Analyze Community Notes Consensus Patterns in Bulk

TL;DR — The X Community Notes Scraper exports filtered slices of the public notes dataset — with per-note reasons, source flags, and showing status — as JSON. That's the raw material for studying when crowdsourced correction actually reaches users.

Why consensus patterns matter

Community Notes' promise is that good corrections surface. Whether they do — and which kinds of claims get corrected versus contested forever — is an empirical question answerable from the public dataset X publishes.

The job-to-be-done is a corpus with the denominator: not just showing notes, but all notes with each one's status — so "what fraction of corrections reach users" is computable.

How does this compare to the alternatives?

How to build the analysis corpus in 4 steps

How do I pull a denominator corpus?

import os
from apify_client import ApifyClient

client = ApifyClient(os.environ["APIFY_TOKEN"])
run = client.actor("thirdwatch/x-community-notes-scraper").call(
    run_input={
        "classification": "all",
        "includeStatus": True,
        "includeHistorical": True,
        "sinceDate": "2026-06-01",
        "maxResults": 1000,
    }
)
notes = list(client.dataset(run["defaultDatasetId"]).iterate_items())

How do I compute the showing rate?

import pandas as pd
df = pd.DataFrame(notes)
rate = df.groupby("isMisleading")["isShowingOnX"].mean()
print(rate)

The gap between misleading-claim corrections that show and those that stall is the consensus finding.

How do I profile correction types?

reasons = df.explode("reasons")
tbl = reasons.groupby("reasons")["isShowingOnX"].agg(["count", "mean"])
print(tbl.sort_values("count", ascending=False).head(10))

Reason categories correlate with consensus odds — sourced factual errors resolve; subjective framings linger contested.

How do I control for sources and media?

print(df.groupby(["trustworthySources", "isMediaNote"])["isShowingOnX"].mean())

Source-cited notes on media posts behave differently than unsourced ones — the dataset's design lets you measure it.

Sample output

{"noteId": "1834567890123456789", "tweetId": "1834560000000000000",
 "summary": "Official figures show the opposite trend…",
 "classification": "MISINFORMED_OR_POTENTIALLY_MISLEADING",
 "isMisleading": true,
 "reasons": ["misleadingFactualError"],
 "trustworthySources": true, "believable": "yes",
 "validationDifficulty": "easy",
 "isMediaNote": false, "isCollaborativeNote": false,
 "createdAt": "2026-07-02T14:05:00Z",
 "status": "NEEDS_MORE_RATINGS", "isShowingOnX": false,
 "statusLabel": "Needs more ratings"}

Common pitfalls

Status reflects the snapshot — a contested note may show later; longitudinal runs catch transitions. believable/harmful/validationDifficulty are contributor judgments, not ground truth. Volume is large — filter first, analyze second. The actor reads X's public data; rater-level data lives in a different X file beyond this export.

Related use cases

• Export Community Notes for misinformation research
• Build a dataset of visible fact-check notes
• Guide to scraping social media
• Blog hub