Divergent Emotional Patterns in Disinformation on Social Media?

An Analysis of Tweets and TikToks About the DANA in Valencia

an Arcos

, Paolo Rosso

1,2

and Ram

on Salaverr

ıa

PRHLT Research Center, Universitat Polit

ecnica de Val

encia, Valencia, Spain

ValgrAI Valencian Graduate School and Research Network of Artiﬁcial Intelligence, Spain

School of Communication, Universidad de Navarra, Pamplona, Spain

Keywords:

Disinformation, DANA, Extreme Weather Events, Social Media, Journalism, Computational Linguistics,

Emotional Patterns.

Abstract:

This study investigates the dissemination of disinformation on social media platforms during the DANA event

(DANA is a Spanish acronym for Depresi

on Aislada en Niveles Altos, translating to high-altitude isolated

depression) that resulted in extremely heavy rainfall and devastating ﬂoods in Valencia, Spain, on October 29,

2024. We created a novel dataset of 650 TikTok and X posts, which was manually annotated to differentiate

between disinformation and trustworthy content. Additionally, a Few-Shot annotation approach with GPT-4o

achieved substantial agreement (Cohen’s kappa of 0.684) with manual labels. Emotion analysis revealed that

disinformation on X is mainly associated with increased sadness and fear, while on TikTok, it correlates with

higher levels of anger and disgust. Linguistic analysis using the LIWC dictionary showed that trustworthy

content utilizes more articulate and factual language, whereas disinformation employs negations, perceptual

words, and personal anecdotes to appear credible. Audio analysis of TikTok posts highlighted distinct patterns:

trustworthy audios featured brighter tones and robotic or monotone narration, promoting clarity and credibility,

while disinformation audios leveraged tonal variation, emotional depth, and manipulative musical elements

to amplify engagement. In detection models, SVM+TF-IDF achieved the highest F1-Score, excelling with

limited data. Incorporating audio features into roberta-large-bne improved both Accuracy and F1-Score,

surpassing its text-only counterpart and SVM in Accuracy. GPT-4o Few-Shot also performed well, showcasing

the potential of large language models for automated disinformation detection. These ﬁndings demonstrate the

importance of leveraging both textual and audio features for improved disinformation detection on multimodal

platforms like TikTok.

1 INTRODUCTION

In the digital era, social media platforms such as X

(formerly Twitter) and TikTok revolutionized com-

munication and information dissemination. These

platforms facilitate instant sharing and broad interac-

tions, enabling users to engage with content in real-

time and reach vast audiences with unprecedented

speed. While this democratization of information ac-

cess offers numerous beneﬁts, it also presents signiﬁ-

cant challenges, particularly concerning the spread of

false or misleading information.

Disinformation, deﬁned by European Commis-

sion (2019) as “veriﬁably false or misleading infor-

mation created, presented, and disseminated for ﬁ-

nancial gain or intentional deception of the public”,

https://ﬁrstdraftnews.org/latest/coe infodisorder/

Figure 1: Relationship among Disinformation, Misinforma-

tion, and Malinformation (Wardle and Derakhshan, 2017).

has become a pervasive issue on social media. Un-

like misinformation, which involves the unintentional

Arcos, I., Rosso, P. and Salaverría, R.

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia.

DOI: 10.5220/0013392800003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 925-936

ISBN: 978-989-758-737-5; ISSN: 2184-433X

925

sharing of incorrect information without malicious

intent, disinformation is strategically crafted to ma-

nipulate public perception, inﬂuence opinions, and

achieve speciﬁc agendas. Additionally, malinforma-

tion refers to the dissemination of genuine informa-

tion with the intent to harm or manipulate, further

complicating the information ecosystem. These three

categories are illustrated in Figure 1.

The design of social media platforms is charac-

terized by algorithms that prioritize engagement, the

formation of echo chambers that reinforce existing

beliefs, and the ease of content sharing. These fea-

tures facilitate the rapid and widespread propagation

of disinformation. This environment allows malicious

actors, including those whose aim is to spread disin-

formation to increase their number of followers, as

well as automated accounts and coordinated misinfor-

mation campaigns, to effectively amplify false narra-

tives. Furthermore, the viral nature of social media

content means that disinformation can quickly reach

and inﬂuence large segments of the population, often

outpacing fact-checking and regulatory efforts aimed

at curbing its spread.

Understanding the mechanisms behind the dis-

semination of disinformation is crucial for develop-

ing effective countermeasures. Emotional and lin-

guistic patterns play a signiﬁcant role in how disinfor-

mation resonates with audiences and spreads across

platforms. Analyzing these patterns can provide in-

sights into the narrative employed by those who create

and distribute false information, as well as the emo-

tional triggers that make such content more likely to

be shared and believed (McLoughlin et al., 2024).

This study focuses on the spread of disinformation

during the DANA event in Valencia, Spain, on Octo-

ber 29, 2024, which caused severe ﬂooding, resulting

in 224 deaths and three missing persons. By examin-

ing emotional and linguistic patterns in 650 annotated

posts from TikTok and X, we aim to elucidate the

strategies used in disseminating false narratives and

their impact on public sentiment. The research em-

ploys both traditional machine learning models and

advanced language models, such as transformers and

Large Language Models (LLMs), to detect disinfor-

mation, highlighting the strengths and limitations of

each approach.

The remainder of this article is organized as fol-

lows: Section 2 reviews related literature on disinfor-

mation detection in social media. Section 3 details

the data collection and annotation processes. Sec-

tion 4 presents the emotion analysis conducted on

the dataset. Section 5 explores the linguistic patterns

using the LIWC dictionary. Section 6 discusses the

analysis of audio variables in TikTok posts. Section

7 describes the models used for disinformation detec-

tion and their performance. Finally, Section 8 con-

cludes the study and outlines directions for future re-

search.

2 RELATED WORK

The spread of misinformation and disinformation on

social media has garnered signiﬁcant research atten-

tion due to its profound impact on public opinion,

democratic processes, and public health (Wang et al.,

2022). The viral nature of social media platforms fa-

cilitates the swift dissemination of both true and false

news (Vosoughi et al., 2018).

A seminal study by Vosoughi et al. (2018) ana-

lyzed the diffusion patterns of true and false news

on X from 2006 to 2017, revealing that false news

spreads signiﬁcantly farther, faster, deeper, and more

broadly than true news, particularly in political con-

texts. This enhanced virality is attributed to the nov-

elty and emotional content of false stories, which

evoke fear, disgust, and surprise, thereby increasing

user engagement. Recent research has found that out-

rage is a key emotion that facilitates the spread of

misinformation particularly on X (McLoughlin et al.,

2024). Complementing these ﬁndings, Allcott and

Gentzkow (2017) examined the role of fake news in

the 2016 US presidential election, highlighting its

substantial reach and inﬂuence, especially for stories

favoring speciﬁc political candidates. The study un-

derscored that users are more inclined to believe fake

news that aligns with their preexisting beliefs, particu-

larly within ideologically segregated social networks.

Addressing the challenges posed by misinforma-

tion, researchers have proposed various interventions.

Pennycook et al. (2019) introduced the concept of

the implied truth effect, demonstrating that attach-

ing warnings to a subset of fake news headlines can

inadvertently increase the perceived accuracy of un-

tagged false headlines, highlighting the complexity of

designing effective misinformation mitigation strate-

gies.

The COVID-19 pandemic further highlighted the

dangers of misinformation, giving rise to an “info-

demic.” Islam et al. (2020) conducted a global so-

cial media analysis to assess the impact of false infor-

mation related to COVID-19, ﬁnding that misinfor-

mation about treatments, transmission, and mortality

rates fueled public panic and hindered effective public

health responses.

In the realm of fake news detection, Shu et al.

(2017) provided a comprehensive review of text min-

ing techniques employed to identify false news on so-

EAA 2025 - Special Session on Emotions and Affective Agents

926

cial media, proposing a conceptual framework that

leverages both content-based and context-based fea-

tures to enhance detection accuracy.

Multimodal approaches utilize both textual and vi-

sual data to enhance fake news identiﬁcation. For in-

stance, Rashid et al. (2024) introduced MuAFaNI, a

framework that combines text and image representa-

tions using RoBERTa and ResNet-50 models to as-

sess news authenticity. Similarly, Qu et al. (2024)

proposed QMFND, a quantum multimodal fusion-

based model that integrates image and textual fea-

tures through a quantum convolutional neural net-

work, thereby improving detection accuracy. Fur-

ther advancing the ﬁeld, Lee et al. (2024) explored

the emotional effects of multimodal disinformation,

demonstrating that video-based misinformation sig-

niﬁcantly increases anxiety and misperceptions com-

pared to text-only content, highlighting the enhanced

persuasive power of multimodal disinformation and

its implications for public trust and decision-making.

Other notable contributions include Kumar and

Taylor (2024), which developed a multimodal fake

news detection model incorporating textual and visual

features alongside feature engineering techniques to

capture the behavior of fake news propagators. Shan

et al. (2024) employed similarity inference and ad-

versarial networks to fuse textual and visual features,

thereby enhancing detection accuracy on social me-

dia platforms. Additionally, Pan et al. (2024) and Lu-

vembe et al. (2024) introduced models that leverage

graph convolutional networks and complementary at-

tention mechanisms to align and fuse multimodal fea-

tures, achieving signiﬁcant accuracy improvements

on benchmark datasets.

The emergence of deepfakes has further com-

plicated the disinformation landscape. Vaccari and

Chadwick (2020) investigated the impact of synthetic

political videos, revealing that deepfakes contribute to

uncertainty and reduce trust in news sources, thereby

undermining the integrity of public discourse.

While multimodal approaches represent a signiﬁ-

cant advancement by leveraging both textual and vi-

sual data, the application of multimodal techniques

speciﬁcally to TikTok videos remains underexplored.

This study aims to address this gap by focusing on

the multimodal characteristics of TikTok content, ini-

tially leveraging transcriptions of video audio, with

the intention to incorporate video and audio features

in future work.

3 DATASETS

3.1 Data Extraction

The datasets used in this study were extracted from

TikTok and X using crawlers provided by Apify

For TikTok, the clockworks/tiktok-scraper

was em-

ployed, while for X, the fastcrawler/Tweet-Fast-

Scraper

was utilized. Additionally, the datasets will

be made public for research purposes to facilitate fur-

ther studies and enhance transparency in disinforma-

tion research.

The data extraction process was conducted using

keywords related to the DANA phenomenon and as-

sociated disinformation narratives. The keywords,

summarized in Table 1, were carefully selected to

capture content that potentially spread disinformation

or expressed emotional reactions during the DANA

phenomenon.

Table 1: Keywords Used for Data Extraction and Their

Translations.

Keyword (Spanish) Translation (English)

DANA conspiraci

on DANA conspiracy

DANA fallecidos ocultaci

on DANA deceased concealment

DANA enga

no DANA deception

DANA manipulaci

on DANA manipulation

DANA mentiras DANA lies

DANA ataque clim

atico DANA climate attack

DANA manipular la verdad DANA manipulate the truth

DANA desinformaci

on DANA disinformation

DANA falsedades DANA falsehoods

DANA ocultaci

on DANA concealment

DANA Rub

en Gisbert DANA Rub

en Gisbert

DANA Alvise P

erez DANA Alvise P

erez

DANA Iker Jim

enez DANA Iker Jim

enez

DANA Vito Quiles DANA Vito Quiles

DANA Cruz Roja falsa ayuda DANA Cruz Roja fake aid

DANA Bonaire cementerio DANA Bonaire cemetery

DANA ayuda rechazada DANA rejected aid

DANA pron

ostico incorrecto DANA incorrect forecast

DANA radar sin funcionar DANA radar not working

DANA provocada presas DANA caused by dams

The extraction focused on ﬁltering posts that

matched these keywords, aiming to collect content

that potentially spread disinformation or expressed

emotional reactions during the DANA phenomenon.

For TikTok videos, the transcription of spoken

content was extracted using the Whisper-medium

model

, which is designed for efﬁcient and accu-

rate speech-to-text transcription. Additionally, text

https://console.apify.com/

https://apify.com/clockworks/tiktok-scraper

https://apify.com/fastcrawler/tweet-fast-scraper/api/

javascript

https://huggingface.co/openai/whisper-medium

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia

927

appearing in the videos was extracted using Pad-

dleOCR

, a state-of-the-art Optical Character Recog-

nition (OCR) tool. These methods ensured the com-

prehensive collection of both spoken and visual tex-

tual content for analysis.

3.2 Bias Assessment

To assess potential biases introduced by keyword-

based data collection, we applied the Weirdness Index

(WI) (Poletto et al., 2020), which measures how much

certain words in a dataset deviate from their expected

frequencies in a reference corpus.

As a reference, we collected social media content

using broad, non-targeted keywords: “DANA Valen-

cia”. This corpus captures general discussions on the

DANA phenomenon, including neutral and informa-

tive posts, rather than those explicitly linked to disin-

formation. By comparing our dataset against this ref-

erence, we evaluated whether our keyword selection

introduced thematic biases.

Table 2 presents the WI statistics for both datasets.

A high WI indicates that certain words are dispropor-

tionately represented compared to the reference cor-

pus.

Table 2: Weirdness Index (WI) results comparing datasets.

Metric Tweets TikToks

WI Global (Mean) 4.86 1.76

WI Global (Median) 2.37 0.79

WI Standard Deviation 45.63 3.23

Words with WI > 2 55.64% 24.90%

The tweets dataset shows a signiﬁcantly higher

WI, indicating that keyword ﬁltering has ampliﬁed

speciﬁc narratives. Given the constraints in data avail-

ability and manual annotation time, our objective was

to capture a broad range of disinformation-related

themes rather than general discussions on the DANA

phenomenon. The keyword selection was designed to

encompass various disinformation topics, including

conspiracy theories, exaggerated impacts, and mis-

leading emergency response claims.

In contrast, the TikTok dataset exhibits a lower

WI, suggesting a more diverse linguistic distribution.

Although the same keyword selection strategy was

used, TikTok’s transcription-based content structure

results in a dataset that, while still containing disinfor-

mation, is less narrowly focused compared to tweets.

https://github.com/PaddlePaddle/PaddleOCR

3.3 Data Annotation

The collected TikToks and tweets were further anno-

tated to create a labeled subset for analysis. Each post

was examined to determine its relation to the DANA

phenomenon and the spread of disinformation. Fact-

checking was conducted using the reputable websites

Maldita

, Newtral

, and EFE Veriﬁca

The annotation process involved assigning one of

two labels to each post:

• Label 0. Assigned to posts that discuss the DANA

phenomenon without spreading disinformation or

falsehoods. This category also includes posts that

actively report or debunk disinformation .

• Label 1. Assigned to posts that directly spread

disinformation or falsehoods about the DANA

phenomenon.

Table 3 shows the distribution of labels for Tik-

Toks and tweets. Efforts were made to ensure a bal-

anced representation between the two classes (0 and

1) to facilitate unbiased analysis.

Table 3: Label distribution for TikToks and Tweets.

Label TikToks Tweets Total

0 (Trustworthy) 131 177 308

1 (Disinformation) 137 205 342

Total 268 382 650

To better illustrate the difference between disin-

formation and trustworthy content, Table 4 presents

one example of each category. These examples high-

light the type of content analyzed and the criteria used

for labeling.

The word clouds in Figure 2 highlight that the

vocabulary across both social networks is strikingly

similar. On X, terms related to trustworthy often in-

clude bulo (hoax), tragedia (tragedy), falso (false),

and desinformaci

on (disinformation), alongside men-

tions of public ﬁgures such as Iker Jim

enez

, Vito

Quiles

, Rub

en Gisbert

, and Alvise P

erez

. These

terms often appear in posts criticizing or exposing

falsehoods. In disinformation, words such as embalse

(reservoir), presa (dam), HAARP

, S

anchez (Span-

https://maldita.es

https://newtral.es

https://veriﬁca.efe.com/

Iker Jim

enez is a journalist.

Vito Quiles is a pseudojournalist.

Rub

en Gisbert is a pseudojournalist.

Alvise P

erez is the leader of Se Acab

o La Fiesta (The

Party is Over).

Alleged climate attacks generated from HAARP

(High-Frequency Active Auroral Research Program).

EAA 2025 - Special Session on Emotions and Affective Agents

928

ish Prime Minister Pedro S

anchez), UME

, demolido

(demolished), cementerio (graveyard), and AEMET

are frequently observed.

On TikTok, the focus of terms associated with

trustworthy shifts towards community and natu-

ral disaster responses, with words such as inun-

daciones (ﬂoods), voluntario (volunteer), fen

omeno

(phenomenon), tormenta (storm), ayudar (help), and

zona (area) standing out. Disinformation in this con-

text is more closely linked to terms like Cruz Roja

(Red Cross), ropa (clothing), verdad (truth), gobierno

(government), radar (radar), and cad

averes (corpses).

Table 4: Examples of Disinformation and Trustworthy

Tweets (and their translation into English).

Trustworthy Tweet (Label 0)

Miles de voluntarios se movilizan desde Les Arts hacia

las zonas afectadas. Mat

ıas Prats informa que no hay

ıctimas mortales en el t

unel.

(Thousands of volunteers are mobilizing from Les Arts

towards the affected areas. Mat

ıas Prats reports that there

are no mortal victims in the tunnel.)

Disinformation Tweet (Label 1)

Me acaban de comunicar que ya han sacado 86 falleci-

dos del parking de Bonaire. El aparcamiento del centro

comercial tiene 5700 plazas.

(I have just been informed that 86 deceased have already

been removed from the Bonaire parking lot. The shop-

ping center’s parking lot has 5700 spaces.)

3.4 Few-Shot Annotation with GPT-4o

A Few-Shot annotation approach was tested using

GPT-4o, a variant of GPT-4 optimized for enhanced

efﬁciency and performance. While all annotations in

this study were conducted manually, GPT-4o was ap-

plied to a subset of tweets to explore the potential of

large language models for automated annotation.

Although the primary focus of our analysis was

on two main categories (Trustworthy information

and Disinformation), we discovered that reﬁning the

prompt to include four detailed categories and then

collapsing them into two yielded higher agreement

values between GPT-4o and manual annotations. This

strategy allowed for a more nuanced understanding of

the tweets during the initial annotation process, which

improved the precision and reliability of the ﬁnal bi-

nary labels.

UME (Unidad Militar de Emergencias) is the Spanish

Military Emergencies Unit.

AEMET (Agencia Estatal de Meteorolog

ıa) is the

Spanish national meteorological agency.

(a) Word Clouds for X: Disinformation and Trustworthy

Posts

(b) Word Clouds for TikTok: Disinformation and Trustwor-

thy Posts

Figure 2: Word clouds from X and TikTok.

3.4.1 Prompt Design for GPT-4o

To classify the tweets into predeﬁned categories, the

following clean and structured prompt was provided

to GPT-4o. For readers interested in the full details of

the prompt structure, please refer to Appendix A.

3.4.2 Results and Evaluation

Table 5 shows the annotation agreement matrix for the

four predeﬁned categories, where GPT-4o achieved

a Cohen’s kappa of 0.684, indicating a substantial

agreement with manual annotations. This highlights

the effectiveness of GPT-4o in capturing nuanced dif-

ferences between the categories.

Table 5: Annotation Agreement for Four Categories.

0 1 2 3

0 15 1 0 2

1 1 13 1 1

2 3 4 28 5

3 4 2 3 36

When the categories were reduced to two groups,

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia

929

Trustworthy (0) and Disinformation (1), the Cohen’s

kappa further improved to 0.695 (Table 6).

Table 6: Annotation Agreement for Two Categories.

0 1

0 66 8

1 9 36

These results demonstrate the power of GPT-4o,

achieving substantial agreement with manual anno-

tations. Such capabilities can signiﬁcantly reduce

the time and effort required for annotating large-scale

datasets, particularly in the context of disinformation

detection during events like DANA.

4 EMOTION ANALYSIS IN

TWEETS AND TIKTOKS

This study analyzed the emotions present in tweets

and TikToks using the twitter-xlm-roberta-emotion-

es model

. The model identiﬁes six emotions: sad-

ness, joy, anger, surprise, disgust, fear, and others.

These categories align with the basic emotions pro-

posed by Paul Ekman, which are considered univer-

sal across cultures and rooted in human evolution Ek-

man (1992). Non-parametric tests (Mann–Whitney

U) were performed to determine emotional patterns

in disinformation across both X and TikTok.

4.1 Emotional Patterns in X

Table 7 presents the results for emotional analysis in

tweets. Notably, tweets with disinformation showed

signiﬁcantly higher sadness (p = 0.000061). This of-

ten corresponds to claims exaggerating the number of

casualties, e.g., “El enga

no sobre el n

umero de muer-

tos es cruel, indignante e inhumano” (“The deception

about the number of deaths is cruel, outrageous, and

inhumane”).

Disinformation also showed higher fear levels (p

= 0.031), linked to claims about catastrophic events,

such as dam failures, e.g., “Lo que pasa es que el em-

balse de Tous est

a al 97%. Si abren las compuertas,

viene una riada” (“What’s happening is that the Tous

reservoir is at 97%. If they open the gates, a ﬂood is

coming”).

https://huggingface.co/daveni/

twitter-xlm-roberta-emotion-es

Table 7: Emotional patterns in X.

Emotion Mean (0) Std (0) Mean (1) Std (1) p-value

Sadness 0.077 0.131 0.176 0.235 0.000061

Fear 0.012 0.007 0.014 0.008 0.031030

Others 0.622 0.296 0.522 0.309 0.002906

4.2 Emotional Patterns in TikToks

Table 8 shows emotional patterns in TikToks. Un-

like X, sadness is more prevalent in TikToks with-

out disinformation (p = 0.000380), often focusing on

the tragedy itself, e.g., “Lamentablemente las grandes

inundaciones que se est

an provocando en Valencia

Espa

na han dejado hasta la fecha m

as de 2500 de-

saparecidos” (“Unfortunately, the severe ﬂoods hap-

pening in Valencia, Spain, have left over 2,500 people

missing so far”).

Disinformation is associated with higher levels of

anger (p = 0.000003) and disgust (p = 0.000003), re-

ﬂecting frustration or exaggeration, e.g., “Las ayudas

de los 6.000 euros para los afectados por la DANA

las tienes que devolver en tres meses” (“The C6,000

aid for those affected by DANA must be repaid within

three months”), or “Luego digan que no hay muertos

en las calles, se

nor. Que no hay muertos en las calles.

Acaban de tapar uno aqu

ı con una toalla. Despu

es de

cuatro d

ıas recogiendo y amontonando cosas entre la

mierda. ¿Eh? Y no hab

ıa muertos. ¿Eh? No hab

ıan

muertos” (“Then they say there are no dead bodies in

the streets, sir. That there are no dead bodies in the

streets. They just covered one here with a towel. Af-

ter four days of picking up and piling things among

the ﬁlth. Huh? And there were no dead bodies. Huh?

There were no dead bodies”).

Table 8: Emotional patterns in TikToks.

Emotion Mean (0) Std (0) Mean (1) Std (1) p-value

Sadness 0.294 0.196 0.218 0.191 0.000380

Anger 0.165 0.143 0.262 0.187 0.000003

Disgust 0.027 0.018 0.040 0.025 0.000003

5 LINGUISTIC ANALYSIS USING

LIWC

In addition to emotional analysis, linguistic pat-

terns were examined using the LIWC dictionary Pen-

nebaker et al. (1999)

, which groups words into di-

mensions such as Linguistic Dimensions, Psycholog-

ical Processes, Personal Concerns, and Spoken Cate-

gories.

https://www.liwc.app/

EAA 2025 - Special Session on Emotions and Affective Agents

930

5.1 Linguistic Patterns in TikToks

Table 9 summarizes the linguistic patterns in Tik-

Toks. When disinformation is not present, the dis-

course tends to be more articulate, with a higher use

of prepositions and terms related to health, often in-

forming about infection risks or providing warnings

e.g., “El riesgo de infecci

on en las zonas afectadas por

la DANA sigue siendo muy alto debido a las aguas

estancadas y contaminadas...” (“The risk of infection

in areas affected by the DANA remains very high

due to stagnant and contaminated water...”). Such

videos also include more references to time and spe-

ciﬁc events e.g., “¡Atenci

on! Una nueva DANA traer

humedad y precipitaciones a partir del martes por la

noche” (“Warning! A new DANA will bring humidity

and rainfall starting Tuesday night”).

Conversely, videos spreading disinformation often

use more negation and profanities (Maldec), reﬂect-

ing frustration or skepticism. These videos also em-

ploy more perceptual words (Percept) to suggest ﬁrst-

hand observations or unveriﬁable claims e.g., “En las

noticias no sale toda la verdad, est

an desaguando el

centro comercial de aqu

ı dicen que hay m

as de 800 me

ha dicho un guardia civil...” (“The news isn’t telling

the whole truth; they’re draining the mall here, and

they say there are over 800 [bodies], a Civil Guard

told me...”).

Table 9: Linguistic Patterns in TikToks Using LIWC.

Category Mean (0) Std (0) Mean (1) Std (1) p-value

Prepos 10.68 3.66 8.61 3.66 0.00001

Health 0.25 0.42 0.18 0.36 0.0424

Time 2.56 2.74 2.01 1.64 0.0261

Negation 1.37 1.28 1.93 2.58 0.0105

Maldec 0.11 0.30 0.20 0.42 0.0130

Percept 2.02 1.36 2.56 1.94 0.0336

5.2 Linguistic Patterns in X

Table 10 highlights the linguistic patterns observed in

tweets. Similar to TikToks, tweets without disinfor-

mation also employ a more articulate discourse, using

more prepositions (p = 0.000001) and negation (p =

0.036797).

In tweets spreading disinformation, there is a no-

table increase in the use of ﬁrst-person singular pro-

nouns (I). These tweets often cite unveriﬁable fam-

ily connections as sources of information (Family),

such as “Repito, tengo un familiar directo militar en

el parking de Bonaire y est

an sacando muertos sin

parar” (“I repeat, I have a direct family member in

the military at the Bonaire parking lot, and they are

removing dead bodies non-stop”).

The results indicate clear differences in linguistic

Table 10: Linguistic Patterns in X Using LIWC.

Category Mean (0) Std (0) Mean (1) Std (1) p-value

Prepos 14.08 5.73 11.06 6.56 0.000001

Negation 0.86 1.67 1.32 2.21 0.0368

I 0.31 1.14 0.54 1.42 0.0403

Family 0.02 0.21 0.13 0.79 0.0320

patterns between disinformation and trustworthy con-

tent on both platforms. TikToks and tweets without

disinformation tend to use more articulate language

with a higher focus on factual information, while dis-

information often incorporates perceptual language,

negations, and personal anecdotes to enhance credi-

bility

6 AUDIO FEATURES IN TIKTOKS

Audio is a key component of TikTok videos, con-

tributing signiﬁcantly to their emotional and persua-

sive impact. In this study, we extracted a wide range

of audio features to analyze the role of sound in differ-

entiating disinformation and trustworthy content. The

extracted features include:

• Zero Crossing Rate (ZCR). It measures the rate

at which the signal changes its sign, reﬂecting the

noisiness or percussive nature of the sound.

• Root Mean Square Energy (RMS). It captures

the loudness or energy of the audio signal, often

linked to the intensity of the content.

• Spectral Centroid. It indicates the center of grav-

ity of the spectrum, associated with the perceived

brightness of the audio.

• Spectral Rolloff. It represents the frequency be-

low which a certain percentage (e.g., 85%) of the

spectral energy is concentrated, often used to dis-

tinguish harmonic and percussive elements.

• Chroma Features. They measure the energy dis-

tribution among the 12 pitch classes, representing

harmonic content.

• Spectral Bandwidth. It quantiﬁes the range of

frequencies present in the signal, reﬂecting its

tonal richness.

• Spectral Flatness. It describes the noisiness of

the signal by measuring how ﬂat the power spec-

trum is.

• Mel-Frequency Cepstral Coefﬁcients

(MFCCs). They capture the timbral aspects

of the audio, essential for distinguishing speech

and music characteristics.

• Harmonics-to-Noise Ratio (HNR). It indicates

voice quality and clarity by comparing harmonic

and noise components.

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia

931

• Tempo. It estimates the speed of the audio, which

can evoke urgency or calmness in the listener.

• Onset Strength. It measures the intensity of

sound onsets, reﬂecting rhythmic and dynamic

variations.

• Pitch. It refers to the perceived fundamental fre-

quency of the signal, often linked to emotional

tones such as excitement or seriousness.

• Spectral Contrast. It represents differences in

amplitude between peaks and valleys in the power

spectrum, providing insights into the dynamic

range of the audio.

• Tonnetz Features. They capture tonal harmonic

relationships, offering information about the mu-

sical nature of the audio.

For each feature, both the mean and standard

deviation (std) were calculated to capture the central

tendency and variability of the audio properties.

The analysis revealed signiﬁcant differences in au-

dio features between trustworthy (0) and disinforma-

tion (1) groups, as shown in Table 11, reﬂecting dis-

tinct acoustic and tonal patterns:

• Zero Crossing Rate (ZCR, ZCR STD). Trust-

worthy audios exhibited higher values, indicating

more frequent transitions across the zero axis in

the waveform. This reﬂects more energetic and

articulated speech, characteristic of clear and di-

rect communication.

• Spectral Features:

– Spectral Centroid. Higher in trustworthy con-

tent, representing brighter tones with more em-

phasis on higher frequencies.

– Spectral Bandwidth. Wider in trustworthy con-

tent, indicating a broader range of frequencies

around the centroid.

– Spectral Rolloff. Higher in trustworthy posts,

reﬂecting stronger energy in higher frequen-

cies, which enhances tonal clarity.

• Spectral Flatness. Through the visualization of

videos with high spectral ﬂatness values, we ob-

served that they frequently featured narrators de-

livering information in a neutral, monotone, or

robotic style, often associated with objective news

reporting about tragedies or serious events. This

explains why spectral ﬂatness was higher in trust-

worthy audios. In contrast, disinformation audios

displayed lower spectral ﬂatness, indicating more

tonal content and greater variation. These varia-

tions are likely aimed at dramatizing events, evok-

ing emotional responses, and manipulating the lis-

tener’s perception.

Table 11: Signiﬁcant Audio feature comparison between

trustworthy (0) and disinformation (1) groups.

Feature Mean

(0)

Std (0) Mean

(1)

Std (1) p-

value

zcr 0.094 0.031 0.085 0.029 0.019

zcr std 0.072 0.030 0.060 0.029 0.006

spectral

cen-

troid

1902.41 465.65 1682.81 440.07 0.0002

spectral

cen-

troid

std

935.15 313.20 798.22 337.14 0.0049

spectral

rolloff

3692.10 996.91 3244.93 918.65 0.0010

spectral

rolloff

std

1764.06 470.15 1570.43 583.68 0.019

spectral

band-

width

1984.54 339.32 1799.34 348.43 0.0001

spectral

ﬂatness

0.022 0.029 0.017 0.028 0.007

spectral

ﬂatness

std

0.048 0.063 0.044 0.069 0.030

onset

strength

1.80 0.36 1.66 0.40 0.006

mfcc2

mean

111.10 25.04 120.40 28.08 0.022

mfcc5

mean

3.16 10.85 -0.44 11.02 0.010

mfcc6

std

16.56 3.77 15.54 3.97 0.021

contrast7

mean

49.03 2.60 50.13 3.67 0.025

contrast6

std

7.58 1.55 7.15 1.47 0.015

tonnetz4

mean

-0.001 0.051 0.022 0.074 0.003

• Onset Strength. Higher onset strength was ob-

served in trustworthy audios, reﬂecting strong and

clear acoustic events such as sharp changes in am-

plitude or the beginning of words and sounds. In

contrast, lower onset strength in disinformation

audios suggests smoother transitions and less dy-

namic emphasis, which might be used to create a

more conversational or emotionally manipulative

tone.

• MFCCs (Mel-Frequency Cepstral Coefﬁ-

cients):

– MFCC2. Higher in disinformation, captur-

ing lower frequency tonal characteristics, as-

sociated with deeper and emotionally charged

tones.

– MFCC5 and MFCC6. Higher in trustwor-

thy content, indicating richer and clearer tonal

EAA 2025 - Special Session on Emotions and Affective Agents

932

qualities in the mid-frequency range, aligned

with neutral and structured speech.

• Spectral Contrast (Contrast7). Higher in dis-

information, particularly in the highest frequency

bands, helping dramatize content and attract at-

tention.

• Tonnetz (Tonnetz4). Videos with high Tonnetz

values often included eerie or mysterious music.

Observational analysis of such videos revealed

their use in disinformation to create a suspenseful

or manipulative tone, amplifying the emotional

impact.

Trustworthy audios are characterized by brighter

tones, dynamic events, and monotone or robotic nar-

ration styles, optimized for clarity and credibility.

Disinformation audios, on the other hand, emphasize

tonal variation, emotional depth, and harmonic com-

plexity, often accompanied by manipulative musical

elements designed to captivate and inﬂuence the lis-

tener.

7 MODELS

To train the models for disinformation detection,

tweets and transcriptions of TikToks were combined

into a uniﬁed dataset. To minimize bias, hashtags

and user mentions were removed from the text to pre-

vent the models from relying on these features for

predictions. Given the limited size of the dataset,

we employed three main approaches: a traditional

machine learning model, ﬁne-tuning of transformer-

based models, and a GPT-4o Few-Shot learning ap-

proach.

The models used were roberta-base-bne

roberta-large-bne

, xlm-roberta-large-twitter

longformer-base-4096-bne-es

, fake-news-

detection-spanish

, SVM+TF-IDF, and GPT-4o

Few-Shot. We also experimented with ﬁne-tuning

roberta-large-bne by incorporating audio features

from TikTok. These features were normalized and

combined with text embeddings from the model’s

[CLS] token. The fused representations were passed

https://huggingface.co/PlanTL-GOB-ES/

roberta-base-bne

https://huggingface.co/PlanTL-GOB-ES/

roberta-large-bne

https://huggingface.co/sdadas/

xlm-roberta-large-twitter

https://huggingface.co/PlanTL-GOB-ES/

longformer-base-4096-bne-es

https://huggingface.co/Narrativaai/

fake-news-detection-spanish

through a classiﬁer, leveraging both textual and audio

information to enhance disinformation detection in a

multimodal context.

To ensure robust evaluation, 5-fold stratiﬁed

cross-validation was employed for the traditional and

transformer-based models, maintaining the class dis-

tribution in each fold. This approach provided a com-

prehensive assessment of model performance, min-

imizing the risk of overﬁtting and ensuring reliable

generalization to unseen data. For the GPT-4o Few-

Shot approach, a single evaluation was conducted.

The performance of each model was evaluated us-

ing the following metrics: Accuracy, F1-Score, Preci-

sion, and Recall. The results are summarized in Table

12, where the best-performing model for each metric

is highlighted in bold.

The SVM+TF-IDF model demonstrated the high-

est effectiveness in terms of F1-Score and Recall,

achieving scores of 0.7949 and 0.8831, respectively.

This superiority can be attributed to the Support

Vector Machine’s ability to efﬁciently handle high-

dimensional feature spaces even with small datasets,

thereby avoiding overﬁtting and providing robust gen-

eralization. The simplicity and efﬁciency of tradi-

tional models like SVM make them particularly suit-

able when data resources are limited.

On the other hand, transformer-based models such

as roberta-base-bne and roberta-large-bne exhibited

competitive performance, albeit slightly inferior to

SVM in this scenario of limited data. However, it

is important to note that these models possess sig-

niﬁcant potential for scalability and improved perfor-

mance when larger training datasets are available.

Notably, by incorporating audio features into

roberta-large-bne, we observed a signiﬁcant im-

provement in both Accuracy and F1-Score, achiev-

ing 0.7646 in accuracy (an increase of 1.34 percent-

age points over the standard roberta-large-bne) and

0.7664 in F1-Score (an increase of 1.44 percentage

points). Furthermore, this multimodal approach out-

performed the SVM+TF-IDF model in terms of Ac-

curacy, demonstrating the potential of combining tex-

tual and audio features to enhance performance in dis-

information detection tasks.

Finally, the Few-Shot approach using GPT-4o

achieved the highest accuracy of 0.7834, demonstrat-

ing the effectiveness of large-scale language mod-

els when provided with well-designed and speciﬁc

prompts. This result highlights the potential of LLMs

for automated annotation and classiﬁcation tasks,

even without domain-speciﬁc training.

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia

933

Table 12: Performance Metrics of Different Models for Disinformation Detection.

Model Accuracy F1-Score Precision Recall

roberta-base-bne 0.7497 ± 0.0506 0.7565 ± 0.0575 0.7668 ± 0.0313 0.7486 ± 0.0839

roberta-large-bne 0.7512 ± 0.0177 0.7520 ± 0.0270 0.7885 ± 0.0248 0.7220 ± 0.0543

roberta-large-bne + audio features 0.7646 ± 0.0292 0.7664 ± 0.0272 0.7726 ± 0.0432 0.7634 ± 0.0484

xlm-roberta-large-twitter 0.7274 ± 0.0175 0.7358 ± 0.0192 0.7505 ± 0.0167 0.7229 ± 0.0352

longformer-base-4096-bne-es 0.7266 ± 0.0473 0.7035 ± 0.0681 0.8069 ± 0.0361 0.6285 ± 0.0943

fake-news-detection-spanish 0.7298 ± 0.0454 0.7377 ± 0.0416 0.7541 ± 0.0486 0.7222 ± 0.0370

SVM+TF-IDF 0.7604 ± 0.0173 0.7949 ± 0.0133 0.7227 ± 0.0165 0.8831 ± 0.0087

GPT-4o Few-Shot 0.7834 0.7867 0.8150 0.7602

8 CONCLUSIONS AND FUTURE

WORK

8.1 Conclusions

This study analyzed the spread of disinformation on

TikTok and X during the DANA event in Valencia,

Spain, which was marked by severe ﬂooding. A

dataset of 650 posts was manually annotated to dis-

tinguish between disinformation and trustworthy con-

tent. Additionally, a separate experiment was con-

ducted using GPT-4o in a Few-Shot setup to evaluate

its potential as a tool for automated annotation. GPT-

4o demonstrated substantial agreement with the man-

ual labels (Cohen’s kappa of 0.684 for four categories

and 0.695 for two categories).

Emotion analysis revealed that disinformation on

X is primarily associated with heightened sadness and

fear, while on TikTok, it correlates with increased

anger and disgust. The identiﬁcation of this diver-

gence in predominant emotions across different so-

cial networks is a distinctive result of our study. First,

our work reinforces the conclusion, already reached

by many previous studies, that appealing to emotions

is a recurring strategy in deliberately misleading mes-

sages Osmundsen et al. (2021); McLoughlin et al.

(2024). However, thanks to the comparative approach

of our study—where we analyzed the prevailing emo-

tional traits in posts on X and TikTok—we have iden-

tiﬁed a phenomenon that has been scarcely investi-

gated until now: the fact that, on each platform, dis-

information content operates under a different emo-

tional paradigm. This conclusion underscores the

need to further understand the impact of emotional di-

mensions on credulity toward disinformation content,

as well as the predisposition to share it Pennycook and

Rand (2021). Linguistic analysis using the LIWC dic-

tionary indicated that trustworthy content tends to use

more articulate and factual language, whereas disin-

formation employs negations, perceptual words, and

personal anecdotes to enhance credibility.

The analysis of TikTok audio features revealed

distinct patterns: trustworthy audio showed brighter

tones, stronger onset strength, and higher spectral

ﬂatness, reﬂecting clarity and neutrality, while dis-

information audio used tonal variations and deeper

MFCCs to evoke emotions and manipulate percep-

tions. These ﬁndings highlight the value of audio

analysis in multimodal disinformation detection.

In terms of detection models, the traditional

SVM+TF-IDF model outperformed transformer-

based models and the GPT-4o Few-Shot approach

in F1-Score and Recall, beneﬁting from the lim-

ited dataset size. However, transformer-based mod-

els showed competitive performance and potential for

scalability with larger datasets. Notably, by incorpo-

rating audio features into roberta-large-bne, we ob-

served an improvement in both Accuracy and F1-

Score compared to its text-only counterpart, as well as

a higher accuracy than the SVM+TF-IDF model. This

demonstrates the value of multimodal approaches in

disinformation detection tasks. The GPT-4o Few-

Shot model demonstrated high accuracy, highlighting

the effectiveness of LLMs in automated classiﬁcation

tasks without extensive domain-speciﬁc training.

8.2 Future Work

Future research can expand on this study by:

• Increasing Dataset Size. Incorporate more

tweets and TikToks, as well as additional social

media platforms.

• Multimodal Analysis. Include images and

videos to capture a broader range of disinforma-

tion strategies.

• Emotional Divergences. Examine how distinct

emotional triggers inﬂuence disinformation dis-

semination on different social media platforms.

• Advanced Models. Explore sophisticated trans-

former architectures and ensemble methods for

enhanced performance.

EAA 2025 - Special Session on Emotions and Affective Agents

934

• Cross-Linguistic Studies. Analyze disinforma-

tion patterns across different languages and re-

gions.

• Network Analysis. Investigate user behaviors

and network structures that facilitate disinforma-

tion spread.

• Reﬁning Few-Shot Techniques. Improve prompt

designs and methodologies for large language

models to reduce reliance on manual annotations.

• Intervention Strategies. Develop effective

counter-disinformation measures based on identi-

ﬁed emotional and linguistic patterns.

• Intelligent Agents. Design and implement intel-

ligent agents capable of real-time detection and

mitigation of disinformation. These agents could

combine machine learning techniques with rule-

based systems to analyze linguistic and emotional

cues, detect malicious content, and take auto-

mated countermeasures. For instance, intelligent

agents could:

– Flag potential disinformation for review or fur-

ther analysis.

– Provide users with context or veriﬁed informa-

tion to counter false claims.

– Interact with social media algorithms to limit

the spread of harmful content while promoting

veriﬁed, accurate information.

Addressing these areas will enhance the detec-

tion and mitigation of disinformation on social me-

dia, thereby strengthening the integrity of information

ecosystems during critical events.

ACKNOWLEDGEMENTS

This work was done in the framework of the Iberian

Digital Media Observatory (IBERIFIER Plus),

co-funded by the EC under the Call DIGITAL-

2023-DEPLOY-04 (Grant 101158511), and of

the Malicious Actors Proﬁling and Detection in

Online Social Networks Through Artiﬁcial In-

telligence (MARTINI) research project, funded

by MCIN/AEI/10.13039/501100011033 and by

NextGenerationEU/PRTR (Grant PCI2022-135008-

2).

REFERENCES

Allcott, H. and Gentzkow, M. (2017). Social media and

fake news in the 2016 election. Journal of Economic

Perspectives, 31(2):211–236.

Ekman, P. (1992). An argument for basic emotions. Cogni-

tion and Emotion, 6(3-4):169–200.

European Commission (2019). Tackling online disinforma-

tion.

Islam, M., Sarkar, T., Khan, S., Mostofa Kamal, A., Hasan,

S., Kabir, A., Yeasmin, D., Islam, M., Amin Chowd-

hury, K., Anwar, K., Chughtai, A., and Seale, H.

(2020). Covid-19-related infodemic and its impact

on public health: A global social media analysis.

American Journal of Tropical Medicine and Hygiene,

103(4):1621–1629.

Kumar, A. and Taylor, J. W. (2024). Feature importance

in the age of explainable ai: Case study of detecting

fake news & misinformation via a multi-modal frame-

work. European Journal of Operational Research,

317(2):401–413.

Lee, J., Hameleers, M., and Shin, S. Y. (2024). The

emotional effects of multimodal disinformation: How

multimodality, issue relevance, and anxiety affect mis-

perceptions about the ﬂu vaccine. New Media & Soci-

ety, 26(12):6838–6860.

Luvembe, A. M., Li, W., Li, S., Liu, F., and Wu, X. (2024).

Caf-odnn: Complementary attention fusion with opti-

mized deep neural network for multimodal fake news

detection. Information Processing & Management,

61(3):103653.

McLoughlin, K. L., Brady, W. J., Goolsbee, A., Kaiser,

B., Klonick, K., and Crockett, M. J. (2024). Misin-

formation exploits outrage to spread online. Science,

386(6725):991–996.

Osmundsen, M., Bor, A., Vahlstrup, P. B., Bechmann, A.,

and Petersen, M. B. (2021). Partisan polarization is

the primary psychological motivation behind political

fake news sharing on twitter. American Political Sci-

ence Review, 115(3):999–1015.

Pan, Z., Mao, Y., Xiong, L., Pang, T., and Ping, P. (2024).

Mfae: Multimodal fusion and alignment for entity-

level disinformation detection. Pattern Recognition

Letters, 184:59–65.

Pennebaker, J., Francis, M., and Booth, R. (1999). Linguis-

tic inquiry and word count (liwc).

Pennycook, G., Bear, A., and Collins, E. (2019). The im-

plied truth effect: Attaching warnings to a subset of

fake news headlines increases perceived accuracy of

headlines without warnings. Management Science.

Pennycook, G. and Rand, D. G. (2021). The psychology of

fake news. Trends in cognitive sciences, 25(5):388–

402.

Poletto, F., Basile, V., Sanguinetti, M., Bosco, C., and Patti,

V. (2020). Resources and benchmark corpora for hate

speech detection: a systematic review. Language Re-

sources and Evaluation, 55:477 – 523.

Qu, Z., Meng, Y., Muhammad, G., and Tiwari, P. (2024).

Qmfnd: A quantum multimodal fusion-based fake

news detection model for social media. Information

Fusion, 104:102172.

Rashid, J., Kim, J., and Masood, A. (2024). Unraveling the

tangle of disinformation: A multimodal approach for

fake news identiﬁcation on social media. In Compan-

ion Proceedings of the ACM Web Conference 2024,

Divergent Emotional Patterns in Disinformation on Social Media? An Analysis of Tweets and TikToks About the DANA in Valencia

935

pages 1849–1853, New York, NY, USA. Association

for Computing Machinery.

Shan, F., Sun, H., and Wang, M. (2024). Multimodal social

media fake news detection based on similarity infer-

ence and adversarial networks. Computers, Materials

& Continua, 79(1).

Shu, K., Sliva, A., Wang, S., Tang, J., and Liu, H. (2017).

Fake news detection on social media: A data mining

perspective. In Proceedings of the IEEE International

Conference on Data Mining Workshops (ICDMW),

pages 799–804.

Vaccari, C. and Chadwick, A. (2020). Deepfakes and disin-

formation: Exploring the impact of synthetic political

video on deception, uncertainty, and trust in news. So-

cial Media + Society, 6(1):2056305120903408.

Vosoughi, S., Roy, D., and Aral, S. (2018). The spread of

true and false news online. Science, 359(6380):1146–

1151.

Wang, S., Su, F., Ye, L., and Jing, Y. (2022). Disinfor-

mation: A bibliometric review. International Jour-

nal of Environmental Research and Public Health,

19(24):16849.

Wardle, C. and Derakhshan, H. (2017). Information dis-

order: Toward an interdisciplinary framework for re-

search and policymaking, volume 27. Council of Eu-

rope, Strasbourg.

A DETAILED FEW-SHOT

ANNOTATION PROMPT FOR

GPT-4o

The following prompt was originally provided in

Spanish to align with the language of the data but for

the sake of comprehension for non-Spanish speakers,

it has been translated into English in this article:

Context: The Depresi

on Aislada en Niveles Al-

tos (DANA) that affected Valencia and other areas of

Spain in October 2024 caused severe ﬂooding and the

spread of disinformation. Common types of disinfor-

mation included:

• False claims about dam openings or breaks,

such as Forata.

• Conspiracy theories about climate manipula-

tion (e.g., HAARP, Morocco, geoengineering).

• Rumors of hidden deaths or transport in re-

frigerated trucks (e.g., parking at Bonaire).

• False accusations of rejected aid or misman-

agement of donations, such as delays or destruc-

tion of aid by the Red Cross.

• Exaggerated numbers of victims, ﬂoods, or

negligence.

• Alarmist content about basic services (e.g., wa-

ter supply, healthcare disruptions).

• Unfounded accusations against political ﬁgures

or governments.

• Claims of Valencia’s meteorological radar be-

ing non-functional (when it was operational).

• Rumors of censorship or media manipulation

by public ﬁgures (e.g., YouTube or Iker Jim

enez).

Labeling Guidelines:

• 0 (Not related to DANA): Tweets unrelated to

DANA, even indirectly.

Example: “Just had the best paella in Valencia.

Highly recommend!”

• 1 (Related to DANA, but not disinformation):

Tweets discussing real events or consequences of

DANA without spreading disinformation.

Example: “Thousands of volunteers are helping

victims of DANA in Valencia.”

• 2 (Disinformation mentioned, but criticized):

Tweets mentioning disinformation about DANA

to criticize or report it, including posts debunking

or ridiculing falsehoods.

Example: “Don’t believe the fake news about

hidden deaths in Bonaire; it has been debunked

by ofﬁcial sources.”

• 3 (Disinformation): Tweets actively spreading

false information, rumors, or conspiracy theories

about DANA, or unfounded accusations against

political ﬁgures.

Example: “They’re hiding the truth! Over 800

bodies were found in the Bonaire parking lot, and

no one is reporting it.”

Tweet: [TWEET HERE]

Return only the number:

EAA 2025 - Special Session on Emotions and Affective Agents

936