost’smain
claimsarefalseormisleadinginawaythatcouldmisinformareasonable
reader.
```

```
Post:
```

```
{post}
```

```
Additionalcontextinformationaboutthepost:
{post_context}
```

22 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

```
Evidence:
{evidence}
```

```
Outputonlyoneof:"WRITENOTE","NONOTENEEDED",or"NOTENOUGHEVIDENCE".
ErronthesideofNONOTENEEDEDifunsure."""
```

## **E.3 Note Generation Prompt** 

## **System prompt:** 

```
"""Youareahelpfulfact-checkingassistant.
```

```
YourgoalistowritegoodCommunityNotesthatwouldbeapprovedhelpfulby
peoplewithdifferentviewpoints.
```

```
Donotinventfactsormakeclaimsthatarenotsupportedbytheprovided
evidence."""
```

## **User prompt:** 

```
"""Task:Writeacommunitynoteforthetargetpostbelow.Imagesorvideo
previewsfromthetargetpostmaybeprovided;ifpresent,analyzeanylegible
text(OCR),recognizablelogos/landmarks,andconfidentlyidentifiablepublic
figurestoinformthenote.Additionalcontextprovidespostauthordetailsand
audiencereactions.Ifthetargetpostquotes/repliestoanotherpost,useit
onlyforcontextandfocusonthetargetpost.
```

```
HardConstraints:
```

`1. The note is written to explain why the post is misleading and add additional context to the post. Focus on primary claim(s) of the post rather than trivial details.` 

`2. The note must be grounded in the provided evidence and should cite the URL of the evidence it uses. At least one URL must be cited.` 

`3. Keep the note strictly under 280 characters. Stay neutral and clear.` 

`4. No hashtags, emojis, unnecessary words. No markdown, brackets, or parentheses around URLs. Do not mention "this note" or "the prompt."` 

```
Targetpost:
‘‘‘
{post}
‘‘‘
```

```
Additionalcontextaboutthepost:
‘‘‘
{post_context}
‘‘‘
```

```
Allowedevidencesources:
‘‘‘
```

23 

AI Fact-Checking in the Wild: A Field Evaluation of LLM-Written Community Notes on X 

```
{evidence}
‘‘‘
```

```
Outputonlythefinalnote(atmost280characters)."""
```

## **E.4 Topic Classification Prompt** 

```
"""ClassifythefollowingX/Twitterpostintoexactlyoneofthesecategories:
```

- `Politics & Elections: U.S./global elections, politicians, voting, partisan claims, policy debates.` 

```
-Geopolitics&InternationalConflicts:Wars,foreignpolicy,terrorism,
diplomacy,country-specificevents(e.g.,Ukraine,Israel-Palestine,
China-Taiwan).
```

```
-Health&Medicine:Diseases,treatments,vaccines,publichealthpolicies,
medicaladvice,COVID.
```

```
-Social/CulturalIssues:Abortion,gender/LGBTQ+,race/DEI,guns,
crime/justice,education,religion.
```

```
-Economy&Finance:Inflation,jobs,taxes,crypto,markets,inequality.
```

```
-Science&Technology:Climatechange,AI,space,gadgets(non-healthtech).
-Conspiracy&Pseudoscience:Generalconspiracies,electionfraudclaimsnot
tiedtoactivepolitics,QAnon-style,flatearth,etc.
```

```
-Celebrity/Entertainment/Viral:Non-politicalhoaxes,celebritydrama.
-Sports:Sportsevents,sportsscandals,andrelevantdiscussion.
```

```
-AIgeneratedcontent:AIgenerated/modifiedcontent.
```

```
-Other/Miscellaneous:Everythingelse(weather,personal,ads,neutralnews
withoutcontroversy).
```

```
Posttext:
{tweet_text}
{context_section}
Answerwithonlytheexactcategorynamefromthelistabove
(e.g.,"Politics&Elections"or"Other/Miscellaneous")."""
```

24