=Paper=
{{Paper
|id=Vol-3917/paper50
|storemode=property
|title=AI-agent-based system for fact-checking support using large language models
|pdfUrl=https://ceur-ws.org/Vol-3917/paper50.pdf
|volume=Vol-3917
|authors=Leonid Kupershtein,Oleksandr Zalepa,Volodymyr Sorokolit,Serhii Prokopenko
|dblpUrl=https://dblp.org/rec/conf/cs-se-sw/KupershteinZSP24
}}
==AI-agent-based system for fact-checking support using large language models==
https://ceur-ws.org/Vol-3917/paper50.pdf
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
AI-agent-based system for fact-checking support using
large language models
Leonid Kupershtein1 , Oleksandr Zalepa1 , Volodymyr Sorokolit1 and Serhii Prokopenko2
1
Vinnytsia National Technical University, 95 Khmelnytske Hwy., Vinnytsa, 21021, Ukraine
2
Simon Kuznets Kharkiv National University of Econimics, 9A Nauky Ave., Kharkiv, 61166, Ukraine
Abstract
In today’s world, the problem of disinformation is becoming increasingly relevant due to the speed of information
dissemination and the influence of social media. This article examines the impact of fake news on society and its
political, economic and social consequences. Special attention is paid to the use of large language models (LLMs)
to automate the fact-checking process. Describes the capabilities of LLMs in verifying information, including text
analysis, comparison with reliable sources, and contextualization. At the same time, the risks of using LLMs to
create fake news are highlighted. Proposes an architecture of an AI-based disinformation detection tool, which
includes query processing modules, a database, work with web resources, and results analytics. This approach is
aimed at improving the efficiency and accuracy of information verification.
Keywords
fake news, disinformation, fact-checking, large language model, AI-agent, RAG
1. Introduction
In today’s world, the flow of information is constantly growing, and with it the number of unverified
claims and myths. Social media and online platforms greatly facilitate this process, as each user can
easily create and disseminate information. In the context of such a rapid spread of disinformation,
the problem of identifying and refuting it becomes extremely important. Fake news, or deliberately
fabricated false information, is used to mislead the audience or manipulate public opinion, which can
have serious consequences for society. To combat disinformation, there is a need to develop effective
tools for fact-checking.
Big language models represent one of the most advanced advances in artificial intelligence that can
be used to automate fact-checking. These models are capable of analyzing and generating text with high
accuracy, which makes them effective tools for verifying information. Using them for fact-checking
will significantly improve the quality and speed of information verification, as well as integrate them
into existing systems.
2. The impact of fakes on society
Studies show that fake news spreads faster and more widely than the truth due to the structure of
social media, which rewards users for frequent sharing of information, regardless of its accuracy [1].
Fake news has a significant impact on society, causing a number of social, political, and economic
consequences. Their harm can manifest itself in various forms and areas of human activity.
For example, the spread of false information about vaccination has led to growing doubts and fear
among the population, which has complicated governments’ efforts to fight epidemics. Fakes can also
CS&SE@SW 2024: 7th Workshop for Young Scientists in Computer Science & Software Engineering, December 27, 2024, Kryvyi
Rih, Ukraine
" kupershtein.lm@gmail.com (L. Kupershtein); sashasun2002@gmail.com (O. Zalepa); sorokolitvovan@gmail.com
(V. Sorokolit); prokopenko.serhii@gmail.com (S. Prokopenko)
0000-0003-4712-3916 (L. Kupershtein); 0009-0008-2847-2006 (O. Zalepa); 0009-0001-9177-3054 (V. Sorokolit);
0000-0003-4712-3916 (S. Prokopenko)
© 2025 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
321
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
contribute to the spread of social conflicts, enmity and hatred by deliberately manipulating people’s
emotions [2].
In the political sphere, fakes are used to discredit opponents or manipulate public opinion before
elections. They can influence election results by distorting the electoral process. Falsification of
information creates the illusion of distrust in decision-making processes and the government in general,
which can lead to political instability [3].
Fakes also have the potential to affect the economy. For example, false reports on the financial
condition of companies can seriously affect their stock prices and investment attractiveness. Such
information attacks can lead to significant financial losses not only for investors, but also for ordinary
consumers who may lose their jobs due to market instability. False information on the Internet about
companies can cause significant fluctuations in the market, which negatively affects economic stability
[4].
Disinformation is becoming an increasingly common problem in today’s information environment,
especially in the context of international politics. Ukraine is often one of the main targets of such
attacks. According to the second report of the European External Action Service ‘Manipulation of
Foreign Information and the Threat of Interference’, a significant number of disinformation cases were
analyzed between December 2022 and November 2023 (table 1) [5].
Table 1
Analysis of recorded disinformation incidents.
Indicator Number of cases Share (%)
Incidents against Ukraine 160 21.3%
Incidents against the United States 58 7.7%
Incidents against Poland 33 4.4%
Incidents against Germany 25 4.1%
Incidents against France 25 3.3%
Attacks against the EU 142 19%
Attacks on NATO 113 15%
Attacks on the Armed Forces of Ukraine 105 14%
Attacks on the UN 22 3%
One of the most high-profile examples of disinformation that has had a significant impact on interna-
tional politics and relations is the spread of fakes about events in Ukraine by Russia. Two of the most
significant fakes were published in an article by the European Commission [6].
The first fake is that Ukraine and its government are the aggressor in relations with Russia. This
false claim was used to justify Russia’s military aggression against Ukraine, including the annexation of
Crimea in 2014. Russian disinformation was aimed at convincing the international community and its
citizens that Ukraine was allegedly provoking the conflict and posing a threat to Russia. However, the
reality was that it was Russia that invaded Ukraine, violating its sovereignty and territorial integrity, as
confirmed by numerous international agreements and legal documents.
The second fake is the genocide of the Russian-speaking population in Ukraine. This propaganda
thesis has been used to justify Russia’s criminal military actions in Ukraine, to strengthen support from
the Russian population, and to create a negative image of Ukraine in the international arena. However,
numerous international organizations, including the Council of Europe, have found no evidence of
systematic violations of the rights of the Russian-speaking population in Ukraine. This myth has been
dispelled as part of Russia’s disinformation campaign aimed at destabilizing the situation in Ukraine
and justifying its aggression.
These examples of disinformation have serious consequences not only for the domestic political
situation in Ukraine, but also for international relations, as they contribute to the escalation of the
conflict and influence international support and actions of other states towards the conflict. Distortion
of facts and manipulation of information during military conflicts are among the most harmful aspects
of modern information warfare.
322
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Fact-checking is a critical process, as the speed of information dissemination can lead to the massive
spread of disinformation. This process involves the use of various techniques and approaches that have
evolved over time from traditional verification methods to modern technological solutions [7].
Initially, fact-checking depended mainly on manual verification of information, which required
significant time and effort. Traditional methods included verification of primary sources of information,
such as official documents, scientific studies or reports of independent organizations, cross-checking
information from different sources, where data is confirmed through several independent and reliable
resources, expert opinions from industry experts to analyze complex, or specialized topics that require
deep knowledge in a particular area [8].
3. Possibilities of large language models for fact-checking
Large language models (LLMs) are one of the most advanced advances in artificial intelligence aimed at
understanding and generating human speech. They are based on deep neural networks and use large
amounts of text data for training, which allows them to perform complex language tasks with high
accuracy. The basic concept of LLMs is to train the model to predict the next word in a sentence based
on the context of the previous words.
LLMs can perform a wide range of tasks, namely [9]:
• generate text, which can be useful for writing articles, creating content, or even writing code;
• translate text from one language into another, taking into account the context and peculiarities
of each language
• determine the emotional tone of the text, which can be useful for analyzing reviews, comments
and other textual data;
• check information for accuracy, thanks to their ability to process large amounts of text and
compare information from different sources.
LLMs have significant potential for use in fact-checking due to their ability to process and analyze
large amounts of textual data. They can perform several key tasks that make them effective tools for
verifying the accuracy of information [10]:
• quickly analyze textual materials, identifying potentially questionable or false statements;
• compare statements with available reliable sources of information;
• take into account the context in which the statement was made, which helps to avoid out-of-
context quotes and manipulations;
• automate a part of the verification process, which significantly reduces the time and effort required
to identify disinformation.
However, LLMs can be used not only for fact-checking, but also for creating fakes. They are capable
of generating seemingly plausible text, which makes it quite easy to create articles, news and social
media posts that can mislead readers. Such models can automatically formulate answers to questions
by changing only key elements to create fake news that looks authentic. This process may include
replacing facts or details in the original texts to form a new article containing disinformation [11].
For example, one way to create fake news is to first ask a question about a real news story, get an
answer, and then modify that answer to create a new article. This article is then checked to see if the
new answer matches the original question. If it differs significantly, the generated article is saved as fake
news. This method allows you to create text that looks plausible but contains inaccurate information
that can mislead readers [12].
Thus, while LLMs can be powerful tools for combating disinformation, they can also be used to create
it, which underscores the importance of developing mechanisms for detecting fake news to ensure the
information security of society.
323
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Despite the potential harm, LLMs can also be used to detect text generated by different language
models, based on stylistic and structural features of automated generation, which helps to identify fakes.
They can be configured to identify texts that have such features. For example, language models often
demonstrate certain patterns in word usage, parts of speech, and emotional tone, which can be detected
by analyzing the syntactic and semantic features of a text. Studies show that machine-generated text
has less vocabulary variation and is less emotionally charged than text written by humans [13].
Such methods include the use of pre-trained models that can distinguish between human-generated
and LLM texts by analyzing word frequency, sentence length, syntactic structure, and other characteris-
tics. Using these features, models can be trained on data containing both human and generated texts
[14].
Several leading LLMs can be used for effective fact-checking, and they are highly effective due to
their powerful algorithms and ability to process large amounts of textual data. A comparison of the
technical characteristics and capabilities of the most advanced LLMs is shown in table 2 [15].
Table 2
Technical characteristics and capabilities of modern LLMs.
Presenting an
Context Search for information Connect Support Listing of
LLM answer in a
window size on the Internet files for RAG found sources
given form
GPT-3.5 16k tokens No Yes Limited No Yes
GPT 4o 128k tokens Yes Yes Yes Yes Yes
Llama 3.1 128k tokens No Yes No No No
Claude 3 200k No Yes Yes Yes Yes
Mistral 4k tokens No Yes Yes Yes Yes
Gemini 1.5 1 million tokens Yes Yes Yes Yes Yes
4. Architecture of AI-agent-based system for fact-checking
The architecture of an AI-based fact-checking system consists of several main modules that interact with
each other to ensure efficient operation. Each module performs a specific function, and their coordinated
interaction ensures high accuracy and speed of information verification. The main components of the
tool’s architecture are:
• user interface (UI);
• user request processing module;
• database;
• module for working with web resources;
• LLM;
• results analysis module.
Scheme of work an AI agent-based system to support fact-checking is shown in figure 1.
The interaction scheme of these components is shown in figure 1. The user query processing module
is responsible for receiving queries from users through the interface, pre-processing them and passing
them to the large language model for analysis. The module also provides feedback by returning the
results of the check to users. Thanks to this module, the tool can quickly respond to queries, ensuring
high accuracy and efficiency of fact-checking.
The database is the main repository of reliable sources of information or already verified facts used
for verification. It contains a large amount of verified data that is constantly updated to ensure that the
information is up-to-date. The vector database with a large language model implements the so-called
RAG system. It allows quickly find and compare facts with existing ones, which ensures high accuracy
and reliability of the results of checking those news items that have already been checked.
324
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Figure 1: Architecture of AI-powered system for fact-checking support.
The web resources module is responsible for integrating the tool with external information sources.
It makes queries to various web resources, extracting the necessary data for analysis and comparison if
the data is not in the existing database. This module ensures that the application’s checks are up-to-date,
allowing it to use the latest data from the Internet. It also interacts with other modules of the system,
transferring the received data for further processing and analysis.
LLMs are the key component of the application responsible for analyzing textual data and checking
its validity. They use deep neural networks to understand and generate human speech, allowing them
to effectively detect potentially false statements. LLMs will be used to parse the user’s query and search
for relevant information in the database. If the relevant data is not found in the database, the model
converts the user’s query into a search query for searching the Internet. After receiving the results, LLM
analyzes extracted data and makes a decision on its reliability, determining whether the information is
fake. This allows you to quickly and accurately verify a large amount of information, reducing the time
and resources required for manual verification.
The results analysis module processes data obtained from a large language model and web resources,
generates reports and provides recommendations on the reliability of information. It analyses the results
of the check, determines the level of data reliability, and draws conclusions that are then provided to the
user. The module can also detect patterns and trends in the data, which helps to identify new sources of
disinformation and develop strategies to counter them. The module provides users with accurate and
detailed verification results.
5. Development technologies and tools
To develop the fact-checking support system, we used modern tools that ensure efficiency, reliability
and scalability. The main component is OpenAI’s GPT-4o large language model, which is integrated
through the OpenAI API. This choice was made due to GPT-4o’s ability to quickly analyze large amounts
of text, perform comparisons, take into account context, and provide accurate conclusions.
The system was implemented in the Python programming language, which provides flexibility and
325
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
a wide range of libraries for text processing, process automation, and neural networks. To process
queries and retrieve relevant data from external sources, we used the google search and «serpapi»
libraries. They provide integration with search engines, allowing us to collect relevant information
for verification. The «htmldate» library was used to automatically determine the dates of webpage
publications, which guarantees accurate extraction of chronological information for analysis.
The user interface was developed with the help of the Streamlit library, which allows creating
interactive web applications without the need to use HTML, CSS or JavaScript. This greatly accelerated
the implementation of a user-friendly interface for user interaction with the system.
Combining these tools with Python capabilities made it possible to implement the functionality of
automatic text data verification, real-time fact analysis, and integration with other analytical systems.
This ensured the creation of an efficient and reliable fact-checking system capable of meeting modern
requirements.
6. Experiments
This section presents the results of experimental studies of a fake news detection software tool based
on an AI agent system using large language model. The aim of the study is to evaluate the functionality,
performance, and reliability of the developed system. The experiments were conducted in three stages:
user interface research, testing of interaction with the database, and evaluation of information search
on the Internet.
The first stage was to study the user interface to assess its intuitiveness, usability and correctness of
interaction with the user. Figure 2 shows a view of the user interface.
Figure 2: User interface design.
We checked the possibility of entering user queries and the correctness of the system’s processing of
the entered data. Figure 3 shows an example of a query used for this test. The results showed that the
interface is intuitive and provides correct interaction with the user.
Figure 3: Test query view for checking the work user interface.
The second stage was to test the system’s interaction with the verified facts database. The purpose
326
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
of this study was to make sure that the tool interacts correctly with the database and retrieves the
necessary information. Figure 4 shows a fragment of the database.
Figure 4: View of the DB fragment.
A test was conducted with an initial query that matches the database record exactly. The results
showed that the system successfully finds the relevant information. Next, the query was modified to be
different from the database records (figure 5).
Figure 5: View of the modified query for testing work with DB.
The system still successfully identified the relevant information, demonstrating the ability to under-
stand the context of the query and find the necessary data even with the changed wording.
The third stage involved testing the functionality of information search on the Internet. This allowed
us to assess whether the program correctly forms queries to search engines, whether it can find relevant
information and analyze it correctly.
First request: Verification of information about the introduction of electricity blackouts in Ukraine
lasting up to 20 hours a day starting June 5, 2024. The results (figure 6) showed that the program
correctly refuted this statement, providing a detailed conclusion with links to reliable sources.
Figure 6: The first test query to check how the web search works.
Second query: Checking the information that English has become the second official language in
Ukraine. The program refuted this claim, noting that English was granted the status of a language of
international communication, but was not recognized as the state language (figure 7).
327
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Figure 7: The second test query to check how the web search works.
Third request: Verification of information about the Kyiv Regional CCC’s drawing of mobilization
armor. Initially, the program gave a false answer, confirming this statement (figure 8). After a second
request, the program gave the correct answer, refuting the fake and providing links to reliable sources
(figure 8).
Figure 8: The third test query to check how the web search works.
The information checked in the test queries was also refuted by the fact-checking organization Gvara
Media (figure 9). For example, they confirmed that the Kyiv Regional Military Commissariat does not
sell reservations for mobilization, emphasizing that this information is fake.
Figure 9: Refutation of the fake about the armor raffle from the mobilization by the fact-checking organization
Gvara Media.
Figure 10 shows an erroneous result of the tool. The sources were found to be correct, but LLM gave
an erroneous analysis result. This once again emphasizes the need for human control over the work of
language models, as they can experience so-called “hallucinations”.
After trying to make this test query a second time, the program gave the correct answer (figure 11).
328
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Figure 10: Incorrect response from the program.
Figure 11: The correct answer of the program, after the second attempt of verification.
The results of the accuracy assessment show that the fake news detection tool has a high accuracy of
90%. The first and second types of errors amount to 10%, which indicates a relatively low level of false
positives. The program’s errors occurred due to the language model’s “hallucinations” and because
it sometimes received not the best sources for analysis. In general, the system demonstrates reliable
performance, but there is still room for further improvement in detecting both true and false statements.
Testing has shown that the fake detection tool works correctly, meets the specified requirements,
and is ready to be used in real-world conditions.
7. Conclusion
In this article, we develop and present a tool for automating fact-checking based on the use of large
language model. The proposed system includes basic modules such as a user interface, a query processing
module, a database of verified facts, a module for working with web resources, and a module for analyzing
results.
The analysis of existing methods and tools for fact-checking has revealed the main problems and
limitations of traditional approaches, which are often not effective enough in the modern information
environment. The integration of LLM proved to be a promising solution and improved the quality and
speed of information verification by automating the analysis of large amounts of text and searching for
reliable sources.
Testing of the system on real data has confirmed its efficiency and flexibility in handling different
329
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
types of queries. The tool demonstrates a good ability to quickly adapt to different types of queries and
provide accurate answers based on the analysis of a large number of sources. This demonstrates their
versatility and ease of use for different categories of users.
Declaration on Generative AI: The authors have not employed any Generative AI tools.
References
[1] D.-M. Ordway, Fake news and the spread of misinformation: a research
roundup, 2017. URL: https://journalistsresource.org/politics-and-government/
fake-news-conspiracy-theories-journalism-research/.
[2] University of Tokyo, New analysis shows anti-vaccination conspiracy theories
gain political weight due to social media, 2024. URL: https://phys.org/news/
2024-02-analysis-anti-vaccination-conspiracy-theories.html.
[3] The real impact of fake news: The rise of political misinformation–and
how we can combat its influence, 2024. URL: https://sps.columbia.edu/news/
real-impact-fake-news-rise-political-misinformation-and-how-we-can-combat-its-influence.
[4] M. Niessner, Does Fake News Sway Financial Markets?, 2018. URL: https://insights.som.yale.edu/
insights/does-fake-news-sway-financial-markets.
[5] N. Lototska, Ukraina naichastishe ye zhertvoiu dezinformatsii ta feikiv, – zvit yevropeiskoi sluzhby
zovnishnikh sprav, 2024. URL: https://lb.ua/society/2024/01/25/595364_ukraina_naychastishe_ie_
zhertvoyu.html.
[6] Directorate-General for Neighbourhood and Enlargement Negotiations, Dis-
information about the Current Russia-Ukraine Conflict – Seven Myths De-
bunked, 2022. URL: https://neighbourhood-enlargement.ec.europa.eu/news/
disinformation-about-current-russia-ukraine-conflict-seven-myths-debunked-2022-01-24_en.
[7] Y. Baryshev, L. Kupershtein, V. Maidanovych, O. Voitovych, S. Prokopenko, Information System
for the Fact-checker Support, in: A. Anisimov, V. Snytyuk, A. Chris, A. Pester, F. Mallet, H. Tanaka,
I. Krak, K. Henke, O. Chertov, O. Marchenko, S. Bozóki, V. Tsyganok, V. Vovk (Eds.), Selected
Papers of the X International Scientific Conference “Information Technology and Implementation”
(IT&I-2023). Workshop Proceedings, Kyiv, Ukraine, November 20-21, 2023, volume 3646 of CEUR
Workshop Proceedings, CEUR-WS.org, 2023, pp. 127–138. URL: https://ceur-ws.org/Vol-3646/Paper_
13.pdf.
[8] Z. Guo, M. Schlichtkrull, A. Vlachos, A Survey on Automated Fact-Checking, Transactions of the
Association for Computational Linguistics 10 (2022) 178–206. doi:10.1162/tacl_a_00454.
[9] L. Ardaiev, LLM: shcho tse take i yaki vidkryvaie mozhlyvosti, 2023. URL: https://aw.club/global/
uk/blog/what-are-llms-and-what-opportunities-do-they-offer.
[10] N. Van Otten, Fact-Checking With Large Language Models (LLMs): Is It A Powerful NLP Verifica-
tion Tool?, 2024. URL: https://spotintelligence.com/2024/02/26/fact-checking-verification-nlp/.
[11] Y. Sun, J. He, L. Cui, S. Lei, C.-T. Lu, Exploring the Deceptive Power of LLM-Generated Fake
News: A Study of Real-World Detection Challenges, 2024. URL: https://arxiv.org/abs/2403.18249.
arXiv:2403.18249.
[12] J. Su, C. Cardie, P. Nakov, Adapting Fake News Detection to the Era of Large Language Models,
in: K. Duh, H. Gomez, S. Bethard (Eds.), Findings of the Association for Computational Linguistics:
NAACL 2024, Association for Computational Linguistics, Mexico City, Mexico, 2024, pp. 1473–1490.
doi:10.18653/v1/2024.findings-naacl.95.
[13] S. Abdali, R. Anarfi, C. Barberan, J. He, Decoding the AI Pen: Techniques and Challenges in
Detecting AI-Generated Text, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge
Discovery and Data Mining, KDD ’24, Association for Computing Machinery, New York, NY, USA,
2024, p. 6428–6436. doi:10.1145/3637528.3671463.
[14] M. Bethany, B. Wherry, E. Bethany, N. Vishwamitra, A. Rios, P. Najafirad, Deciphering Textual
Authenticity: A Generalized Strategy through the Lens of Large Language Semantics for Detecting
330
Leonid Kupershtein et al. CEUR Workshop Proceedings 321–331
Human vs. Machine-Generated Text, in: 33rd USENIX Security Symposium (USENIX Security
24), USENIX Association, Philadelphia, PA, 2024, pp. 5805–5822. URL: https://www.usenix.org/
conference/usenixsecurity24/presentation/bethany.
[15] S. Kazi, A. Elmahdy, Top Large Language Models (LLMs): GPT-4, LLaMA
2, Mistral 7B, ChatGPT, and More, 2023. URL: https://vectara.com/blog/
top-large-language-models-llms-gpt-4-llama-gato-bloom-and-when-to-choose-one-over-the-other/.
331