IA as a tool to combat disinformation. Approaching a model focused
on hoaxes in an electoral context
La IA como herramienta para combatir la desinformación. Planteamiento de un
modelo enfocado en los bulos en un contexto electoral
doxa.comunicación |
nº 41, pp. 511-533
| 511
July-December of 2025
ISSN: 1696-019X / e-ISSN: 2386-3978
How to cite this article:
Herrero de la Fuente, M.; Sancho Belinchón, C. and Sedeño López, J. (2025). IA as a tool to combat disinformation. Approaching a model
focused on hoaxes in an electoral context.
Doxa Comunicación
, 41, pp. 511-533.
https://doi.org/10.31921/doxacom.n41a2840
Mercedes Herrero de la Fuente.
PhD in Information Sciences (Complutense University of Madrid), Master’s Degree in
Radio Communication (Spanish National Radio-Complutense University of Madrid) and Master’s Degree in Linguistics
Applied to the Teaching of Spanish as a Foreign Language (Antonio de Nebrija University). Researcher with an active
six-year research period and member of the INNOMEDIA research group. Principal researcher of the Chair in Cinema,
Women and Education, promoted by EGEDA (Audiovisual Producers’ Rights Management Association) and Platino
Educa. Member of the R+D+I research project COM2GENDER, focusing on digital divides in university education. She
has previously participated in other competitive research projects with public funding, including COMPENSA, focused
on the labour insertion of people with disabilities in the audiovisual sector. He has been a research fellow at Cornell
University (USA), Saldford University (UK), Radboud Universiteit (e Netherlands) and Univerzita Karlova (Czech
Republic). She publishes articles in highly indexed journals focusing on: application of technology to news discourse,
new professional proles and women’s participation in the audiovisual sector. She is currently Coordinator of the PhD
in Innovation in Digital Communication and Media and an accredited lecturer at the Antonio de Nebrija University. She
has been a news producer at Telemadrid for fteen years
.
Nebrija University, Spain
[email protected]
ORCID:
0000-0002-5361-9056
Celia Sancho Belinchón.
PhD in Audiovisual Communication, Advertising, and Public Relations. She currently directs
the Master’s Degree in Digital and Data Journalism at Antonio de Nebrija University and teaches in the Journalism
degree program and the Master’s Degree in Political Communication and Crisis and Emergency Management. Her
research interests include journalism, digital communication, social media, social media strategy, fact-checking, and
advertising.
Nebrija University, Spain
[email protected]
ORCID:
0000-0001-5979-1853
is content is published under Creative Commons Attribution Non-Commercial License. International License CC BY-NC 4.0
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
512 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
Recibido: 20/11/2024 - Aceptado: 14/06/2025 - En edición: 19/06/2025 - Publicado: 01/07/2025
Resumen:
La inteligencia articial (IA) ha contribuido a la desinformación por
su capacidad para generar contenidos falsos. Pero el potencial de esta
tecnología puede también enfocarse en diseñar un prototipo de herra
-
mienta que detecte los bulos, en concreto aquellos amplicados en redes
sociales y en contextos electorales. Este artículo analiza los principales
patrones seguidos por las noticias falsas lanzadas en X durante las úl
-
timas elecciones catalanas (12 mayo 2024), siguiendo criterios como la
temática, el formato, el origen o su difusión, entre otros. Con la infor
-
mación obtenida se elabora de forma preliminar un recurso de IA con
capacidad de reconocer tales contenidos. Partimos de estos resultados
concretos: el tema más recurrente es la inmigración, predomina el for
-
mato texto más fotografía, en la mayoría de los casos procede de perles
registrados como un ciudadano cualquiera y los medios convencionales
no participan, en general, en su propagación. Sobre estas pautas plan
-
teamos las principales características de un sistema IA que combina
patrones de difusión con análisis de texto, imágenes y sentimiento, que
junto con la vericación en tiempo real de hechos nos permita ltrar
con un grado suciente de sensibilidad (proporción de bulos correcta
-
mente identicados) y especicidad (proporción de contenidos veraces
erróneamente clasicados como bulos).
Palabras clave:
Inteligencia articial; desinformación; vericación; elecciones
catalanas; algoritmo.
Received: 20/11/2024 - Accepted: 14/06/2025 - Early access: 19/06/2025 - Published: 01/07/2025
Abstract:
Articial intelligence (AI) has contributed to disinformation through
its ability to generate false content. But the potential of this technology
can also be focused on designing a prototype tool that detects hoaxes,
particularly those amplied in social networks and in electoral
contexts and moments of political relevance. is article analyses
the main patterns followed by the fake news launched on X during
the last Catalan elections (12 May 2024), following criteria such as
subject matter, format, origin and dissemination, among others.
With the information obtained, an AI resource with the capacity to
recognise such content is preliminarily developed. We start from these
specic results: the most recurrent topic is immigration, the text plus
photograph format predominates, in most cases it comes from proles
registered as any citizen, and the conventional media do not generally
participate in its propagation. Based on these guidelines, we propose
the main characteristics of an AI system that combines dissemination
patterns with analysis of text, images and sentiment, which, together
with real-time verication of facts, allows us to lter with a sucient
degree of sensitivity (proportion of hoaxes correctly identied) and
specicity (proportion of truthful content erroneously classied as
hoaxes).
Keywords:
Articial Intelligence; disinformation; verication; Catalan elections;
algorithm.
Jorge Sedeño López.
PhD in Computer Engineering, Master’s Degree in Information and Communications Technology
Management and Computer Engineering from the University of Seville. He has more than twenty years of professional
experience in dierent Public Administrations and belongs to the Senior Corps of Information Systems and Technologies
of the State Administration and to the Corps of Experts in Information Technology of the Bank of Spain. His lines of
work are oriented towards digital transformation through agile methodologies, e-Government, Data Governance
and Articial Intelligence. He belongs to the ES3 research group (Engineering and Science in Software System) and
is a member of the EQUAVEL Project PID2022-137646OB-C31 funded by MICIU/AEI /10.13039/501100011033 and by
FEDER, EU. He is a lecturer on the Master’s programme in Digital and Data Journalism at the University of Nebrija, as
well as at other public and private institutions.
Sevilla University, Spain
[email protected]
ORCID:
0000-0002-5368-5547
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 513
1. Introduction
Articial intelligence (AI) is present in many areas of our society, especially in the journalistic sector (Lopezosa-García et al.,
2024). e rst examples of its use in writing date back to 2014, when the Associated Press agency began to use it in sports
summaries and business reports (Badgamia, 2023). However, it is in the present time when the debate about generative AI
has gained strength, due to its potential for aggravating the phenomenon of disinformation. e present work is based on this
problem, but it is centered on the possibilities of AI as a tool that can help in the detection of fake news.
According to the UNESCO (United Nations Educational, Scientic and Cultural Organization) (2021), AI is the “simulation of
human intelligence processes by machine”. ese developments include learning, reasoning, and self-correction. With respect
to this denition, Blanco-Marañón (2023) emphasizes the term “simulation”, arming that imitating does not mean being
equal. Criado-Grande understands that the former refers to the possibility that machines attain “some kind of rationality
through the perception of the environment with which they interact” (2021, p.351), using sensors, obtaining and processing
data, reasoning on them, and adopting decisions.
In the area of communication, one of the changes that has been accelerated by AI is the increasingly liquid character of
information. us, “we will see texts that will turn into images, audio, or video, which will imply the alteration of production
models and processes, distribution, and monetization, as never experienced before” (Cerezo-Guilarranz, 2024, p.49). e
potential of AI is presented as a double-edged sword that can function as an ecient tool for journalism, but it also turns into
a threat due to its capacity to generate fake contents in any format.
From the technical point of view, it will be necessary to identify sources of data, which in the case of disinformation, will be
related to the web and social networks, in addition to knowing how this information will be stored and accessed. According
to the study by the IDC (International Data Corporation), the “data sphere” will grow to 175 ZB
1
in 2025 (Reinsel et al., 2018,
p.3) which is a great challenge. It is also necessary to know who this information will be treated, a eld in which data science
(articial intelligence and machine learning)) comes into play for developing indicators of the degree of falseness/truthfulness
of the information. Lastly, the way of extracting and presenting the value obtained from the previous treatment and its degree
of quality will have to be dened.
1.1. Disinformation and AI
e EC (European Commission) denes disinformation as “all forms of false, inaccurate, or misleading information designed,
presented and promoted to intentionally cause public harm or for prot” (2018, p.3). e emphasis is therefore on the intention
of causing harm to obtain prot, even if it is done in an unethical manner. For Alandete-Ballester (2019), fake news do not have
to be an outright lie; they can have some link with what is occurring, but they are characterized by distorting reality in pursuit
of sensationalism.
1
1 ZB (Zettabyte) = 1.000.000.000.000 GB (Gigabyte).
514 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
e phenomenon of disinformation is a serious concern for democratic countries (Rodríguez-Martelo et al., 2023). rough
the use of fake news, the intent is to manipulate the population and undermine the main political institutions (Arrieta-Castillo
and Rubio-Jordán, 2023). e White Paper against Disinformation published by the Spanish Government in 2022 warns against
the threat it implies to political stability and national security, “due to its potential to corrupt public debate, erode trust in the
institutions, manipulate public opinion, and condition foreign policy” (2022, p.9).
Disinformation is increasingly aecting public opinion, according to dierent national and international reports. e Reuters
Institute arms that this concern has grown in 2023 by two points with respect to the previous year, and 56% of those surveyed
fear not being able to distinguish what is true and what is false, when reading news on the Internet (Newman, 2024, p.17).
AI possesses sophisticated tools that can be used to amplify this undesirable phenomenon. e best known in recent years
have been the automatic generation of text and bots. Both have been used in social networks for the massive fabrication of
misleading texts and their dissemination through fake proles, widening their reach.
e most novel element as of today is the deepfake, or a video and audio in which the images and sound (normally both)
have been manipulated (Herrero-De-La-Fuente and Ríos-Calvo, 2022). As Deeptrace (2019) points out, the rst creations of
this kind emerged in November, 2017, with the creation of a forum in Reddit with the same name centered on the use of deep
learning programs to edit pornographic videos. Since then, the availability of AI tools for this type of edits has continuously
grown, with some of them being easy to access and manage. It can be said that a proof of this is that the circulation of deepfakes
has grown by 5505% between 2019 and 2023, according to the online security organization Home Security Heroes (2024).
Deepfakes represent a loss trust between society and images (Jacobsen and Simpson, 2023), and due to this they represent a
change in paradigm, where “seeing is believing” ceases to make sense
2
.
As pointed out, social networks are an essential agent in the dissemination of disinformation. In the past few years,
communication media have lost their monopoly on the distribution of news, so that informational contents have proliferated
outside of media circuits and are propagated by millions of accounts created by individuals, political groups, businesses, or
any organization (González-Quintero and Cardona-Restrepo, 2023), who may or may not show their real identity. In Spain,
the number of daily Internet users (87%) is already higher than those who watch television (81%) (AIMC, 2024, p.38 and p.64).
Access to current information was among the main uses of the Internet in 2023 (as declared by 60% of those polled), at the
same time that 70% used it to navigate social networks, and 97% for instant messaging (AIMC, 2024, p.67). It must also be
mentioned that due to its closed nature, it constitutes one of the main channels of dissemination of fake news (Díez-Garrido
et al
., 2021).
In addition, the news created in the journalistic area, that is, by journalists or experts, lose relevancy within the set of supposedly
informational content that is consumed in social networks, especially by the youth (Newman, 2024, p.11). According to the
last report by the Reuters Institute, “while mainstream journalists often lead conversations around news in Twitter and
Facebook, they struggle to get attention in newer networks like Instagram, Snapchat, and TikTok” (Newman, 2024, p.13), where
2
One of the most notorious in recent years was the manipulated video of Ukrainian President Volodymyr Zelensky declaring his country’s surrender to Russia in
the rst days of the war that began in February 2022.
https://cutt.ly/0w2geeXl
<p>[Enlace de URL / hc (has AS)]</p>
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 515
inuencers are more prominent. erefore, this data shows a trend that was also observed by the Edelman Trust Barometer,
which detected, in its last global study, an increase in mistrust in the media in fteen out of twenty-eight countries consulted.
Italy, Germany, and Brazil were the three countries in which this media discredit is greatest, with Spain found only four places
behind (2024, p.43).
1.2. AI as an instrument to combat disinformation
Peña-Fernández
et al
. (2023) point out that among the main applications of articial intelligence is the development of tools to
detect disinformation. As García-Marín explains, “AI makes it possible to determine the credibility of news sources based on
their reputational analysis, while oering a powerful answer to identify false proles on social networks” (2021, p.53). It is also
able to detect disinformation contents through the use of computational linguistics (with semantic and syntactic models) and
non-linguistic models, to discover image manipulation (photographs or videos).
Along with big data, AI can be an instrument to unmask fake content, as pointed out by recent studies (Moreno-Espinosa et al.,
2024; García-Marín, 2021; Flores-Vivar, 2019). Among the most utilized devices to nd fake news, dierent types of bots stand
out, developed in many cases with collaborations between universities, companies, and media. ey are based on aspects
such as adaptive algorithms, which examine sources and diusion patterns, mainly. Some examples are Fact Machine (from
the Brazilian verier Aos Fatos), TruthBuzz (promoted by the International Center for Journalists) or Les Décodeurs (from the
newspaper
Le Monde
). Recently, many AI tools have been developed based on machine learning, which work with linguistic
patterns through automated learning classiers, indicating the veracity of news bit as a function of dierent variables (Luengo-
Cruz and García-Marín, 2020). Here we mention, to cite one of them, Fakebox, which discriminates articles written in a similar
manner to real news articles, and texts that do not follow these guidelines, providing a score (Telefónica Tech, 2018). Other
novel systems include ClaimBuster (
https://idir.uta.edu/claimbuster/
) and Full Fact (
https://fullfact.org/
) with the latter
especially designed for political content with real-time systems and databases of veried facts (although not in Spanish). In
the present article, in the presentation of our results, an in-depth analysis of these resources will be performed.
Social network platforms lead many projects that develop AI systems to automatically eliminate malicious content through
text-based analysis; some of them include Facterbot or Projeto Lupe (Flores, 2019). However, many publications are pictures,
videos, or audios, and the verication methods for these are still not very developed (Moreno-Espinosa et al., 2024). is implies
that the identication of disinformation is an important challenge in the eld of automated learning (machine-learning) and
articial intelligence. It can be said that there is not only one “better” algorithm for this, as the adequate approach depends on
many factors, such as the type of data and the available formats, the complexity of the problem and the specic characteristics
of fake news that are to be detected, which points to the need for a multidisciplinary approach (Ruo et al., 2023).
e European Union has already started to promote initiatives to combat disinformation from dierent approaches. Since
2018, a series of areas of action were established, in: specic research on this phenomenon in the dierent areas involved
(the project Fandango has existed since 2018), the creation of independent networks of veriers (FactCheckEU was created in
2019), as well as the promotion of media literacy, among others. e community institutions are committed to tackling fake
news with a cross-cutting strategy, which goes beyond the creation of technological tools for the detection of fake news. ese
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
516 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
resources must rely on information professionals and all citizens. “e key is to have citizens that understand the importance
of obtaining quality information from reliable sources, that are capable of identifying potentially false content, and, in short,
that value the truth” (Sádaba and Salaverría, 2023, p.27). On this respect, the new Regulation (EU) 2024/1689 of the European
Parliament and of the Council of 13 June 2024, which establishes harmonized rules on AI, mentions the potential negative
eects “on the democratic framework, civic discourse, and electoral processes” (BOE, 12 julio 2024).
2. Method
e main objective of this study is to extract the main characteristics of the hoaxes spread in the last regional elections in
Catalonia (May 12th, 2024) and to extrapolate these rules to design a prototype of an AI tool that will be able to detect fake news
related with this typology in the electoral context.
For this, the present study was developed in three phases. e rst was descriptive, and was centered on consulting sectorial
reports and previous studies. A review was performed of the literature related with our object of study and the most current
data, to establish the basis for a reected directed towards creating a conceptual framework on which our work is based.
e second phase consists of a content analysis of the most popular fake news on the X (formerly Twitter) website during the
pre-campaign period (starting on April 2nd), campaign (from May 3rd to May 10th) and the day of the Catalonian elections
(May 12
th
, 2024). e selection of the fake news (fourteen) was performed starting with an article from the blog Maldito Bulo
(Maldita, 2024), which collects the ones that had the greatest repercussion in social networks (specically X). It is believed
that this sample represents a larger set, as it gathers the most signicant ones. Credit is given to the compilation carried out by
Maldita, as it a foundation that is dedicated to verication, of recognized prestige in Spain
3
. Nevertheless, intentional sampling
is proposed, which follows a criteria of ease, as it is an accessible sample to which Maldita has already applied some lters. It
is considered to be a starting point for this initial phase of the development of an AI tool, as pointed out in the explanation of
the third phase of the study.
It must also be added that all the contents analyzed were written in Spanish, except for those published by the television
channel TV3, which were in Catalan, although this media outlet re-tweets texts in Spanish.
e analysis pattern that was followed to analyze the fourteen highlighted fake news was the following:
3
Maldita was founded as a non-prot digital native media outlet in 2018 and has since its inception collaborated with the European Commission on several
initiatives to combat disinformation. It is a pioneer in fact-checking work in Spain and enjoys international recognition, being a member of the International
Fact-Checking Network.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 517
Table 1. Hoax analysis variables
Variable
Description
Date
Day the hoax was published
Subject matterematic area to which the hoax refers
*
Political partyPolitical party subject of the hoax
Format
Text; text and photograph; edited photograph; text and video;
video; text and audio
Writing of the hoax
e way in which information is addressed (direct, indirect,
explanatory, sensationalist)
Type of prole that created the hoax
Political spokesperson (public oce, party oce, or related
organization); media outlet; unknown citizen
Dissemination of the hoax by the media
Yes/No
Digital media are considered in social network X. Traditional
media are analyzed, such as: newspaper (
El País, El Mundo,
ABC, La Razón, El Diario, Público
), radio stations (
SER,
COPE, Onda Cero, RNE
) and television channels (
La 1, AT3,
Telecinco, La Sexta and TV
3)
* We chose to leave this variable open, without dening categories that could predetermine it.
Source: prepared by the authors
e third phase is exploratory, and the intention is, once the selected fake news have been analyzed and the common patterns
analyzed, the design of a prototype AI tool that includes all the mechanisms to identify these patterns, in order to detect them
in later news articles and to determine if they are fake or not, with a sucient sensitivity and specicity.
e sensitivity refers to the rate of true positives, and the specicity to the rate of false positives, so that for a binary model of
classication, as the one proposed (false information, true information), the greater the sensitivity (more positives) and the
lower the specicity (less mistakes), the closer we are to a perfect classier.
For each of these patterns, a series of algorithms would be proposed that will be tted by with hyperparameters. A
hyperparameter is a conguration that is adjusted externally, that is not learned from the data, but instead is established
before the training process of the model, that is a set of parameters outside of the algorithm itself, that can improve or adapt its
performance (Simanjuntak et al., 2024) with respect to the problem proposed.
518 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
Likewise, given the myriad of combinations and adjustments, the selection of the performance metrics related with sensitivity
and specicity in a binary classication model (false information-fake news, true information-not fake news) is especially
relevant.
Lastly, to address the problem of classication of fake news, and to understand why certain decisions are made, the explanatory
(or interpretable) models are crucial. ese models must not only be precise, but must also provide a clear justication for
their predictions. erefore, the system must have the capacity to explain the decision taken (Hashmi et al., 2024).
3. Results
3.1. e main fake news s in X with respect to the regional elections of Catalonia
We begin the presentation of the rst stage of the results by showing a table with the identication data that are closely related
with the content of the fourteen main hoaxes detected, as shown in Table 2.
Table 2. Identication sheet of the main hoaxes in the Catalan elections (date, topic, party)
Denominatio
n
Description of the hoax
DateTopic
Political party
4
Casa Tarradellas
False statements by the company’s
founder explaining that he only hires
Catalans
02/04/2024Catalonia/BrandNone
Jordi Évole joins the ERC
Fake news about the journalist
joining the ERC
09/04/2024
Party membershipsERC/ Government
Grandfather Puigdemont
Misattributed photograph of
Falangist Gregorio Martín Mariscal
17/04/2024Candidates
Junts+ Puigdemont
COVID Vaccination Illa
Distorted statements about his
vaccination
22/04/2024
Candidates/ Health
PSC
Ringworm epidemic
False health alert due to ringworm
epidemic
24/04/2024HealthNone
“Islamic emirate” posters
False image of welcome signs placed
at the entrance to various towns
25/04/2024Immigration
Frente Obrero
Illa’s ag event
Real image taken out of context
of a march for Spanish unity with
dierent political parties present
26/04/2024CandidatesPSC
4
ERC: Esquerra Republicana de Catalunya; PSC: Partido Socialista de Catalunya, part of the PSOE: Partido Socialista Obrero Español; Junts: Junts per Catalunya.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 519
Social assistance to
Moroccans
Fake news claims that Moroccan
families in Catalonia are living solely
on aid.
26/04/2024
Subsidies/
Immigration
ERC / Government
Guaranteed Income
False information about aid to
Moroccan families from the former
PIRMI (
03/05/2024
Subsidies /
Immigration
ERC / Government
Implementation of Arabic
in schools
False subsidy from the Generalitat to
implement the language
03/05/2024
Education/
Immigration
ERC / Government
Buying of votes by PSOE
advisor
Alleged vote buying via the advisor’s
WhatsApp
10/05/2024Corruption
PSOE and PSC
Private swimming pools
expropriated
False news about the expropriation
of private swimming pools due to the
drought
10/05/2024Drought
ERC / Government
PIRMI (Minimum
Insertion Income) Aid
Social aid to families of any
nationality, which no longer exists in
Catalonia
10/05/2024
Subsidies /Education /
Immigration
ERC / Government
Rodalies Sabotage
5
Alleged theft of copper cables from
the Rodalies (local trains) paralyzes
the service
12/05/2024
Transport / Corruption
PSOE
Source: prepared by the authors
As the table shows, fake news were observed spread out from April 2
nd
, 2024, to May 12
th
, 2024, although not evenly, given that
they were not published every day. Only the hoaxes found in Table 2 were found, with April 10
th
being the date with the most
fake news. e period encompasses the period from pre-campaign to election day, as the confrontational tone in recent years
has amplied the partisan diatribe over time. It is striking that the climate of tension intensies on election day itself.
It was observed that the most recurrent thematic areas were related to social themes, and the most predominant issue was
immigration, which was behind ve contents related with subsidies/nancial aid, and education. Within the social topic, we
also nd health, but only the case of the false ringworm epidemic is treated as such, without it being used as an excuse for other
issues. We also found allusions to candidates and new members from dierent parties (four). Likewise, we found hoaxes that
are apparently recurrent in matters outside of politics, such as a brand (Casa Tarradellas), or the problem of drought, to launch
contents related to Catalonian nationalism or economic policies against private property, all of which to discredit the central
government. With the same purpose, we found discussions about buying votes via WhatsApp, or incidences local transport
during election day.
5
Rodalies is the public service of commuter and medium-distance regional trains in Catalonia. It was transferred by the Ministry of Public Works to the Catalan
autonomous region in 2010 and 2011.
520 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
As Table 2 shows, only two hoaxes were found that did not directly correspond to a specic political party. Of the remaining
twelve: six alluded to the ERC party (which governed Catalonia when the research took place), three to the PSC (one of them
shared with the PSOE), one to the PSOE, one to Junts, and another one to Frente Obrero. All of this seemed to also obey a
strategy of wearing down the institutions, when trying to erode the parties that lead the regional and central executive branches.
Within the sample analyzed, the most utilize format was the combination of text and photos, with a total of ten cases (the
image was not edited and it did not have any superimposed text). It must be claried that in two, plus the body of the tweet,
there was only a photograph (without text), and it appeared with encrusted text.
We found another tow formats in the fourteen hoaxes analyzed, which were: audio accompanied by text, which showed
the spoken testimony of an immigrant from the Maghreb, who received social aid (PIRMI) in an illegal manner; and video
accompanied by text, in the tweet about the rejection of Salvador Illa, a socialist candidate, to become vaccinated against
COVID-19.
It must be pointed out that some formats also existed for which no results were found, such as only text or only video. In any
case, none of the publications seemed to be associated to a user with a high level of digital skills.
In general, the texts tended to be sensationalist, using upper case text, and excessive use of punctuation marks, and spelling
mistakes in some cases. In most of them, these texts can help to capture more attention from the spectator, as they use bright
colors such as red. erefore, it can be said that this simple aesthetic and based on visual aspects achieves a very eective
message that is easily assimilated by the audience, which contributes to the deception. Likewise, no hoaxes were found that
were written indirectly, but instead all of them were written in a direct manner, based on simplistic arguments to capture
the attention of the spectator. No explanations were included, as the strategy is that of simplication and the presentation of
maximalist arguments.
When considering the X proles that published the hoaxes analyzed, an almost complete coincidence was observed, as most
(eleven) originated in this social network from unknown users in the public sphere. ree cases did not t this pattern: the
rst, created by the leader of Frente Obrero; the second, created by political spokepersons of the VOX party; and the third
by political spokepersons of the Partido Popular. In these three cases, the member of these political parties did not have any
public positions during the dissemination of these hoaxes.
Within the three hoaxes published by political ocials of the indicated parties, the rst, published by Roberto Vaquero, shows
welcome signs at road entrances to the Autonomous Community of Catalonia that read “Islamic Emirate of Catalonia”. It came
from the X prole of this Frente Obrero leader (extreme right party). At the same time, citizens that agree with this information
disseminate it in their personal accounts.
e second is focused on the boycott to Catalan local trains by the PSOE, to prevent citizens from travelling and exercising
their right to vote. PP leaders launch accusations in this sense, which are re-tweeted by members of the ERC, Junts, and Frente
Obrero.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 521
e last hoax refers to a supposed investment of the Catalonian regional government to implement the study of the Arabic
language at schools. Publications of this fake news have been found in X proles belonging to VOX spokepersons and followers
of this party (not identied as members).
With respect to the publications of the fake news analyzed in other communication media, in general terms, the media did not
echoed them in their ocial X proles; but two exceptions were found to this pattern.
e rst was the fake news about Salvador Illa and the supposed fact of not having gone to a health center to be vaccinated
against COVID-19. In this case, it was a Honduran television medium, “Girasol TV”, which included it in its X prole.
e second was related to the alleged sabotage of the Rodalies local trains. e traditional media (indicated in the pattern of
analysis), as well as the regional television of Catalonia (TV3) resorted to the clickbait, publishing in X headlines with the terms
“alleged sabotage” or “election sabotage”. en they refer to the theft of copper suered by the commuter rail network, but in
no way do they reproduce the core of the hoax, that is, that the Partido Socialist is behind the regional train (Rodalies) travel
problems during election day.
In order to quantify the dierent variables analyzed, a summary is shown in Table 3:
Table 3. Quantication of results in analysis variables (not directly related to: topic, date, party)
VariableDescription
Number of hoaxes
Format
Text0
Text and photograph
10
Edited photograph
2
Text and video
1
Text and audio
1
522 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
Writing of the hoax
Direct14
Indirect0
Explanatory0
Sensationalist14
*
Type of prole that created the hoax
Political spokesperson
3
**
Communication media
0
Unknown use
11
Dissemination of the hoax by the media
Yes, they disseminate
12
Do not disseminate
2
* All hoaxes fall into both categories: direct and sensationalist
** None of these political leaders hold public oce
Source: prepared by the authors
To provide a greater clarity about the repercussion of the hoaxes analyzed in the sample, the related data can be consulted in
the following table:
Table 4. Impact of hoaxes in X
Denomination
Prole that
originates the hoax
Dissemination of the
hoax by the media
Number of
retweets
Comments
Casa TarradellasUnknown users
n/a
*
134,000875
Jordi Évole joins the
ERC
Unknown users
n/a
423
Grandfather
Puigdemont
Unknown users
n/a84,00028
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 523
COVID Vaccination IllaUnknown usersGirasol TV
145,0001,714
Ringworm epidemicUnknown users
n/a45854
“Islamic emirate”
posters
Roberto Vaquero
(Frente Nacional)
n/a100
Illa’s ag eventUnknown users
n/a2,600369
Social assistance to
Moroccans
Unknown users
n/a2,6988
Guaranteed IncomeUnknown users
n/a1,29383
Implementation of
Arabic in schools
Political spokepersons
of VOX
n/a1,25825
Buying of votes by
PSOE advisor
Unknown users
n/a1,000256
Private swimming pools
expropriated
Unknown users
n/a1,4556
PIRMI (Minimum
Insertion Income) Aid
Unknown users
n/a98718
Rodalies Sabotage
Political spokepersons
of the Partido Popular
TV3, El País, ABC, La 1,
AT3, Telecinco, La Sexta
and TV3
162,000493
* Not applicable
Source: prepared by the authors
As observed on the table, there are three publications that are clearly notable due to their expansion as compared to the rest,
and that allude to topics related with Catalonia, trying to discredit a businessman and a politician of that community, and
thereby fueling the anti-Catalan sentiment. e one with the greatest reach attacks the local Catalonian commuter trains, but
in this case the objective is the Government of Spain and the creator of the hoax is the Partido Popular.
524 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
3.2. Algorithms and features for developing a tool to detect hoaxes in regional elections in Spain
e patterns detected in the analysis of the previous phase serve as the basis for the third stage, which centered on the design
of an AI resource, concatenating dierent algorithms so that it is able to identify this type of hoaxes selected. Given that the
sample suers from a numerical limitation, the proposed tool is considered to be in an initial stage of development.
In this way, and based on the hoaxes analyzed
a posteriori
in the context of some regional elections in Spain, a detection
system will be outlined based on the patterns identied, for which the use of AI paradigms will be needed, through algorithms
and features (an individual characteristic or property of the input data that help the model recognize patterns and make
predictions) that are able to:
–
To analyze the dissemination of media, to detect the origin of the information that will be processed (anonymous or by
political spokepersons without a public oce).
–
To detect formats and styles in the information and to analyze the sentiment (multi-format information with a direct and
sensationalist style, that seeks to provoke polarization).
–
Fact-checking.
ese concepts will be broken down in more detail below.
3.2.1. Media dissemination
Within the analysis of media dissemination, the use of the algorithm PageRank, which is based on the concept of eigenvector
centrality, is proposed
6
, to identify inuencing accounts and patterns of dissemination of hoaxes.
As the main feature, we opted for the analysis of graphics, to model the relationship between these accounts and the
dissemination patterns, given that the generation of hoaxes took place through the X network, and we seek anonymous
accounts or accounts from political spokepersons that do not have any public oce.
3.2.2. Formats and styles
Given that the format pattern identied in the hoaxes in phase 2 combine text and another element, such as photographs
(edited or not), videos or audio (which will be later transformed to text), the following algorithms are proposed:
–
For the text, a transforming model will be used, that is, a type of neural network architecture for natural language proces
-
sing, due to its capacity to eectively manage text sequences and establish complex relationships between the words in
a sentence. Both BERT (Bidirectional Encoder Representations from Transformers) and RoBERTa (A Robustly Optimized
BERT Approach) can be adapted to detect disinformation (Zhang et al., 2024). Pre-trained models will also be utilized (Hu
-
6
PageRank is a sophisticated measurement of centrality that combines the idea of node importance with a probabilistic navigation model. It simulates the behavior
of a user randomly browsing the web and uses that simulation to calculate the importance of each node, based on the quantity and quality of the links pointing to
it. Eigenvector centrality means that the importance of a node depends not only on how many other nodes connect to it, but also on the relevance of those nodes.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 525
gginFace, FakeBERT, or FakeNewsBERT) along with the built-in
embeddings
7
for hoax analysis to improve their detection
(Yang et al., 2018). In this way, we will have hybrid models focused to the context of Spanish regional elections. is is what
is known as RAG architecture
8
.
–
For images and videos, EcientNet will be used, a family of convolutional neural networks (CNN) that have been shown
to be very eective in the classication of these contents, with the ability to capture ne details, to be very critical for the
detection of manipulations (edited photos), and which can be easily integrated with transforming models, to created a
robust multi-modal detection system. Likewise, the EcientNet B4 model has shown to have a detection precision of 92%,
of faces manipulated in videos, as it analyzes many parameters, such as facial expressions and irregularities in the images,
to dierentiate between real and fake videos (Priyaa et al., 2024).
–
For the audio, the Whispers model will be used, which is well known for its cutting-edge performance in audio to text con
-
version, achieving a high precision in the transcription (Haz et al., 2023), and afterwards, once the text has been obtained,
it will be processed just as the text from the hoax.
–
As described in phase two of the study, the hoaxes are written in a non-neutral manner (direct and sensationalist langua
-
ge). Currently, the most eective algorithms for the analysis of style and sentiment, due to the ability to capture the bi-di
-
rectional context of the words in a sentence, are BERT and RoBERTa. For this, classiers will be added. For example, for
the proposed study the sensationalist classier can be used, so that: 0 = non-sensationalist and 1 = sensationalist, creating
dierent classiers to indicate a direct study and to verify if they create a sentiment of polarization or not.
3.2.3. Fact-checking
Due to the speed with which hoaxes are spread, and the short period of time in which elections take place (in the hoax
identication sheet, we can observe that they even exist on the same day as the elections), it will be necessary for the chosen
system to work in real-time, and to be integrated with dierent communication media, and that it contains alert systems for
human veriers. erefore, ClaimBuster and Full Fact will be used (the latter, in addition, provides the verication of political
discourses and specic communication media, and it has a database of contrasted statements). At the same time, the BERT-
QA will be used to extract facts, with the most adequate feature being contextualization, which is dened as the verication
process against said databases.
7
Natural language processing technique that converts human language into mathematical vectors and is the basis of generative AI models, whose answer with the
highest number is closest to the question.
8
Retrieval-Augmented Generation (RAG) architecture is a hybrid model that combines information retrieval and text generation approaches. In this type of
architecture, the system rst retrieves relevant information from a database or external source and then uses that information as context to generate more
coherent and accurate responses or content. is integration allows the model to generate responses not only from its internal knowledge but also from its own
internal knowledge.
526 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
3.3. Process
Figure 1. Analysis process diagram
Source: prepared by the authors
e design of the proposed system can be observed in Figure 1, which shows the process described below, so that once the
hoax has been received, it will be tokenized
9
,
to become the input in the following sub-processes.
–
An inquiry will be made about the type of format, so that:
▪
If it is text, it will be analyzed with BERT-RoBERTa.
▪
If it is an audio, it will be transcribed with Whisper and posteriorly analyzed with BERT-RoBERTa.
▪
If it is a video, it will be analyzed with EcientNet B4.
–
Format and sentiment will be extracted
9
A token is a basic unit of text that the aforementioned algorithms use to process and analyze information. It can be a word, a subword, etc. Tokenization is the
process of dividing text into these basic units.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 527
–
Dissemination will be analyzed with PageRank to obtain source and disseminator.
–
e facts will be veried with ClaimBuster, and will be conrmed with Full Fact.
Lastly, all the information from each sub-process will be unied to provide a nal binary response [0,1] to decide if it is a hoax
or not.
Given the above, two fundamental aspects remain to validate the system proposed: the analysis of sensitivity and specicity,
and the assessment of explainability.
3.4. Sensitivity and specicity analysis
e True Positive Rate (TPR) and False Positive Rate (FPR) are key metrics used in the evaluation of classication models,
especially in binary classication problems. Our system intends to perform the binary classication (“hoax”, “not hoax”) of
each of the input pieces of information (Jeni et al., 2013). us:
–
TPR is also known as sensitivity. It measures the ratio of positive examples that are correctly identied by the model. It is a
measure of how well the model captures positive cases. TPR is calculated as TP / (TP+FN), where TP (True positives) is the
number of hoaxed correctly classied as hoaxes, and FN (False Positives) is the number of hoaxes incorrectly classied as
“not hoaxes”. us, the closer to 1 (all the hoaxes are detected) the better the sensitivity.
–
FPR, also known as specicity, measures the ratio of negative examples that are incorrectly classied as positives. It is a
measurement of the mistakes committed with respect to the negative examples. e FPR is calculated as FP / (FP+TN),
where FP (False Positives) is the number of “not hoaxes) erroneously classied as “hoaxes” and TN (True Negative) is the
number of “not hoaxes” correctly classied as “not hoaxes”. erefore, the closer to a 0 value (no truths are detected as a
hoax), the higher the specicity.
3.5. Explainability
Although not part of the system itself, explainability, that is, the way in which the system explains how it arrived to the
conclusion that a piece of information is a hoax or not, it an important component to grant it with reliability.
For this, SHAP (
SHapley Additive exPlanations
) will be used, which is based on game theory and provides consistent and
additive explanations. is means that the sum of the contributions of all the characteristics provides the prediction of the
model as a result, which is intuitive and easy to understand. SHAP can be applied to any type of model, for local interpretations
(individual prediction), as well as global ones (throughout the entire set of data), which is crucial for understanding both
specic cases and general trends in the context of fake news with political content, as the SHAP explanations can manage
complex interactions between characteristics. e latter is relevant in the analysis of fake news, where multiple factors (such
as the text content, the topic, the author, the source, and the date) can interact in non-trivial manners.
528 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
4. Discussion and conclusions
After the presentation of the results, and the previous analysis, it can be said that AI has a great potential in the area of
journalism, but it shown to be a double-edged sword. In the past few years, its high capacity to generate fake content has
been shown, which have been used profusely to manipulate public opinion, especially in electoral contexts. e damaging
consequences of disinformation for the correct functioning of democracies have led to a reection on the need to use this
same technology to more eciently combat the proliferation of hoaxes.
e study starts from a specic situation, such as the pre-campaign and campaign periods in the last regional elections in
Catalonia, to analyze the patterns followed in the creation and dissemination of fake news. In line with what other works such
as those by González-Quintero and Cardona-Restrepo (2023) have corroborated, the focus was placed on social networks as
the main way to propagate them, specically in the X network.
e analysis of the fourteen hoaxes detected by Maldito Bulo (2024) which were used as our sample, and which we consider
them to be representative of a broader set of this type of content, did not nd any cases of deepfakes, in spite it being a growing
trend worldwide, as stated in the report by Home Security Heroes (2024), a predominance was observed of supposedly 100%
fake news (ten out of fourteen), so that our results do not correspond to the scheme of distortion based on partially-true
information described by Alandete-Ballester (2019) (so common, on the other hand, in this type of phenomena).
e most recurring topics were social in nature, and in most of the cases connected with immigration, although they are
presented as news about diverse areas, among which we nd education, or social aids. In this aspect, we nd sensationalism
and the desire to discredit the institutions, as referred to by Arrieta-Castillo & Rubio-Jordán (2023). e attacks mainly directed
to ERC and the Partido Socialista (through its Catalonian division), at the helm of the Catalonian and Spanish governments,
respectively, demonstrate this purpose, which was intensied on the same day as election day, with the hoax about the
sabotage of the Rodalies local trains. For this, the aggressive tone shown by our data is believed to contribute to this aim, with
a predominance of exclamation marks or upper case letters in the published texts. All of this incites tension and mistrust,
and increases the risk for democratic stability, thereby illustrating one of the risks pointed out in the White paper against
Disinformation (Gobierno de España, 2022).
e disinformation came, in eleven of the fourteen samples studied, from proles that did not belong to someone recognizable
(such as a politician or a public servant), and which corresponded, apparently, from any citizen. In fact, the formatting was
simple (text plus a photograph in most the cases), and a minimum level of skill was needed only for the edited images. However,
it is possible that hidden beneath these seemingly innocuous proles, we may nd activists serving one party or another, with
those aliated with VOX being the most active. Outside of this pattern, we nd the far-right leaders (VOX and Partido Obrero),
who created some of these hoaxes (two) in our sample. e PP, on its part, added a third, the already-mentioned one about the
malfunctioning of the local trains during election day, which is the most widespread of the observed set.
e media contributes to this as they do not share any of the fake news collected, but in this case, they resort to clickbaits, and
also echo the disruptions to commuter trains with alarmist headlines. Far from being a positive conclusion, as only one case
was detected, this data indicates a harmful practice, which adds to the noise and confusion promoted in the X network.
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 529
After evaluating these results, we moved on to developing an AI tool that, based on the dynamics detected in the fabrication
and spread of hoaxes, may be able to identify attempts at disinformation related to an election.
We share with Ruo et al. (2023) the premise that there is no single “best” algorithm, so that the algorithms and features most
appropriate to the focus of our analysis were combined. e proposed AI system is specically adapted to the typology of fake
content so that, once the dierent components of the same have been determined (Table 3 and Table 4), a system is designed
that combines dierent techniques, algorithms and features, adapted to the selected sample. e dissemination patterns were
analyzed with PageRank (on the X network) and, in accordance with the conclusions of Zhang et al. (2004), we adapted the
treatment of images with EcientNet and the text information with a model of Transformers (BERT and RoBERTa) to our
interests, which also served for conducting a sentiment analysis (such as for example, sensationalism). Once the information
was tokenized, this must be veried with fact-checking tools, such as Claimbuster or Full Fact, which must have real-time
responses.
e proposed AI system aims to perform a binary classication (“hoax”, “not hoax”) of each of the information inputs, through
the process described in Image 1, so it will be crucial to measure sensitivity (ratio of positive examples correctly identied) and
specicity (ratio of negative examples incorrectly classied as positive).
In turn, and because the algorithms used in this system allow it, SHAP will be used to perform the explainability of the system’s
binary decision, where multiple factors (such as text content, topic, format, author, and date), can interact in non-trivial ways.
Among the limitations of this research, we nd the sample size, as the fourteen hoaxes selected by Maldita, while presenting
patterns that could be extrapolated to a larger set, do not constitute a large group. As we noted in the Methodology section,
the sample is a set of hoaxes intentionally chosen in a specic political context, without it being statistically or theoretically
representative.
Likewise, the design and development of the AI resource require rening the features detected in the patterns of the analyzed
hoaxes, which were used as external properties of each algorithm employed in the process. It is also necessary to perform
parameter adjustments on specic data from a broader set of hoaxes related to other Spanish regional elections, in order to
train the model and test the metrics applied for specicity and sensitivity on a larger sample over time.
e proposed AI resource is tailored to the specic characteristics of this type of content (false information in electoral
processes in Spain), which may be similar to other examples of disinformation from diverse situations and locations. erefore,
we consider the application of the model’s accuracy to new national electoral events and its suitability for such events held
outside our country to be the subject of the subsequent study phases, through the measurement of the proposed quality
parameters and the system’s ability to self-explain the results.
5. Acknowledgments
is article has been translated into English by Mario Fon to whom we are grateful for his work.
530 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
6. Specic contributions of each author
Name and Surname
Conception and design of the work
Mercedes Herrero de la Fuente, Celia Sancho
Belinchón and Jorge Sedeño López
Methodology
Mercedes Herrero de la Fuente, Celia Sancho
Belinchón and Jorge Sedeño López
Data collection and analysis
Celia Sancho Belinchón
Discussion and conclusions
Mercedes Herrero de la Fuente, Celia Sancho
Belinchón and Jorge Sedeño López
Drafting, formatting, version review and approval
Mercedes Herrero de la Fuente
7. Conict of interests
e authors declare that there is no conict of interest contained in this article.
8. Bibliographical references
AIMC (2024).
Marco General de los Medios en España 2024
.
https://cutt.ly/Kw2ghBxA
Alandete-Ballester, D. (2019)
Fake news: la nueva arma de destrucción masiva: Cómo se utilizan las noticias falsas y los hechos
alternativos para desestabilizar la democracia
. Deusto.
Arrieta-Castillo, C., y Rubio-Jordán, A. V. (2023). Periodismo de vericación en formato vertical: narrativas multimedia de los
vericadores en TikTok.
Ámbitos. Revista Internacional De Comunicación
, 60, 13-32.
https://doi.org/10.12795/Ambitos.2023.
i60.01
Badgamia, N. (1 mayo 2023). Explained. AI journalism: Can articial intelligence replace journalists?
WioNews
.
https://cutt.
ly/Rw2gWt3x
Blanco-Marañón, N. (noviembre 2023). Qué diría Aristóteles de la inteligencia articial
. Telos
, 123, 35-39.
https://cutt.
ly/6rWQCBxr
Boletín Ocial del Estado (BOE) (12 julio 2024).
Reglamento (UE) 2024/1689 del Parlamento Europeo y del Consejo, de 13 de
junio de 2024, por el que se establecen normas armonizadas en materia de inteligencia articial
.
https://www.boe.es/buscar/
doc.php?id=DOUE-L-2024-81079
Cerezo-Guilarranz, P.(2024). Tendencias 2024. Hacia el nal de la web abierta.
Programmatic Spain
.
https://cutt.ly/Sw2gWQTs
Comisión Europea (2018).
A multi-dimensional approach to disinformation.
https://cutt.ly/Srjh4cNL
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 531
Criado-Grande, J.I. (2021). Inteligencia Articial (y Administración Pública).
Eunomía. Revista en Cultura de la Legalidad
, 20,
348-372.
https://doi.org/10.20318/eunomia.2021.6097
Deeptrace (2019).
e state of deepfakes. Landscape, threats and impact.
https://sensity.ai/reports/
Devlin, J., Chang, M.W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language
understanding.
arXiv,
1810.04805
.
https://doi.org/10.48550/arXiv.1810.04805
Díez-Garrido, M., Renedo-Farpón, C., y Cano-Orón, L. (2021). La desinformación en las redes de mensajería instantánea.
Estudio de las fake news en los canales relacionados con la ultraderecha española en Telegram.
Miguel Hernández
Communication Journal
,
12
(2), 467-489.
https://doi.org/10.21134/mhjournal.v12i.1292
Flores-Vivar, J. M. (2019). Inteligencia articial y periodismo: diluyendo el impacto de la desinformación y las noticias falsas a
través de los bots.
Doxa Comunicación
, 29, 197-212.
https://doi.org/10.31921/doxacom.n29a10
García-Marín, D. (2021). Las fake news y los periodistas de la generación z. Soluciones post-millennial contra la desinformación.
Vivat Academia. Revista de Comunicación
, 154, 37-63.
http://doi.org/10.15178/va.2021.154.e1324
Gobierno de España (2022).
Lucha contra las campañas de desinformación en el ámbito de la seguridad nacional. Propuestas
de la sociedad civil
.
https://cutt.ly/VrWQOXJc
González-Quintero, J. I., y Cardona-Restrepo, P.(2023). Post-truth and Social Networks as Challenges for Journalism in the
Digital.
Ánfora, 30
(55), 332-359.
https://doi.org/10.30854/anf.v30.n55.2023.977
Hashmi, E., Yayilgan, S.Y., Yamin, M.M., Ali, S., & Abomhara, M. (2024). Advancing Fake News Detection: Hybrid Deep Learning
with FastText and Explainable AI.
IEEE Access
, 12, 44462-44480.
https://doi.org/10.1109/ACCESS.2024.3381038
Haz, L., Fajrianti, E.D., Funabiki, N., & Sukaridhoto, S. (14-15 octubre 2023).
A Study of Audio-to-Text Conversion Software
Using Whispers Model.
Sixth International Conference on Vocational Education and Electrical Engineering (ICVEE), Surabaya
(Indonesia).
https://doi.org/10.1109/ICVEE59738.2023.10348186
Herrero-De-la-Fuente, M., y Ríos-Calvo, C. Construcción de un escenario para la posverdad: redes sociales y desinformación
(2022). En A. Pérez-Escoda y J. Rubio-Romero (eds.).
Redes sociales, ¿el quinto poder? Una aproximación por ámbitos al
fenómeno que ha transformado la comunicación pública y privada
(pp.79-97). Tirant lo Blanch.
Home Security Heroes (2024).
2023 State of Deepfakes.
https://cutt.ly/QrWQP1Cu
Jacobsen, B.N., & Simpson, J. (2023). e tensions of deepfakes
. Information, Communication Society, 27
(6), 1095-109.
https://doi.org/10.1080/1369118X.2023.2234980
Jeni, L. A., Cohn, J.F., & De La Torre, F. (12 diciembre 2013).
Facing Imbalanced Data Recommendations for the Use of
Performance Metrics.
Humaine Association Conference on Aective Computing and Intelligent Interaction, Ginebra (Suiza).
https://doi.org/10.1109/ACII.2013.47
López-López, P.C., Lagares-Díez, N., y Puentes-Rivera, I. (2021).
Razón y Palabra
,
24
(111), 5-11.
https://razonypalabra.net/
index.php/ryp/article/view/1891/1681
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
532 |
nº 41, pp. 511-533
|
July-December of 2025
IA as a tool to combat disinformation. Approaching a model focused on hoaxes in an electoral context
ISSN: 1696-019X / e-ISSN: 2386-3978
doxa.comunicación
Lopezosa-García, C., Pérez-Montoro, M., y Rey-Martín, C. (2024). El uso de la inteligencia articial en las redacciones:
propuestas y limitaciones.
Revista de Comunicación
,
23
(1), 279-293.
https://doi.org/10.26441/RC23.1-2024-3309
Luengo-Cruz, M., y García-Marín, D. (2020). e performance of truth: politicians, fact-checking journalism, and the
struggle to tackle COVID-19 misinformation.
American Journal of Cultural Sociology
, 8, 405-427.
https://doi.org/10.1057/
s41290-020-00115-w
Maldito Bulo (17 mayo 2024).
14 bulos y desinformaciones sobre las elecciones autonómicas Cataluña del 12 de mayo de 2024.
https://cutt.ly/uenhODjL
Moreno-Espinosa, P., Abdulsalam-Alsarayreh, R. A., y Figuereo-Benítez, J. C. (2024). El Big Data y la inteligencia articial como
soluciones a la desinformación.
Doxa Comunicación
, 38, 437-451.
https://doi.org/10.31921/doxacom.n38a2029
Newman, N., Fletcher, R., Eddy, K., Robertson, C.T., & Nielsen, R.K. (2024).
Reuters Institute Digital News Report 2023.
Reuters
Institute, University of Oxford.
https://cutt.ly/6rWQSU6x
Peña-Fernández, S., Meso-Ayerdi, K., Larrondo-Ureta, A, y Díaz-Noci, J. (2023). Sin periodistas, no hay periodismo. La
dimensión social de la inteligencia articial generativa en los medios de comunicación.
Profesional de la información
,
32
(2),
e320227.
https://doi.org/10.3145/epi.2023.mar.27
Priyaa, V. G., Harrish, M. J., Udhayakumar, M., Jothieswaran, N., & Suresh, S. (12-14 abril 2024).
EcientNet-Based Deep
Learning Approach for Video Forgery Detection and Authentication.
10th International Conference on Communication and
Signal Processing (ICCSP), Melmaruvathur (India).
https://doi.org/10.1109/ICCSP60870.2024.10544113
Reinsel, J., Gantz, J., & Rydning, D., (2018).
e digitization of the world from edge to core
.
International Data Corporation.
https://cutt.ly/frWQDLdu
Rodríguez-Martelo, T., Rúas-Araújo, J., y Maroto-González, I. (2023). Innovation, digitization, and disinformation management
in European regional television stations in the Circom network.
Profesional de la información, 32
(1).
https://doi.org/10.3145/
epi.2023.ene.12
Ruo, G., Semeraro, A., Giachanou, A., & Rosso P. (2023) Studying fake news spreading, polarisation dynamics, and manipulation
by bots: A tale of networks and language.
Computer Science Review
, 47.
https://doi.org/10.1016/j.cosrev.2022.100531
Sádaba, C., y Salaverría, R. (2023). Combatir la desinformación con alfabetización mediática: análisis de las tendencias en la
Unión Europea.
Revista Latina de Comunicación Social
, 81, 17-33.
https://www.doi.org/10.4185/RLCS-2023-1552
Simanjuntak, A., Lumbantoruan, R., Sianipar, K., Gultom, R., Simaremare, M., Situmeang, S., & Panggabean, E. (2024).
Research and Analysis of IndoBERT Hyperparameter Tuning in Fake News Detection.
Jurnal Nasional Teknik Elektro Dan
Teknologi Informasi
,
13
(1), 60-67.
https://doi.org/10.22146/jnteti.v13i1.8532
Telefónica Tech (26 enero 2018).
Machine Learning contra “fake news”.
https://cutt.ly/RrWQLvLr
UNESCO (2021).
TVETipedia Glossary.
UNESCO International Centre for Technical and actional Education and Training.
https://cutt.ly/Ew2grz3A
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
doxa.comunicación |
nº 41, pp. 511-533
July-December of 2025
Mercedes Herrero de la Fuente, Celia Sancho Belinchón and Jorge Sedeño López
ISSN: 1696-019X / e-ISSN: 2386-3978
| 533
Yang, Y., Zheng, L., Zhang, J., Cui, Q., Li, Z., & Yu, P.S. (2018). TI-CNN: Convolutional neural networks for fake news detection.
arXiv
,1806.00749
.
https://doi.org/10.48550/arXiv.1806.00749
Zhang, Z., Lv, Q., Jia, X., Yun, W., Miao, G., Mao, Z., & Wu, G. (2024). GBCA: Graph Convolution Network and BERT combined
with Co-Attention for fake news detection.
Pattern Recognition Letters
,180, 26-32.
https://doi.org/10.1016/j.patrec.2024.02.014
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>
<p>[Enlace de URL / hc (has AS)]</p>