💻 computer science

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

This paper presents the first field evaluation of an LLM-based fact-checking system deployed on X's Community Notes, demonstrating through a three-month study that AI-written notes achieve higher helpfulness scores and broader cross-partisan consensus compared to human-written notes when controlling for platform dynamics.

Original authors: Haiwen Li, Michiel A. Bakker

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haiwen Li, Michiel A. Bakker

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine social media as a massive, chaotic town square where everyone is shouting. Sometimes, people shout things that aren't true. To keep the square from turning into a riot, the town has a group of volunteers (the "Community Notes" on X) who step in to add context, like a librarian whispering, "Actually, that story about the moon landing is fake; here's a photo of the launch."

For a long time, scientists have wondered: Could a super-smart robot librarian do this job just as well as, or even better than, human volunteers?

Most previous tests were like putting the robot in a quiet classroom and asking it to take a quiz. The robot did well, but we didn't know if it could handle the noise, the timing, and the chaos of the real town square.

This paper is the first time researchers put a robot librarian into the actual town square to see how it performs in the wild.

The Experiment: The Robot vs. The Humans

The researchers built an AI "writer" and let it loose on X (formerly Twitter) for three months.

The Job: When a post looked suspicious, the AI would search the web, check videos, find the truth, and write a short note explaining it.
The Competition: It wrote 1,614 notes. At the same time, real humans wrote 1,332 notes on the exact same posts.
The Judges: Over 42,000 real users voted on whether these notes were "helpful."

The Big Problem: The "Late Arrival" Penalty

Here is the tricky part. The town square has a weird rule: The earlier you arrive, the more people see you.

Humans could jump in the second they saw a lie.
The AI had to wait. It could only start working after enough humans had already flagged the post as "suspicious."

Because the AI was always a few minutes late, it got fewer votes. It's like a comedian who walks on stage 10 minutes after the show starts; the audience is already leaving, so they don't get to hear the joke, even if the joke was funny. The researchers realized that if they just compared the final scores, the AI would look worse just because it arrived late, not because its notes were bad.

The Solution: The "Fair Play" Test

To fix this, the researchers did two things:

The "Individual Vote" Check: They looked at every single person who voted. Did a left-leaning voter like the AI note? Did a right-leaning voter like it?
The "Same Audience" Check: They only looked at the rare cases where the exact same group of people voted on both the human note and the AI note for the same post. This removed the "late arrival" advantage.

The Surprising Results

When they stripped away the timing advantage, the robot librarian won.

The "Bridge" Effect: The goal of Community Notes is to write something that people on the left, the right, and the middle all agree is helpful. The AI was surprisingly good at this. It wrote notes that were rated as "helpful" by people across the entire political spectrum more often than the human notes were.
The "Goldilocks" Zone: The AI was a superstar when checking facts about health, medicine, and conspiracy theories. It could quickly find authoritative sources (like medical journals or fact-checking sites) and synthesize them perfectly.
The Weak Spot: The AI struggled a bit with posts about AI-generated content (like deepfakes). It's a bit ironic: the robot wasn't great at spotting other robots' work because it didn't have the specialized tools to detect deepfakes yet.

How They Wrote

Humans tended to link to other tweets and social media posts (the "gossip" of the square).
The AI tended to link to major news outlets, Wikipedia, and official fact-checkers (the "encyclopedias").
The AI also wrote slightly longer, more detailed notes.

The Takeaway

This study proves that AI can be a powerful partner in fighting misinformation, but it needs to be used wisely.

AI is great at: Rapidly gathering facts, checking against reliable databases, and writing neutral summaries that don't offend anyone. It's like a tireless research assistant who never gets tired.
Humans are great at: Spotting brand-new, weird trends, understanding cultural context, and handling situations where the rules haven't been written yet.

The Bottom Line: We don't need to choose between robots and humans. The best town square is one where the robot librarian handles the heavy lifting of fact-checking, allowing the human volunteers to focus on the complex, nuanced issues that require a human touch. When they work together, the whole community gets better information.

1. Problem Statement

While Large Language Models (LLMs) show promise for contextual fact-checking (researching, synthesizing evidence, and drafting explanations), prior evaluations have been limited to controlled settings (benchmarks, crowdworkers, or expert annotations). These settings fail to capture critical real-world platform dynamics such as:

Timing and Exposure: The order in which notes appear affects their visibility and rating accumulation.
User Attention: Real users have limited attention spans compared to lab evaluators.
Subjective Judgment: Real-world consensus is harder to achieve than simulated "gold standard" labels.
The Gap: There is a lack of evidence regarding how LLM-generated fact-checks perform in situ on a live social media platform, specifically regarding their ability to achieve the "broadly helpful" consensus required by platforms like X (formerly Twitter).

2. Methodology

The authors conducted the first field evaluation of an LLM-based fact-checking system deployed directly on the X platform via the Community Notes "AI writer" API.

Experimental Setup

Deployment Period: November 1, 2025 – January 31, 2026.
System Architecture: A multi-step pipeline using Grok-4-fast for web/X search and evidence gathering, and GPT-5-mini for decision-making and note generation.
- Multimodal Capability: Handles text, images, and videos.
- Workflow: Retrieve flagged posts $\rightarrow$ Search for evidence $\rightarrow$ Assess sufficiency of evidence $\rightarrow$ Generate note $\rightarrow$ Quality control filters (URL validity, length, claim/opinion check) $\rightarrow$ Publish.
Dataset:
- LLM Notes: 1,614 notes on 1,597 unique tweets.
- Human Notes: 1,332 notes on the same tweets.
- Ratings: 108,169 ratings from 42,521 unique raters.
Control: The study focuses on the authors' specific writer to ensure full transparency and reproducibility of the pipeline.

Analytical Strategies

To address the confounding variable of submission timing (LLM notes are submitted later because they wait for human flags, resulting in fewer ratings), the authors employed two complementary strategies:

Rating-Level Analysis (Primary):
- Models individual rater evaluations (Helpful, Somewhat Helpful, Not Helpful) using a Linear Mixed Effects Model.
- Accounts for rater political ideology (derived from the Community Notes "bridging" algorithm) to test if LLM notes achieve cross-partisan consensus.
- Advantage: Uses the full dataset and isolates note quality from platform exposure effects.
Note-Level Analysis (Equal-Exposure):
- Constructs a subset of ratings where raters evaluated both the LLM and human notes on the same tweet.
- Re-calculates "Helpfulness Scores" and "Currently Rated Helpful" (CRH) status using only these matched ratings to neutralize timing/exposure bias.

3. Key Results

A. Cross-Ideological Consensus (Rating-Level)

Finding: LLM notes received significantly more positive ratings than human notes across the entire political spectrum (Left, Neutral, Right).
Magnitude: Among centrist raters, LLM notes saw an approximate 10-percentage-point increase in "Helpful" ratings compared to human notes.
Ideological Gradient: The advantage was largest for neutral raters, followed by left-leaning, and smallest for right-leaning raters. The advantage diminishes as raters become more ideologically extreme.
Statistical Significance: The AI coefficient in the mixed-effects model was $0.104 $($ p < 0.001$).

B. Note-Level Performance (Equal-Exposure)

Finding: When exposure bias was removed, LLM notes achieved significantly higher Helpfulness Scores (Mean: 0.21 vs. 0.18 for humans; $p = 0.010$ ).
Status: A higher proportion of LLM notes reached the "Currently Rated Helpful" (CRH) status (2.40% vs. 1.89%), though this difference was not statistically significant in the smaller equal-exposure sample.
Conclusion: The quality signal favors LLMs, but in the raw platform data, this is obscured by the algorithmic penalty on notes with fewer ratings (which LLMs receive due to later submission).

C. Topic and Modality Heterogeneity

Strongest Performance: Health/Medicine and Conspiracy Theories/Pseudoscience. These domains have clear authoritative sources and factual grounding.
Weakest Performance: AI-generated content detection. The LLM showed minimal advantage here, likely due to a lack of specialized detection tools (e.g., reverse image search, SynthID) in the current pipeline.
Modality: LLMs performed best on text-only posts, followed by image and video posts.

D. Writing and Sourcing Differences

Length: LLM notes were significantly longer (35.8 words vs. 26.9 words).
Sources:
- LLMs: Relied heavily on traditional authoritative outlets (e.g., Reuters, BBC, Snopes) and Wikipedia.
- Humans: Cited platform-native content (X.com) nearly 4x more often than LLMs (18.7% vs. 4.2%).
- Implication: LLMs prioritize established external authority, while humans rely more on community-sourced platform context.

4. Key Contributions

Methodological: First field evaluation of LLM fact-checking on a live social media platform with organic user feedback, moving beyond synthetic benchmarks to provide ecological validity.
Practical: Open-sourced the entire LLM writer pipeline (prompts, code, multi-step architecture), allowing for reproducibility and further refinement.
Empirical: Demonstrated that LLMs can meet or exceed human performance on the demanding standard of "broad helpfulness," achieving higher cross-ideological consensus than human contributors.

5. Significance and Implications

Scalability vs. Quality: The study proves that scaling fact-checking via AI does not necessarily degrade quality or cross-partisan acceptance. This is crucial given the massive volume of misinformation that outpaces human fact-checking capacity.
Complementarity: The results suggest a future where AI and humans are complementary rather than competitive. LLMs excel at synthesizing widely available information at scale, while humans are better at niche topics, novel events, and interpreting nuanced social context.
Platform Dynamics: The study highlights that raw platform metrics (like CRH rates) can be misleading due to structural biases (e.g., submission timing). Researchers must control for exposure dynamics when evaluating AI systems in the wild.
Policy: Platforms should consider integrating AI writers not as replacements for human communities, but as tools to augment the volume of high-quality, evidence-grounded context provided to users.