💻 computer science

How Human Feedback Shapes AI-generated Community Notes

This paper analyzes a large corpus of collaborative notes on X to demonstrate that while human feedback significantly improves the helpfulness of AI-drafted content—particularly through factual corrections and challenging claims—these hybrid notes face adoption bottlenecks due to limited participation and ultimately serve a complementary role by targeting posts overlooked by purely human or AI-only moderation.

Original authors: Soham De, Isaac Slaughter, Jiawei Guo, Qiao-Yun Cheng, Jiayuan Yan, Sruti Banerjee, Martin Saveski

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Soham De, Isaac Slaughter, Jiawei Guo, Qiao-Yun Cheng, Jiayuan Yan, Sruti Banerjee, Martin Saveski

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 a massive town square where people post messages, and sometimes those messages contain lies or half-truths. To keep the square honest, the platform uses a system called Community Notes. Think of this like a neighborhood watch where regular citizens, not professional police officers, step up to write "fact-checks" on misleading posts.

Recently, the platform (X) tried a new experiment: Collaborative Notes. Instead of a human writing a fact-check from scratch, an AI (a robot writer) drafts the note first. Then, human neighbors read that draft, suggest changes, and the AI rewrites it. This happens in a loop until the note is good enough to show everyone.

This paper is like a report card on how well this "Human + Robot" team is working. Here is what the researchers found, explained simply:

1. The Feedback Loop: Who is talking and what are they saying?

When the AI writes a draft, humans step in to critique it. The researchers looked at thousands of these critiques and sorted them into categories:

The "Fact-Checkers": People pointing out errors or adding missing context (e.g., "This source is fake" or "You forgot to mention the date").
The "Editors": People fixing grammar or formatting.
The "Philosophers": People arguing about whether the note should exist or expressing personal moral opinions.
The "Anti-Robot" Crowd: People explicitly saying, "We don't want AI doing this."

The Verdict: The AI is very good at listening to the "Fact-Checkers" and "Editors." If you give it a link to a source or a clear correction, it usually fixes it. However, it largely ignores the "Philosophers" and the "Anti-Robot" crowd. If you just say "I hate this note" without a specific reason, the AI usually keeps the draft as is.

Who is doing the talking? It turns out the people critiquing the AI drafts aren't your average user. They are the "veterans"—people who have been on the platform for years, have rated thousands of notes, and often hold very strong, polarized political views. It's like the town square meeting being dominated by the same 50 loud, experienced neighbors rather than the whole community.

2. The Evolution: Does the note get better?

Imagine a sculpture. The AI starts with a rough block of stone. Humans chisel away, and the AI smooths it out.

Yes, it gets better: On average, the final version of the note is much more helpful than the first draft the AI made.
But it's a bumpy ride: The improvement isn't a straight line up. Sometimes a change makes the note worse before it makes it better. It's like a game of "hot and cold" where the note fluctuates in quality before finally settling on a good version.
More rounds = Better results: Notes that go through many rounds of revision (long chains) end up being much higher quality than those that only get tweaked once or twice.

3. The Reality Check: Is this new system actually working?

Here is the tricky part. Even though the notes do get better with human help, the "Human + Robot" notes are struggling compared to notes written purely by humans or purely by AI.

The "Helpfulness" Gap: Collaborative notes are rated as "helpful" far less often than the other types. Only about 1.8% of them make the cut, compared to 14% for AI-only notes and 9% for human-only notes.
The "Speed" Problem: It takes much longer for a collaborative note to get enough votes to be shown. It's like a slow-moving train compared to the fast cars of the other systems.
The "Crowd" Problem: Why are they failing? Because not enough people are voting on them.
- People seem hesitant to rate AI drafts.
- People get tired. If you rate the first draft, you rarely come back to rate the second or third draft. The "audience" keeps changing, so the note never gets a stable consensus.

4. The Silver Lining: Why keep doing it?

If they are slower and less popular, why bother? The researchers found that these notes play a complementary role.

Think of the town square again. The "Human-only" and "AI-only" teams are busy checking the most obvious, popular lies. But there are thousands of other posts that no one is checking at all. The "Collaborative" team is stepping in to check those forgotten posts. They are filling the gaps where no one else is looking.

Summary

The experiment shows that AI can write a good first draft, and humans can fix it, but the current system has a few bugs:

Too few people are willing to help fix the AI drafts.
The people who do help are mostly the same "veterans" with strong opinions.
The process is too slow, and people lose interest after the first draft.

The paper suggests that to fix this, the platform needs to make it easier for regular people to jump in, perhaps by not forcing them to write long comments, and by encouraging them to come back and vote on the updated versions. If they can get more people to participate, these "Team Human + Robot" notes could become a powerful tool for cleaning up the internet.

Technical Summary: How Human Feedback Shapes AI-generated Community Notes

Problem Statement

Community Notes has emerged as a standard mechanism for crowdsourced fact-checking on social media platforms like X (formerly Twitter), Facebook, and TikTok. While effective at reducing misinformation spread, the system faces persistent challenges regarding coverage (which posts receive notes) and latency (how quickly notes are attached). To address these, X introduced Collaborative Notes, a system where Large Language Models (LLMs) draft initial notes on flagged content, which are then iteratively refined based on human feedback (ratings and free-text suggestions).

The central research problem is to understand the dynamics of this human-AI collaboration: Does the iterative feedback loop actually improve note quality? How does the nature of human input influence these improvements? And ultimately, are these collaborative notes more effective than notes generated solely by humans or solely by AI?

Methodology

The authors conducted a systematic mixed-methods analysis of the complete corpus of Collaborative Notes on X up to May 5, 2026. The dataset comprises 19,146 note versions across 10,600 unique posts and 211,850 instances of human ratings and suggestions.

The analysis was structured around three Research Questions (RQs):

Taxonomy Development (RQ1): To characterize human input, the authors developed a taxonomy of suggestion types. Two authors annotated a random sample of 100 suggestions to derive 13 low-level codes organized into three dimensions: suggestion intent, evidence level, and actionability. This taxonomy was applied to the full dataset using GPT-5.5. They also used a specialized prompt with Grok-4.20-reasoning to determine whether specific suggestions were incorporated into revised note drafts.
Performance Metrics (RQ2 & RQ3):
- Note Helpfulness: Calculated using the Core Matrix Factorization scorer (pre-scoring phase) from the official Community Notes algorithm. This yields a "note intercept" representing how broadly helpful a note is rated across diverse viewpoints.
- Time to Consensus: Measured as the elapsed time from the creation of the first note version to the first transition out of "Needs More Ratings" (NMR) status to either "Currently Rated Helpful" (CRH) or "Currently Rated Not Helpful" (CRNH).
Regression Analysis: Ordinary Least Squares (OLS) regressions with cluster-robust standard errors were performed at the version-to-version transition level. These models analyzed how changes in helpfulness scores correlated with:
- The content of incorporated suggestions (e.g., challenging claims, adding context).
- Rater characteristics (e.g., tenure, rating volume, political lean, writing ability).
Topic Annotation: A multi-label tweet topic classifier was used to categorize the posts receiving collaborative notes, allowing for performance comparisons across different domains.

Key Contributions and Results

1. Characterization of Human Input (RQ1)

Suggestion Types: The most common suggestions (40.1%) challenged the main claims of the post. Other frequent types included meta-policy comments (21.4%) and requests for additional context (15.7%). A notable subset (4.8%) expressed explicit resistance to AI usage.
Incorporation Rates: Suggestions grounded in factual corrections, source critique, and added context were most likely to be incorporated into new drafts. Conversely, subjective policy judgments and expressions of resistance against AI were rarely incorporated.
Rater Demographics: Contributors to collaborative notes differ significantly from those rating standard notes. They are more polarized (higher bimodality in political lean), have longer platform tenure, and possess substantially higher rating volumes. The system is dominated by experienced "veterans" (83.3% have writing privileges).

2. Evolution of Collaborative Notes (RQ2)

Quality Improvement: Human feedback drives meaningful quality gains. Both the final and best versions of collaborative notes are rated significantly more helpful than the initial AI-drafted version ( $p < 0.001$ ).
Non-Monotonicity: Improvement is rarely linear; note trajectories frequently fluctuate before converging on higher quality.
Drivers of Improvement:
- Chain Length: Longer revision chains (7+ versions) are associated with significantly greater quality gains than short chains.
- Suggestion Type: Suggestions that challenge the main claim, offer meta-policy feedback, or include formatting edits are statistically associated with improvements in helpfulness scores.
- Rater Characteristics: Suggestions from high-volume raters lead to larger improvements. Conversely, suggestions from ideologically extreme raters predict smaller improvements, suggesting partisan perspectives may reduce broad helpfulness.

3. Effectiveness and Comparative Performance (RQ3)

Lower Overall Helpfulness: Despite iterative improvements, collaborative notes currently underperform both human-only and AI-only baselines in terms of average helpfulness scores. Only 1.8% of collaborative notes achieve a "helpful" status, compared to 14.1% for AI-only and 9.0% for human-only notes.
Bottlenecks:
- Time to Consensus: Collaborative notes take longer to reach a consensus status, reducing their utility in time-sensitive fact-checking.
- Lack of Ratings: Collaborative notes attract significantly fewer unique raters (mean $\mu=25.7$ ) compared to human-only ( $\mu=75.6$ ) or AI-only ( $\mu=33.1$ ) notes.
- Rater Retention: Engagement drops sharply across versions; only ~41% of raters who evaluated the first version returned to rate the second.
Complementary Role: Despite lower performance metrics, collaborative notes play a complementary role. They predominantly target posts that do not attract human-only or AI-only notes. Of posts with collaborative notes, 76.5% had no other notes proposed, suggesting they expand coverage to otherwise neglected content.

Significance and Claims

The paper claims to provide the first empirical description of a real-world deployment of a collaborative human-AI fact-checking system. Its significance lies in:

Validating the Hybrid Model: It demonstrates that while AI drafts can be improved by human feedback, the process is complex, non-monotonic, and currently limited by structural bottlenecks (specifically, a lack of rater engagement).
Identifying Structural Limitations: The authors highlight that the current system relies heavily on a hyperactive, polarized minority of users, which may introduce noise and instability. The "lack of ratings" is identified as a primary bottleneck preventing collaborative notes from reaching helpful status.
Complementary Utility: The study argues that rather than being a direct substitute for human or AI notes, the collaborative system serves a unique function by covering posts that other mechanisms miss.
Design Pathways: The authors outline concrete design interventions, such as treating ratings cumulatively across versions, guiding raters toward more actionable suggestion types, and reducing friction for new contributors to broaden participation.

The authors maintain a modest stance, noting that as the system is under active development, these findings reflect the current state and that future iterations may alter these dynamics. They emphasize that while the collaborative process improves note quality, it has not yet closed the gap with non-collaborative baselines in terms of overall helpfulness or speed.