=Paper=
{{Paper
|id=Vol-3782/paper6
|storemode=property
|title=Analysis of Climate Change Misleading Information in TikTok
|pdfUrl=https://ceur-ws.org/Vol-3782/paper6short.pdf
|volume=Vol-3782
|authors=Clara Baltasar,Sergio D’Antonio Maceiras,Alejandro Martín,David Camacho
|dblpUrl=https://dblp.org/rec/conf/codai2/BaltasarDC24
}}
==Analysis of Climate Change Misleading Information in TikTok==
https://ceur-ws.org/Vol-3782/paper6short.pdf
Analysis of Climate Change Misleading Information in
TikTok
Clara Baltasar1 , Sergio D’Antonio Maceiras1 , Alejandro Martín1 and David Camacho1
1
Department of Computer System Engineering, Universidad Politécnica de Madrid, Calle de Alan Turing, 28031 Madrid, Spain
Abstract
In today’s digital landscape, social media platforms have become important areas for disseminating information,
ranging from legitimate discourse to misinformation, especially on critical topics such as Climate Change. This
study employs claim detection and clustering techniques to analyze misleading information within an initial
dataset of videos. Initially, the study identified 5,352 videos out of a total of 8,151 that warranted further investiga-
tion. Utilizing clustering methods, it was discovered that the prevalence of misinformation was surprisingly lower
than anticipated. Most of the clusters showcases content that promotes sustainability and raises environmental
awareness, strengthened by corroborated information of fact-checking agency EFE Verifica. Conversely, there
are two clusters that focuses on videos propagating misinformation, conspiracy theories, active discussion and
debate, highlighting the necessity of consuming media with caution. Looking ahead, combating misinformation
necessitates the enhancement of digital literacy and the cultivation of critical thinking skills. This research aims
to leverage technology and verified information from credible organizations to identify, analyze, and mitigate the
influence of misleading content on social media, thus better understanding its dynamics and reducing its adverse
impacts.
Keywords
Climate change, TikTok videos, Misleading information, Environmental issues
1. Introduction
The spread of climate change misinformation on TikTok has become a serious concern for both
environmental scientists and social media analysts. Understanding the dynamics and impact of this
phenomenon requires in-depth analysis of the most viral content on the platform Viral content on
TikTok often has several key characteristics: visually appealing, emotional and easy to understand. In
the context of climate change misinformation, these elements are often used to attract audiences and
spread false narratives. For example, many viral videos contain dramatic images of natural disasters
or melting ice caps combined with misleading or inaccurate explanations. These videos often contain
sensational and apocalyptic predictions, which can evoke strong emotional responses and encourage
users to share the content widely.
A perfect example is videos falsely claiming that climate change is a fraud committed by governments
or corporations for financial gain. These clips often feature conspiracy theory imagery, which is
particularly appealing to viewers who are already skeptical of mainstream science. Additionally, such
content creators often pretend to be tipsters, exploiting the feeling of inside information to increase
their authority and attract more views.
TikTok’s viral misinformation patterns are influenced by the platform’s algorithm, which prioritizes
content that generates high levels of engagement. Videos with lots of likes, comments and shares
appear more frequently on the For You page, which exponentially increases their reach. This creates a
feedback loop in which misinformation continuously reaches new users, thereby increasing its spread.
Collaboration features also play a key role in the spread of misinformation. These features allow
users to directly interact with existing videos by adding their own comments or creating reaction videos.
Proceedings of the 1st Workshop on COuntering Disinformation with Artificial Intelligence (CODAI), co-located with the 27th
European Conference on Artificial Intelligence (ECAI), pages 54–61, October 20, 2024, Santiago de Compostela, Spain
$ cm.baltasar@upm.es (C. Baltasar); sergio.dantonio@upm.es (S. D. Maceiras); alejandro.martin@upm.es (A. Martín);
david.camacho@upm.es (D. Camacho)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
54
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
Not only does this increase the visibility of the original content, but it also creates a sense of community
and brainstorming around misinformation, which further contributes to its virality.
The widespread dissemination of climate change misinformation on TikTok has significant implica-
tions for public perception. Studies such as [1] have shown that repeated exposure to misinformation
can lead to the formation of false beliefs and increased skepticism towards scientific consensus. On
TikTok, where the user base skews younger, the impact is particularly concerning. Young users are
still developing their understanding of complex issues like climate change and their perceptions can be
heavily influenced by the content they consume on social media. Furthermore, viral misinformation
can undermine efforts to promote accurate scientific information.
In this research, we undertake an analysis of misinformation around climate change in TikTok. We
filtered 5,352 from a total of 8,151 videos related to conversations around climate. Then, we extracted
information from the videos, such as keywords and the transcription. Then we applied clustering
algorithms to identify different subtopics and conversations, which allowed us to identify groups
promoting negationist theories about climate change.
The rest of this article is organised as follows. Section II presents an analysis of the state-of-the-art
literature, Section III describes the methodology, Section IV the results, Section V analyses the presence
of hoaxes and finally, Section VI presents a series of conclusions.
2. State of the art
In the digital age, social media platforms have become a prime battleground for the dissemination of
information and the spread of ideas, both legitimate and otherwise. One such arena is the discussion
surrounding climate change, where the impact of social media, AI and algorithmic systems has become a
growing concern [2]. The proliferation of misinformation and disinformation on social media platforms
has garnered significant attention from researchers, especially concerning critical issues like climate
change. This section reviews some relevant contributions in this area, identifying the main trends and
findings in the existing literature.
Several studies as [3], agree that fake news is an old concept; it has existed and will exist as long
as publishers continue to use misleading information to promote their interests and this has been
happening since before the Internet even existed. Nowadays, as [4] points out, misleading information
can come in various forms, such as fake new, disinformation or misinformation, which are easily spread
through social media.
Many studies have investigated the spread of misinformation on social media. [5] examined the
mechanisms through which misinformation is disseminated on platforms like Twitter. They presented
a language model who detects fake news spreaders on Twitter. In the same way, other studies as [6]
examines the stability and evolution of network structure and discussion topics among a group of
prominent climate change deniers. The findings reveal that while the climate change denial network
remains stable in terms of size and core group composition, sub-groups continuously emerge and
dissolve.
[7] explored the dynamics of misleading information on environmental issues on TikTok, a relatively
new and rapidly growing social media platform. This study found that TikTok’s short-form video content
and highly engaging user interface create a fertile ground for the viral spread of both accurate and
inaccurate information. This study, also emphasized the importance of science education in addressing
misleading information.
Also, the purpose of the study [8], was to describe content related to climate change on TikTok.
Their findings indicate that climate change is being represented on TikTok as a legitimate and anxiety
provoking issue. Although only a few videos included in their sample are disinformation, these garnered
millions of views. Therefore, they concluded that the presence of credible professionals is essential on
platforms like TikTok to increase the chances that messages are as complete as time allows, while also
being scientifically sound.
55
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
3. Architecture development
This section describes the methodology employed to extract, filter and analyze all extracted information
from the social network used in this research: TikTok.
3.1. Data collection
The initial phase of our research involved the systematic collection of data using TikTok’s research API
designed for developers [9].
We meticulously extracted metadata from videos associated with a selection of environmentally
relevant hashtags. These hashtags included #climatechange, #ecofriendly, #sustainability and #ecotok.
Upon securing the initial dataset, we conducted a comprehensive analysis to identify the most fre-
quently used hashtags within these videos. This analysis led to the discovery of additional pertinent
hashtags such as #zerowaste, #naturetok, #globalwarming, #climatecrisis, #savetheplanet, #ecology,
#plasticfree, #sustainable, #savetheworld, #recycle, #recycling, #upcycling, #saveourplanet, #upcycle,
#bethechange, #environment, #climateaction, #climateemergency, #climate, #plasticpollution, #plas-
ticwaste, #savetheocean, #saveouroceans and #eco. This iterative process was repeated meticulously
until we had compiled an exhaustive list of the most prevalent hashtags in our dataset, which are those
shown above.
Subsequent to the environmental hashtag analysis, we turned our focus to the investigation of
misinformation related to climate change. We identified and collected metadata associated with hash-
tags that propagate misinformation, such as #climatelies, #climatehoaxx, #climatehoaks, #climatelie,
#climatehoax, #globalwarmingisfake, #globalwarminghoax, #globalwarmingisahoax, #carbonkleptoma-
nia, #globalcooling, #climatechangehoax, #noclimateemergency, #climatescam, #weathermanipulation,
#stopglobalwarming and #globalwarmingsucks. And just like before, we repeated this iterative process
until we had compiled an exhaustive list of the most prevalent hashtags that propagate misinformation
in our dataset, which are those shown above.
Through this rigorous process, we accumulated a total of 8,151 video metadata entries, covering the
period from January 2020 to June 2024.
3.2. Data characterization in TikTok
Once all videos were retrieved, the second step was to extract the audio channel, converting from .mp4
to format .mp3, and then we used the Whisper model [10] which is a pre-trained model for automatic
speech recognition and speech translation, to get the transcription from the audio of the videos. We
obtained a total of 6,998 text transcriptions and we started to work with that.
The next step was to analyze the transcriptions. First, we divided the transcriptions into sentences
and two pre-trained models were loaded: the SentenceTransformer model [11], for creating dense
vector representations of sentences, and the KeyBERT model [12], for extracting keywords from the
transcriptions.
Our aim was to do clustering to all transcriptions and to obtain the most relevant keywords of each
cluster. For this, to find the optimal number of clusters for K-Means, the elbow method and silhouette
scores were used, and it was found that this number was ten clusters.
Then, we applied K-Means clustering to the embeddings and when we analyzed the most relevant
keywords of each cluster we obtained some keywords that didn’t make sense on the issue of climate
change, like “foryourpage”, “fyp” or “for watching”, so we established them as stopwords but the results
obtained were not much better. After trying several things we realized that many of the videos only had
songs or phrases that were not very relevant as sound, so we decided to leave the transcriptions aside.
The second option was to use the “video_description” that we obtained when we extracted metadata
from videos using TikTok’s research API designed for developers [9], which owns the description of
the video, and apply SentenceTransformer [11] and KeyBERT model [12] to this descriptions.
As we mentioned before, we accumulated a total of 8,151 video metadata entries, covering the period
from January 2020 to June 2024, so the data collection we used in the end was larger than expected.
56
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
To find the optimal number of clusters for K-Means, the elbow method and silhouette scores were
used, and it was found that this number was seven clusters. Then, we applied K-Means clustering to
the descriptions of the videos and we obtained seven clusters well-defined, which are shown in the
section 4.
4. Results
The image below, Fig. 1, represents the clustering of descriptions from 8,151 TikTok videos. We employed
K-means clustering and Principal Component Analysis (PCA) to visualize and understand the thematic
grouping of these transcriptions. The scatter plot shows the distribution of seven clusters, differentiated
by colors. This section provides a detailed analysis of these clusters, discussing the most relevant
keywords, potential misinformation and aggregate statistics for likes, views and comments.
Figure 1: Clustering of descriptions from TikTok videos.
The clustering algorithm identified seven distinct clusters among the videos descriptions. Each cluster
represents a group of videos with similar content based on their descriptions. The cluster analysis is
presented below:
• Group 0. This cluster, represented in the scatter plot by the color purple, contains 1,260 videos.
This cluster is characterized by prominent keywords like "sewing", "upcycling", "thrift", "thriftflip"
and "fashion". This suggests a focus on sewing techniques, clothing upcycling and sustainable
fashion practices.
• Group 1. It is identified in the scatter plot by the color dark blue, comprises 1,300 videos. This
cluster is characterized by content related to TikTok, duets, life hacks and fashion. Common
57
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
keywords include "duet with", "duet", "lifehacks" and "tiktok". This indicates a focus on social
interaction and entertainment.
• Group 2. It is represented in the scatter plot by the color blue, contains 1,665 videos. This cluster
is distinguished by its focus on sustainability, environment and ecological awareness. Keywords
include "sustainability", "environment", "eco", "nature" and "ecofriendly". This reflects a strong
interest in sustainable practices and environmental protection.
• Group 3. Identified in the scatter plot by the color turquoise, comprises 1,243 videos. This
cluster is characterized by discussions on weather manipulation, chemtrails and climate-related
conspiracy theories. Keywords include "weathermanipulation", "chemtrails", "geoengineering",
"weathermodification" and "conspiracytiktok".
• Group 4. This cluster, represented in the scatter plot by the color green, contains 1,139 videos.
This cluster focuses on climate change, global warming, climate action and its controversies.
Common keywords include “climatechange”, “globalwarming”, “climatecrisis”, “climateaction”
and “gretathunberg”. Also included are terms that deny or question climate change (“climatelies”,
“globalwarmingisfake” and “climatehoaxx”). The activist “gretathunberg” is mentioned, suggesting
discussions about her influence.
• Group 5. This cluster, identified in the scatter plot by the color light green, comprises 636 videos.
This cluster is characterized by its focus on climate action, saving the planet and environmen-
tal awareness. Keywords include “stopglobalwarming”, “savetheworld”, “savetheplanet” and
“saveouroceans”.
• Group 6. This last cluster, represented in the scatter plot by the color yellow, contains 908 videos.
This cluster is characterized by content related to recycling, waste reduction and sustainable
practices. Keywords include “recycle“, “recycling”, “zero waste”, “reuse” and “crafts”.
4.1. Keyword Frequency, videos count, likes, views and comments Analysis
To provide the big picture of the results, Table 1 lists the most frequent keywords for each cluster, The
total number of videos in each cluster and the total likes, views and comments for videos in each cluster.
This data highlights the distribution of videos across the seven thematic clusters and provides insight
into the engagement levels of videos in different thematic groups.
An analysis of the keywords and content themes across clusters reveals distinct focal areas and
potential areas of misinformation. Clusters like Cluster 0 and Cluster 6 emphasize practical advice and
advocacy for environmental issues, showcasing keywords related to sewing, upcycling, recycling and
sustainable living. For instance, Cluster 0 includes terms like “sewing”, “upcycling” and “thrift” while
Cluster 6 is characterized by “recycle”, “recycling” and “plasticfree”. These clusters tend to present more
straightforward and educational content, making them less likely to contain misinformation.
In contrast, Cluster 3 stands out due to the prevalence of terms such as "weathermanipulation",
"chemtrails" and "geoengineering", pointing to content related to conspiracy theories. This cluster is
more likely to mix factual information with misleading claims, which can negatively influence public
perception.
Also Cluster 4, despite containing terms as "climatechange", "globalwarming", "climatecrisis" and
"climateaction", it also contains terms as "climatelies", "globalwarmingisfake" and "climatehoaxx; which
means it focuses on climate change, global warming, climate action and its controversies.
The focus on conspiracy theories contrasts with the more practical and educational nature of the
content in Clusters 0 and 6.
When considering the total likes, views and comments of these clusters, significant differences
emerge. Cluster 5, for instance, which includes keywords like "stopglobalwarming" and "savetheworld"
has the highest number of likes (3,053,581). This suggests that videos in this cluster receive higher
acceptance and engagement from users, possibly due to the urgency and global appeal of climate change
messaging.
Despite having fewer likes, Cluster 6 has the highest number of views (31,976,184). This indicates that
videos in this cluster may be viewed more frequently, possibly due to a growing interest in recycling and
58
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
Cluster Keywords and Frequency Videos Count Likes Views Comments
sewing (37)
upcycling (24)
0 thrift (24) 1,260 1,160,419 11,666,224 15,321
thriftflip (20)
fashion (19)
duet with (59)
duet (47)
1 bethechange (28) 1,300 839,619 6,303,510 40,536
tiktok (25)
lifehacks inspiration (19)
sustainability (63)
environment (40)
2 sustainable (40) 1,665 1,778,118 18,963,898 39,140
ecology (37)
eco (27)
weathermanipulation (338)
chemtrails (136)
3 weathermodification (69) 1,243 681,355 3,707,237 57,980
weathermanipulation lexky (60)
coveringthesun weathermanipulation (51)
climatechange (238)
globalwarming (116)
4 global warming (67) 1,139 1,511,484 10,797,769 48,182
climate change (47)
climatecrisis (36)
stopglobalwarming (162)
savetheworld (87)
5 savetheplanet (32) 636 3,053,581 17,783,375 52,676
globalwarming (29)
saveouroceans (18)
recycle (126)
recycling (118)
6 recycled (27) 908 1,621,028 31,976,184 17,701
plasticfree (18)
waste (15)
Table 1
Clustering analysis results
zero-waste topics. The higher view count suggests that the audience for recycling content is substantial,
even if individual engagement through likes is lower.
Cluster 1, by the other hand, characterized by keywords such as "duet with", "duet" and "bethechange"
highlights a significant trend in social media engagement. With 1,300 videos, Cluster 1 focuses on
collaborative content, challenges and inspirational messages.
Cluster 3, which includes conspiracy-related keywords, has the highest number of comments (57,980),
and Cluster 4 which also mentions "gretathunberg", has the third higher number of comments (48,182),
indicating active discussion and debate. This could be due to the controversial nature of the topics,
which often elicit strong opinions and engagement from viewers, as we saw in 1. In comparison, Cluster
2, which focuses on sustainability and ecology, has a lower number of comments (39,140) but still shows
significant engagement, reflecting interest in environmental issues.
59
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
5. Disinformation in TikTok
In the digital age, the spread of misinformation, which is often mistakenly believed to be reliable, is a
serious problem. As claim detection technology advances, the models for the detection of potentially
misleading content have become more accurate. We used an automated claim detection model from
Huggingface1 flagged a significant number of videos, suggesting they were worthy of fact-checking.
Specifically, 5,352 videos were initially identified as requiring review due to concerns of possible
misrepresentation.
However, after further analysis using clustering techniques, we discovered an interesting finding:
the actual number of videos related to disinformation and fake news was significantly lower than the
original estimate, as we saw in 4.1. This difference highlights the importance of using more diverse
methods to distinguish reviewable content from non-reviewable content.
Using a clustering algorithm allows us to divide tagged videos into distinct groups based on their
subject content and inferred features. It’s worth noting that we’ve divided these videos into seven
groups, each revealing unique interaction patterns and thematic focus.
Group 0 emerged as a hub for sutainable fashion practices, clothing upcycling and sewing techniques.
Our validation process, including consultation of correborated information of fact-checking agency EFE
Verifica [13], confirmed the absence, mostly, of misinformation associated with these keywords. Instead,
it underscored a commitment to promoting responsible consumption practices. Group 1 indicates a
focus o social interaction, practical tips and entertaiment, which, for the most part, does not contain
misinformation
Group 2 is distinguished by its focus con sustainability, environment and ecological awareness. It was
underscored a commitment to promoting responsible sustainability and ecological discourse. Group 3
is characterized by discussions on weather manipulation, chemtrails and climate-related conspiracy
theories. Despite not all videos in this cluster perpetuating misinformation, those that did sparked
intense debates and polarized discussions among viewers. Group 4 focuses on climate change, global
warming, climate actions and its controversies. Although many of the videos are based on climate
actions, there are also some other certain controversies and debates.
Group 5 is characterized by its focus on climate action, saving the planet and environmental aware-
ness. Our validation process, confirmed the absence, mostly, of misinformation associated with these
keywords. Group 6 is characterized by content related to recycling, waste reduction and sustain-
able practices. The commitment to promoting responsible sustainability and recycling practices was
highlighted.
6. Conclusion
As we have seen, in the digital era, social media platforms have become critical battlegrounds where
misinformation about critical issues like climate change proliferates. Our study utilized claim detection
and clustering techniques to analyze a substantial dataset of videos. Initially flagging 5,352 videos for
potential misinformation, we found through clustering that the actual prevalence of misinformation,
particularly in Cluster 3 and some in Cluster 4, despite existing, was lower than expected.
The rest of clusters predominantly featured content promoting sustainability and environmental
awareness, corroborated by information of fact-checking agency EFE Verifica [13]. In contrast, Cluster
3 and 4 contained videos with keywords associated with misinformation and conspiracy theories.
Collaboration with multilingual fact-checking platforms underscored the need for vigilant media
consumption practices.
In summary, our study contributes to understanding and addressing the challenges posed by misinfor-
mation in digital media, emphasizing the importance of informed media consumption and collaborative
efforts in safeguarding information integrity.
1
https://huggingface.co/Nithiwat/xlm-roberta-base_claim-detection
60
Clara Baltasar et al. CODAI Workshop Proceedings 54–61
Acknowledgements
This work has been funded by Grant PLEC2021-007681 (XAI-DisInfodemics), by the “European Union”
or by the “European Union NextGenerationEU/PRTR”, by grant PCI2022-134990-2 (MARTINI) of the
CHISTERA IV Cofund 2021 program, funded by MCIN/AEI/10.13039/501100011033 and by the “European
Union NextGenerationEU/PRTR”, by Calouste Gulbenkian Foundation, under the project MuseAI -
Detecting and matching suspicious claims with AI and by European Comission under IBERIFIER Plus -
Iberian Digital Media Observatory (DIGITAL-2023-DEPLOY- 04-EDMO-HUBS 101158511).
References
[1] U. K. H. Ecker, S. Lewandowsky, J. Cook, et al., The psychological drivers of misinformation
belief and its resistance to correction, Nature Reviews Psychology 1 (2022) 13–29. URL: https:
//doi.org/10.1038/s44159-021-00006-y. doi:10.1038/s44159-021-00006-y.
[2] V. Galaz, H. Metzler, S. Daume, A. Olsson, B. Lindström, A. Marklund, Ai could create a perfect
storm of climate misinformation, 2023. arXiv:2306.12807.
[3] M. Aldwairi, A. Alwahedi, Detecting fake news in social media networks, Procedia
Computer Science 141 (2018) 215–222. URL: https://www.sciencedirect.com/science/article/pii/
S1877050918318210. doi:https://doi.org/10.1016/j.procs.2018.10.171, the 9th Inter-
national Conference on Emerging Ubiquitous Systems and Pervasive Networks (EUSPN-2018) / The
8th International Conference on Current and Future Trends of Information and Communication
Technologies in Healthcare (ICTH-2018) / Affiliated Workshops.
[4] P. N. Petratos, Misinformation, disinformation, and fake news: Cyber risks to business,
Business Horizons 64 (2021) 763–774. URL: https://www.sciencedirect.com/science/article/pii/
S000768132100135X. doi:https://doi.org/10.1016/j.bushor.2021.07.012, cIBER SPE-
CIAL ISSUE: CYBERSECURITY IN CRISIS.
[5] J. Huertas-Tato, A. Martín, D. Camacho, Bertuit: Understanding spanish language in
twitter with transformers, Expert Systems 40 (2023) e13404. URL: https://onlinelibrary.
wiley.com/doi/abs/10.1111/exsy.13404. doi:https://doi.org/10.1111/exsy.13404.
arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/exsy.13404.
[6] A. Yang, Exploring the network and topic stability in climate change deniers’ disinformation
network: A longitudinal study (2024).
[7] B. Braga Pereira, S. Ha, Environmental issues on tiktok: Topics and claims of misleading informa-
tion, Journal of Baltic Science Education 23 (2024) 131–150. doi:10.33225/jbse/24.23.131.
[8] F. J. Basch CH, Yalamanchili B, #climate change on tiktok: A content analysis of videos, J
Community Health 47(1) (2022) 163–167. doi:10.1007/s10900-021-01031-x.
[9] T. Company, Tiktok api, 2024. URL: https://developers.tiktok.com/.
[10] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, I. Sutskever, Robust speech recognition
via large-scale weak supervision, 2022. URL: https://arxiv.org/abs/2212.04356. doi:10.48550/
ARXIV.2212.04356.
[11] N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks,
in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,
Association for Computational Linguistics, 2019. URL: https://arxiv.org/abs/1908.10084.
[12] M. Grootendorst, Keybert: Minimal keyword extraction with bert., 2020. URL: https://doi.org/10.
5281/zenodo.4461265. doi:10.5281/zenodo.4461265.
[13] Agencia EFE, EFE Verifica, ???? URL: https://efe.com/.
61