# c3: how long a fact-check, a Community Note, a label or a takedown takes, against how fast a fake spreads
Researched 2026-11-18. Each source was fetched and read itself unless marked (summary).

## 1. Community Notes on X

### 1a. Timeliness, Consensus, and Composition of the Crowd: Community Notes on X
URL: https://arxiv.org/html/2510.12559v2 (arXiv 2510.12559, Oct 2025)
- Sample: "approximately 1.8 million notes" (retrieved 1 Jul 2025); timeliness analysis on notes created 1 Jan 2023 to 31 May 2025.
- Post to first helpful note: "an average of 65.7 hours (about 2.7 days) and a median of 15.3 hours". Notes finally rated not helpful: average 94.3 h, median 18.6 h.
- "only 11.5% of notes reach agreement on publication" (raters "reached consensus on 11.5% of all notes").
- Says CNs "primarily function as post-hoc correctives rather than preventative interventions"; cites earlier work that moderation "frequently lags during the initial and often most viral phase". Does NOT itself measure views before the note.
SAYABLE as: "On X, across notes written from 2023 to mid-2025, the median wait from a misleading post to a note showing was about 15 hours; the average was nearly 66 hours, because slow ones drag it up." (median 15.3 h, mean 65.7 h, helpful notes only, X, arXiv Oct 2025)

### 1b. ISD, "The trust-consensus paradox"
URL: https://www.isdglobal.org/digital-dispatch/the-trust-consensus-paradox-why-decentralized-fact-checking-faces-challenges-on-polarizing-topics/
- Period 19 Mar 2024 to 19 Mar 2025. "a median delay of 15.6 hours" (key findings: "more than 15 hours") between a misleading post and a helpful Note. Sample count of posts not stated.
SAYABLE as: "A second study (ISD, March 2024 to March 2025) also found a median of about 15 hours." Two independent datasets agree on roughly 15 h median.

### 1c. Chuai et al., Community Notes on X (arXiv 2409.08781, 2024 preprint)
URL: https://arxiv.org/html/2409.08781v1
- Sample: 237,677 community-fact-checked posts, reposted "more than 431 million times", created 6 Oct 2022 to 11 Jun 2024. 36,073 had displayed notes, 201,604 did not.
- "the average response time was 61.4 hours, with a median of 18.1 hours". "Only 5.2% of community notes were displayed within 4 hours since post creation." "75.9% of all notes deemed helpful were displayed ... within 36 hours".
- "Only a relatively small fraction of 10.1% of all notes was actually displayed"; 83.3% stayed in "Needs More Ratings".
- Repost half-life 5.75 h (within 36 h); "only 13.5% of all helpful notes displayed before this time point".
- NOTE: a secondary wiki claimed "96.7% of reposts before the note"; that figure is NOT in the paper. Do not use.
SAYABLE as: "In a 2022 to 2024 study of 237,000 posts, about one in ten notes was ever shown, and only about one in twenty (5.2%) showed within 4 hours of the post." (X, preprint)

### 1d. Slaughter, Peytavin, Ugander, Saveski, PNAS 2025
URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC12478135/
- 40,078 posts with proposed notes; 6,757 (16.9%) got a helpful note attached (treatment group). Data 16 Mar to 23 Jun 2023.
- After a note appears: repost growth down 46.1%; lifetime reposts down 11.6%; view growth after attachment down 13.5%; lifetime views down 5.5%.
- Speed matters: absolute repost effect by quartile of attachment time (1-12 h, 12-23 h, 23-47 h, 47+ h) was -673, -285, -95, -2.
- No median time and no pre-note repost share given.
SAYABLE as: "A note that arrives within 12 hours cuts about 673 reposts; one that arrives after 47 hours cuts about 2. Over a post's whole life, a note cuts total views by only about 5.5% (PNAS 2025, X data from 2023)." Also: a note cuts reposts after it appears by 46%, but most of the damage is before.

### 1e. DDIA "A Deep Dive into X's Community Notes", July 2025 (via Fortune, summary of the report; report itself not fetched)
URL: https://fortune.com/2025/07/10/more-than-90-x-community-notes-are-never-published-stuck-limbo-study/
- Whole public dataset of 1.76 million notes, Jan 2021 to Mar 2025. "more than 90 percent never reach the public". English publication rate 9.5% in 2023 to 4.9% in early 2025.
- Average time from note submission to publication: over 100 days (2022) to 14 days (2025). Report: "far too slow for the reality of viral misinformation"; such content "spread within hours, not weeks."
- Caution: 14 days is an average of submit-to-publish for published notes, a different clock from post-to-note. A "7 h median / 38.5 h mean" attributed to DDIA on a wiki page was NOT verified; do not use.
SAYABLE as: "Over 90% of notes on X are never shown; in English the share fell to under 5% in early 2025 (DDIA, July 2025)."

### 1f. HEC Paris article citing Amariles, Renault, Clement (April 2026) (summary, article only)
URL: https://www.hec.edu/en/dare/strategy-leadership/fake-news-spreads-fast-platforms-must-act-faster
- Database of about 240,000 notes. "around 50% of retweets happen within the first six hours of a tweet's life, and the median note takes over 18 hours to be published." Paper itself not read; title in the source list is inconsistent.
SAYABLE as: "One 2026 study says half of all reposts happen in the first six hours, while the median note takes over 18 hours (summary)."

## 2. Professional fact-checkers

### 2a. Political fact-checking efforts are constrained by deficiencies in coverage, speed, and reach
URL: https://arxiv.org/html/2412.13280 (arXiv 2412.13280). Read first 100,000 of ~110,700 characters.
- 135 false or misleading narratives (ElectionRumors2022) from 1.8 million X posts, in 446 million election posts; 5 Sep to 1 Dec 2022 (US midterms).
- "median delay" of first fact-check: four days after the narrative first appeared (paper says "on average" once in the intro; speed section says median). Median 3 days after the narrative's peak.
- 47% of narratives (63 of 135) fact-checked at all.
- 79% of a narrative's misinformation posts were already shared when the median first fact-check was published.
- Fact-check posts: 12,510 against 1,045,373 others, about 1.2%.
SAYABLE as: "In a study of the 2022 US midterms on Twitter, the typical first fact-check came 4 days after a false story appeared, by which time 79% of the sharing was done, and fewer than half of the false stories were fact-checked at all." (US, 2022, text posts and narratives, not video)

### 2b. Older benchmark: Hoaxy/Shao et al. 2016 (summary, via search results only)
- About 13 hours between false report and debunk (Twitter, 2016, 71 sources over three months). Old; use only as history.

### 2c. India and AI video
- The News Minute, https://www.thenewsminute.com/news/deepfakes-voice-clones-and-ai-images-amplified-disinformation-on-india-pak-conflict : a deepfake of Pakistan's ISPR chief "spread rapidly on May 7 night, when Operation Sindoor began," and reached prime-time TV debates "before being debunked." Most viral item, an AI image of a bombed Rawalpindi stadium, "racked up over 9.6 million views but was not labelled or fact-checked on the platform." No debunk delay given. Names Alt News and BOOM as active.
- BOOM report May 2025 (search snippet only, summary): 68% of BOOM's fact-checks that month were about Operation Sindoor and the conflict; BOOM checks AI voice with detection tools.
- DFRLab, Carney deepfake, https://dfrlab.org/2025/06/19/deepfake-video-of-canadian-prime-minister-reaches-millions-on-tiktok-x/ : TikTok video (creator-labelled AI) gained traction a day before the 28 Apr 2025 Canadian election; Radio-Canada said "viewed over three million times in one day"; AFP Canada debunked it; notes "frequently added"; one X repost 4 May got about 2.4 million views WITH a "manipulated media" label; another a few hours later got about 1.5 million; a TikTok of 5 May got almost 4 million views. Debunk and removal dates not given.
SAYABLE as: "Even labelled, the Canadian PM deepfake kept getting millions of views: an X repost carrying a 'manipulated media' label still reached about 2.4 million (DFRLab, 2025)." And: "An AI image of a bombed stadium in the May 2025 India-Pakistan conflict got over 9.6 million views with no label and no fact-check on the platform (News Minute)."

NOT FOUND: any measured median turnaround for Alt News, BOOM, Factly, Newschecker, PIB Fact Check or DFRAC; any peer-reviewed median time-to-debunk for AI or deepfake video; Duke Reporters / IFCN / Reuters Institute delay numbers; CCDH, Washington Post or Poynter timing numbers (not located). Full Fact/Westminster (search summary, not read): AI content was 24% of Full Fact's fact checks, up from 8% a year earlier; no timing.

## 3. Spread speed
- Chuai et al. (above): repost half-life 5.75 h within the first 36 h (X, 2022-24, preprint).
- HEC citing Amariles et al. 2026: about 50% of reposts in the first 6 h (summary).
- Midterms paper: 79% of a narrative's posts already shared before the median fact-check (4 days).
- Bulletin of Atomic Scientists piece via search summary only: "median half-life of engagement on viral posts can be under two hours". Not read; do not use without a fetch.
- NOT FOUND: an Instagram-specific first-hour/first-day share; a "share of views gone before the note" figure (no source states one; 96.7% was a false lead).
SAYABLE as: "Roughly half of a post's reposts on X come in the first 6 hours, while the median note takes 15 to 18 hours." Compose from two sources with different samples; say "studies find", not "one study found."

## 4. Government clock, India
- TechCrunch, 10 Feb 2026, https://techcrunch.com/2026/02/10/india-orders-social-media-platforms-to-take-down-deepfakes-faster : amendments published 10 Feb, effective 20 Feb 2026; "a three-hour deadline for official takedown orders"; "a two-hour window for certain urgent user complaints"; labelling and traceability of synthetic audio and visual content; platforms must require users to disclose synthetic content and deploy tools to verify. A lawyer (AZB) describes it as removal "within three hours once they become aware of it."
- Search-result summaries (law-firm blogs, Bar & Bench, Mondaq; not each fetched): takedown on a government or court order cut from 36 h to 3 h; non-consensual intimate imagery 24 h to 2 h; urgent complaints 72 h to 36 h; routine grievances 15 days to 7. Sources disagree on sub-rule numbers and on whether the clock runs from an order or from awareness. Official gazette (G.S.R. 120(E)) not read. Platforms called the timelines "operationally unfeasible" (summary).
SAYABLE as: "Since 20 February 2026, India's rules give platforms three hours to remove unlawful content once a government or court orders it, down from 36 hours." Do not say "three hours for every deepfake."
NOT FOUND: any measurement of how fast platforms actually comply in practice; no MeitY or platform transparency data located. Say "I could not find a measurement."

## Suggested one-sentence spine (each part hedged as above)
"Studies of X find about half of reposts come in the first six hours, the median Community Note takes about 15 hours, and over 90% of notes never show; a US-midterm study found the first fact-check came a median 4 days later, after 79% of the sharing."
