Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content

Xucheng Yu; Jiangping Chen; Haohan Wang

cs.SIOct 19, 2026

Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content

Authors: Xucheng Yu, Jiangping Chen, Haohan Wang

Organizations: University of Illinois Urbana-Champaign Champaign, Illinois, USA

Abstract

Much of the manipulation in influencer content is not false but disproportionate, framing accurate facts as historic, urgent, or catastrophic. Existing approaches either verify factual claims or detect clickbait and propaganda in single-platform, single-language text; however, they overlook where such framing originates, because they treat a post and the replies beneath it as one unit. To address this gap, we measure rhetorical distortion with one instrument applied identically across six platforms and two languages, classifying creator posts and their replies separately. Our three-tier pipeline detects five literature-grounded dimensions (Significance Inflation, Anxiety Manufacturing, Novelty Claims, Loaded Language, and Temporal Distortion): bilingual lexical rules propose candidates, negative filters cancel known false positives, and an LLM verifies or relabels the candidates. This design keeps labels transparent and comparable across platforms, calls the LLM on only 14% of posts, and supports validation against human annotators and a second model. On 13,060 posts from 330 accounts on RSS/newsletters, YouTube, Bluesky, Reddit, Weibo, and X/Twitter, platform is far more strongly associated with account-level distortion than content category (partial $η^2=0.49$ vs. $0.09$ ), and X/Twitter scores highest (DI 7.74 vs. 0.78 for RSS). On X/Twitter, the top-5 replies under a post carry 3.5x the distortion of the tweet they answer (1.5x at equal length), while creator tweets alone are indistinguishable from YouTube and Bluesky; two human annotators and GPT-4o confirm this gap. Measured distortion on threaded platforms thus lives largely in the audience layer, which calls for measuring creator and reply text separately.

Figures & tables

Figure 1. Overview of the measurement pipeline. Lexical rules propose candidate dimensions and a confidence $c$ , negative filters cancel weak matches in known false-positive contexts, and the post is routed to one of three paths: the LLM (GPT-4o-mini) classifies it from scratch (Path A), verifies the rule labels (Path B), or is not called (Path C). Numbers on the paths are post counts in the corpus. Post labels are aggregated into per-account rates, the Distortion Index (DI), and the Consistency Score (CS). On X/Twitter, the same pipeline is run on the merged unit, on the tweet alone, and on its replies alone.

Platform	Collection	Classified unit	Accounts	Posts
			(corpus/anal.)	(corpus/anal.)
RSS	public feeds	post text	66 / 62	1,262 / 1,232
YouTube	API	title + transcript	31 / 30	601 / 600
Bluesky	API	post	54 / 53	2,680 / 2,630
Reddit	browser	title + body	57 / 56	2,670 / 2,669
Weibo	browser	post (ZH)	69 / 65	3,240 / 3,222

Table 1. Dataset overview. Corpus = posts with retrievable text; analysis set = accounts with CS

\geq

0.30 and

\geq

10 posts. On X/Twitter, 1,839 corpus posts (1,541 in the analysis set) have stored replies.

		Any	Micro	SI	AM	NC	LL	TD
Creator	F1	0.56	0.46	0.50	0.03	0.09	0.49	0.87
	Prec.	0.62	0.49	0.41	0.02	0.10	0.71	0.77
	Rec.	0.51	0.43	0.66	0.06	0.09	0.38	1.00
Replies	F1	0.71	0.62	–	0.62	–	0.66	–
	Prec.	0.89	0.78	–	0.50	–	0.92	–
	Rec.	0.60	0.51	–	0.83	–	0.51	–

Table 2. Agreement of the final pipeline with two human annotators (mean of the two). Creator text: 198 posts, population-weighted, X/Twitter scored on tweets. Replies: 60 X/Twitter reply units, unweighted; – = too few positives.

Platform	Accts	Posts	Chars	DI	SD	Zero	SI%	AM%	NC%	LL%	TD%
RSS/Newsletter	62	1,232	255	0.78	1.60	59.7%	2.49	0.08	0.24	1.00	0.08
YouTube	30	600	419	3.30	3.43	16.7%	5.17	2.33	0.67	6.83	1.50
Bluesky	53	2,630	191	1.46	1.46	28.3%	0.69	1.00	0.75	3.81	1.02
Reddit	56	2,669	201	0.78	0.74	48.2%	0.47	0.87	0.11	2.39	0.07
Weibo	65	3,222	113	0.97	1.23	43.1%	2.19	0.68	0.43	0.89	0.68
X/Twitter	43	2,150	521	7.74	5.30	9.3%	4.98	9.16	0.74	21.02	2.79

Table 3. Account-level results by platform (

N=309

accounts with CS

\geq

0.30 and

\geq

10 posts). DI is the unrounded mean of the five rates; Zero = share of accounts whose displayed (integer) DI is 0; Chars = mean length of the classified text. X/Twitter units are tweets with their top-5 replies (see § 4.4 for the split). Bold marks the highest value per column.

Panel of Figure 2 — Figure 2. Account-level DI by platform ( $N=309$ ). Boxes show quartiles, whiskers extend to 1.5 IQR, dots are outlying accounts, and diamonds mark platform means. X/Twitter’s lower quartile (4.8) lies above the upper quartile of every other platform except YouTube.

	Flag rate	OR (platform)	OR (+length)	$p$
RSS (ref.)	3.4%	1.00	1.00	–
YouTube	12.7%	4.10	2.82	$<$ 0.001
Bluesky	6.3%	1.91	2.43	$<$ 0.001
Reddit	3.6%	1.05	1.31	0.16
Weibo	4.1%	1.22	2.34	$<$ 0.001
X/Twitter	26.7%	10.31	6.20	$<$ 0.001

Table 4. Post-level logistic regression of “any dimension flagged” (

N=13{,}060

). Odds ratios relative to RSS, without and with log text length.

Panel of Figure 3 — Figure 3. Mean dimension rates by platform ( $N=309$ ). X/Twitter is LL-led, and YouTube has the highest SI.

	DI	SI	AM	NC	LL	TD
Creator tweets	2.17	1.81	1.63	0.48	3.96	2.95
Top-5 replies	7.65	3.44	8.64	0.60	24.97	0.61
$p$	$<$ 10 -6	0.014	$<$ 10 -5	0.95	$<$ 10 -6	0.058

Table 5. X/Twitter tweets vs. their top-5 replies, full pipeline, same posts (43 accounts). Wilcoxon signed-rank tests.

Unit	Chars	DI	LL%	Ratio	$p$
Creator tweet	259	2.17	3.96	–	–
First reply only	105	1.29	5.61	0.59	0.017
Replies, tweet length	224	3.24	12.63	1.50	0.003
Top-5 replies	493	7.50	24.88	3.46	$<10^{-6}$

Table 6. Reply-length control on X/Twitter posts with replies (43 accounts, GPT-4o-mini). Ratio and

p

(Wilcoxon) are relative to the creator tweet.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

	Strong	Weak
SI	changes everything; once in a generation; never happened before; endgame	most important; revolutionary; unprecedented; game changer
AM	you’ll be left behind; hope you’re prepared; point of no return	too late; panic; fall behind; bubble … burst
NC	nobody is talking about this; what the media won’t tell you	exclusive; hidden … truth; hard truths
LL	devastating; catastrophic; insane; insults; profanity	shocking; brutal; staggering; stupid; ALL-CAPS; !{2,}
TD	breaking … news; BREAKING ; this just in	just dropped; happening now; latest update

Appendix

Table 7. Representative Stage-1 patterns (regular expressions abbreviated).

	RSS	YT	BSky	Reddit	Weibo	X
Path C (no hit)	1,165	482	2,384	2,484	2,953	1,769
Path A (fresh LLM)	92	106	288	171	210	791
Path B (verify)	5	13	8	15	77	47
Rule-flagged	92	110	291	172	210	816
LLM cleared all	50	35	124	80	89	139
LLM changed	14	34	55	27	55	336

Appendix

Table 8. Pipeline paths per platform (posts) and LLM effect on rule-flagged posts.

	SI	AM	NC	LL	TD
Proposed by rules	160	73	57	1,190	143
Kept by LLM (%)	67.5	41.1	61.4	54.4	62.9
Added by LLM	158	254	25	53	31
Final labels	266	284	60	700	121
of which LLM-added (%)	59.4	89.4	41.7	7.6	25.6

Appendix

Table 9. Rule proposals versus final labels, analysis set (12,503 posts). Top: per dimension. Bottom: share of posts with any label under the v2 pipeline, the v3 rules alone, and the final v3 pipeline.

	RSS	YT	BSky	Reddit	Weibo	X
ai_ml	0.2 (11)	1.0 (3)	0.0 (2)	0.7 (6)	2.8 (5)	3.6 (4)
culture	1.1 (7)	1.0 (1)	2.4 (9)	0.6 (11)	0.5 (12)	–
education	1.0 (2)	4.3 (6)	–	0.1 (3)	–	–
finance	0.9 (12)	5.8 (4)	0.0 (1)	1.1 (12)	0.8 (18)	7.5 (8)
health	0.0 (3)	–	1.6 (2)	0.0 (1)	0.0 (1)	6.4 (1)
media	–	2.7 (3)	1.3 (20)	0.4 (2)	1.1 (11)	7.8 (10)

Appendix

Table 10. Mean DI by harmonized category and platform (accounts in parentheses;

N=309

). Chinese categories (e.g., finance_cn ) are merged with their English counterparts. Minor categories are merged as follows: crypto into finance; learning into education; lifestyle, rationalism, and sports into culture; marketing into startup; tech_media and tech_edu into tech.

CS	Min. posts	$N$	$\varepsilon^{2}$	Order	X/RSS
$\geq$ 0.0	1	330	0.36	X>Y>B>W>S>R	8.3
$\geq$ 0.0	10	320	0.36	X>Y>B>W>S>R	8.4
$\geq$ 0.0	20	297	0.36	X>Y>B>W>S>R	8.9
$\geq$ 0.3	1	317	0.34	X>Y>B>W>R>S	10.2
$\geq$ 0.3	10	309	0.33	X>Y>B>W>R>S	9.9
$\geq$ 0.3	20	287	0.32	X>Y>B>W>S>R	9.2

Appendix

Table 11. Sensitivity of the platform effect to inclusion criteria. Order is by mean DI (X = X/Twitter, Y = YouTube, B = Bluesky, W = Weibo, R = Reddit, S = RSS).

Panel of Figure 4 — Figure 4. Lorenz curve of flagged posts across the 309 accounts. The bottom 30% of accounts contribute 0.5% of flagged posts; the top 10% contribute 50.5%.

Account	Platform	Category	Leading dim.	DI
GavinNewsom	X/Twitter	politics	LL (70%)	21
tedcruz	X/Twitter	politics	LL (56%)	18
mtgreenee	X/Twitter	politics	LL (58%)	18
jimcramer	X/Twitter	finance	LL (48%)	16
VP	X/Twitter	politics	LL (50%)	16
PeterSchiff	X/Twitter	finance	LL (36%)	14

Appendix

Table 12. Accounts with displayed DI

\geq

12 (

N=309

), with the leading dimension.

Panel of Figure 5 — Figure 5. X/Twitter creator tweets vs. their replies by category (full pipeline, accounts in parentheses). Replies exceed tweets in every category.

Entity	RSS	YT	BSky	X	X tweets	X replies
New York Times	0	–	1	11	6.8	6.4
BBC News	0	–	–	13	0.4	12.0
Reuters	–	–	1	7	0.8	4.8
The Guardian	1	–	0	6	0.0	6.0
Eric Topol	–	–	1	4 †	–	–
Andrej Karpathy	0	0	–	2 †	–	–

Appendix

Table 13. Same entity, different platform (displayed integer DI; X tweets and replies: unrounded full-pipeline DI). † CS

<0.30

Role	Chinese	Gloss
SI marker	史上 / 历史性 / 里程碑	in history / historic / milestone
AM marker	危机 / 崩溃 / 恐慌	crisis / collapse / panic
NC marker	独家 / 揭秘 / 内幕	exclusive / revealed / inside story
LL marker	震惊 / 炸裂 / 塌房	shocking / explosive / scandal collapse
TD marker	突发 / 刚刚	breaking / just now
Filter: retrospective	历史上的今天	on this day in history

Appendix

Table 14. Chinese terms cited in the main text, with English glosses.

Social Media Analysis

Social media analysis studies online posts, media, users, and interactions to understand communication, communities, and public behavior.

See all 221 papers

Content Fuzzing for Escaping Information Cocoons on Digital Social Media
Apr 7, 2026Yifeng He, Ziye Tang, Hao ChenSocial Media Analysis Adversarial Text Attacks
Department of Computer Science, University of California, Davis · Department of Communication, University of California, Davis · University of Hong Kong
RCMN: Understanding Misleadingness in Influential Public Discourse
Aug 27, 2026Peiling Yi
School of Computer Science and Mathematics Faculty of Engineering, Computing and the Environment Kingston University London, United Kingdom
GRID: Modeling Health Influencer Discourse as Argument Networks for Credibility Assessment
Oct 19, 2026Nishanth Sridhar Nakshatri, Eylon Caplan, Dan GoldwasserHealth Misinformation Detection Scientific Claim Verification
Purdue University

Apr 7, 2026J/K move · Enter open · S save

Content Fuzzing for Escaping Information Cocoons on Digital Social Media

Yifeng He, Ziye Tang, Hao Chen

Department of Computer Science, University of California, Davis · Department of Communication, University of California, Davis · University of Hong Kong

Social Media Analysis Adversarial Text Attacks

Open paper AskPDF arXiv

Information cocoons on social media limit users' exposure to posts with diverse viewpoints. Modern platforms use stance detection as an important signal in recommendation and ranking pipelines, which can route posts primarily to like-minded audiences and reduce cross-cutting exposure. This restricts the reach of dissenting opinions and hinders constructive discourse. We take the creator's perspective and investigate how content can be revised to reach beyond existing affinity clusters. We present ContentFuzz, a confidence-guided fuzzing framework that rewrites posts while preserving their human-interpreted intent and induces different machine-inferred stance labels. ContentFuzz aims to route posts beyond their original cocoons. Our method guides a large language model (LLM) to generate meaning-preserving rewrites using confidence feedback from stance detection models. Evaluated on four representative stance detection models across three datasets in two languages, ContentFuzz effectively changes machine-classified stance labels, while maintaining semantic integrity with respect to the original content.

Abstract

Social Media Analysis

Explore similar work

Content Fuzzing for Escaping Information Cocoons on Digital Social Media

RCMN: Understanding Misleadingness in Influential Public Discourse

GRID: Modeling Health Influencer Discourse as Argument Networks for Credibility Assessment