The Challenges and Limitations of Artificial Intelligence as a Tool for Supporting Media Literacy

11/06/2026

The Challenges and Limitations of Artificial Intelligence as a Tool for Supporting Media Literacy

Introduction: AI and Media Literacy in the Age of Disinformation

There is no doubt that, although the dissemination of false news with the aim of shaping public opinion is not a new phenomenon, throughout this century—and especially during the last decade—it has become something entirely different. The widespread use of social media, including as a source of information, has led to the emergence of a hybrid media system (Chadwick, 2017) characterized by echo chambers (Liu et al., 2025). Combined with traditional media outlets experiencing increasing levels of declining legitimacy (Laviana, 2024; Soengas Pérez et al., 2023), this has produced qualitative changes in opinion-formation processes and, more broadly, in the way information dynamics flow (Wahlström & Törnberg, 2021).

The emergence of Artificial Intelligence (AI) in recent years has exponentially intensified this problem. For those unfamiliar with its foundations and operation, this dual nature of AI—as both an amplifier of the very problems it seeks to help solve—enhances its perceived “magical” character. Therefore, this chapter will provide an explanation focused on the technical characteristics of tools used for detecting disinformation. Finally, some of the limitations and challenges faced by AI tools, as well as the role of digital literacy in this task, will be outlined.

2. The Technical Dimension of AI: Possibilities and Limitations

When mainstream media refer to AI today, they are usually referring to a very specific subfield: generative models, particularly those based on neural networks  The goal of a generative model is to produce new samples of a given type of data—text, images, videos, audio, and so on. Neural networks, in turn, are algorithms composed of a set of adjustable parameters that determine their behavior. We can imagine each parameter as the tuning dial of a radio that must be adjusted until it correctly captures a signal. In the machine learning paradigm, instead of adjusting these parameters manually, data are used so that the algorithm itself can optimize them automatically. This process is known as training, and once completed, the resulting algorithm is called a model. Until a few years ago, such models were developed for a specific task—such as generating images of faces, detecting spam emails, or transcribing a vehicle’s license plate from a photograph—and it was necessary to train the model from scratch. This involved collecting a representative dataset, optimizing the model’s parameters, and, after a lengthy process of trial and error, eventually achieving sufficient performance for real-world use.

This paradigm changed with the emergence of foundational models. The main difference lies in the scale and purpose of training  While traditional models were designed to solve specific tasks, foundational models are trained on enormous amounts of data from multiple domains without focusing on a single task. In this approach, quantity became more important than quality, incorporating text from Wikipedia, books, forums, news articles, and virtually any textual content available on the Internet. From this line of research emerged the models we are familiar with today—ChatGPT, Gemini, Claude, DeepSeek, among others. —.Thanks to their development, it is no longer necessary to gather data and train a model from scratch for every task, since general-purpose tools now exist that can perform a wide variety of functions simply by describing the task in natural language, without requiring advanced technical knowledge. This paradigm shift also facilitates the creation of disinformation: it is no longer necessary to master a language in order to produce an article that appears legitimate (Shah et al., 2025); with only a few audio recordings, a voice can be cloned and made to say anything (Genelza, 2024); and with as few as ten carefully selected images of a person, it is possible to train models capable of generating realistic images of that individual in entirely fabricated contexts. The same applies to video and audio generation (Kaur et al., 2024; Loth et al., 2024).

Nevertheless, these same models can also help combat disinformation more effectively when their characteristics are properly understood. These models, based on the Transformer architecture, are trained to generate coherent text, not necessarily truthful text. Therefore, even when they are not used maliciously, their outputs cannot be fully trusted because their responses are often syntactically coherent but invented—a phenomenon known as hallucination (Li, 2023).

As a result, training models to determine whether a statement is true or false is an extremely complex task, and it is even less likely that such an ability would emerge naturally from training on random internet data. In general, it is not advisable to allow any AI model, regardless of its type, to independently decide whether information is accurate or not.

To leverage AI’s capabilities in disinformation detection, the problem must be approached from a different perspective. While determining the truthfulness of a text is beyond the reach of these models due to their architecture and training, those same characteristics enable them to identify whether two pieces of information convey exactly the same meaning. Because of their mastery of language, AI models can compare information against databases of verified hoaxes and misinformation maintained by fact-checking organizations and determine whether a piece of content matches or resembles a known falsehood.

In this way, it is possible to combine the strengths of both humans and machines: information professionals handle the tasks that machines cannot perform—determining whether information is true or false—while machines take on tasks that humans cannot realistically manage—processing the enormous volumes of data generated on the internet.

This is just one example among many possible applications. AI can also be used to identify which pieces of information are suitable for verification and which are not, organize similar content into coherent groups to identify trending topics, or uncover relationships among malicious actors coordinating disinformation campaigns, among many other uses.

Finally, it is important to emphasize that whenever a model of this kind is deployed, a cat-and-mouse game begins. Malicious actors attempt to evade detection systems, while defenders seek to identify and stop them. For example, leet speech(which consists of replacing letters with symbols or numbers, such as using “3” instead of “e”) is currently employed to confuse systems and bypass automated controls. This drives the development of new models capable of recognizing these emerging strategies, while malicious actors continue searching for new ways to circumvent them. Consequently, these systems must be continuously studied, monitored, and updated to minimize response times and maximize effectiveness against malicious uses.

3. Conclusion: Challenges and Limitations of AI

The importance of detecting hoaxes and disinformation in today’s societies cannot be underestimated. AI can not only assist information professionals in performing their work more effectively, but it can also help citizens become better informed.

Disinformation often employs certain psychological tricks to make its message resonate more effectively with the public and spread more rapidly. For example, it uses clickbait—sensationalist headlines that have little or nothing to do with the actual content of the news story—to capture attention. It also appeals to emotions, seeking to provoke anger, alarm, or fear so that people act less rationally. In addition, it relies on logical fallacies such as the straw man fallacy (distorting or exaggerating another person’s argument in order to refute a weaker or false version of it) or anecdotal generalization (using a personal experience or an isolated case as sufficient evidence to support a general claim). AI models have the ability to detect some of these practices and can both warn users and teach them how to recognize them in the future, making them a key component of media literacy among the population. 

To achieve this objective, it is essential, on the one hand, to treat the information systems used by the public as critical national infrastructure requiring constant monitoring to prevent attacks by malicious actors, in the same way that power plants, refineries, or nuclear facilities are protected. On the other hand, it is equally crucial to educate citizens in critical thinking skills, because without a knowledgeable and critical citizenry, it is not possible to develop or deploy these types of systems.

References

Chadwick, A. (2017). The Hybrid Media System: Politics and Power (2.a ed.). Oxford University PressNew York. https://doi.org/10.1093/oso/9780190696726.001.0001 

Genelza, G. G. (2024). A systematic literature review on AI voice cloning generator: A game-changer or a threat? Journal of Emerging Technologies, 4(2), 54-61. https://doi.org/10.57040/ag587791 

Kaur, A., Noori Hoshyar, A., Saikrishna, V., Firmin, S., & Xia, F. (2024). Deepfake video detection: Challenges and opportunities. Artificial Intelligence Review, 57(6), 159. https://doi.org/10.1007/s10462-024-10810-6 

Laviana, J. C. (2024). Recuperar la confianza perdida, el gran desafío de los medios. Cuadernos de Periodistas, 85-92. 

Li, Z. (2023). The Dark Side of ChatGPT: Legal and Ethical Challenges from Stochastic Parrots and Hallucination (No. arXiv:2304.14347). arXiv. https://doi.org/10.48550/arXiv.2304.14347 

Liu, J., Schwarz, A., Risius, M., Hirschheim, R., & Scotter, J. V. (2025). Conceptualizing Echo Chambers and Information Cocoons: A Literature Review and Synthesis of Current Knowledge and Future Directions. The Journal of Strategic Information Systems, 34(2), 101904. https://doi.org/10.1016/j.jsis.2025.101904 

Loth, A., Kappes, M., & Pahl, M.-O. (2024). Blessing or curse? A survey on the Impact of Generative AI on Fake News (No. arXiv:2404.03021). arXiv. https://doi.org/10.48550/arXiv.2404.03021 

Shah, S. B., Thapa, S., Acharya, A., Rauniyar, K., Poudel, S., Jain, S., Masood, A., & Naseem, U. (2025). Navigating the Web of Disinformation and Misinformation: Large Language Models as Double-Edged Swords. IEEE Access, 13, 169262-169282. https://doi.org/10.1109/ACCESS.2024.3406644 

Soengas Pérez, X., Rodríguez Castro, M., & Campos Freire, F. (2023). La credibilidad de los informativos de la televisión pública en España. Comunicar: Revista Científica de Comunicación y Educación, 76, 73-84. https://dialnet.unirioja.es/servlet/articulo?codigo=8947688 

Wahlström, M., & Törnberg, A. (2021). Social Media Mechanisms for Right-Wing Political Violence in the 21st Century: Discursive Opportunities, Group Dynamics, and Co-Ordination. Terrorism and Political Violence, 33(4), 766-787. https://doi.org/10.1080/09546553.2019.1586676 

Ángel Panizo Lledot is an Assistant Professor at the School of Computer Systems Engineering (ETSISI) of the Technical University of Madrid (UPM). He holds a Bachelor’s degree in Computer Science from the Complutense University of Madrid, a Master’s degree in Artificial Intelligence from the Technical University of Madrid, and a PhD in Computer Science from the Autonomous University of Madrid. He is currently a member of the AIDA research group at ETSISI-UPM. His main research interests focus on clustering, graph-based algorithms—particularly those involving temporal graph evolution—social data analysis, and bio-inspired methods. 

Sergio is a sociologist and holds a PhD in Philosophy. He is currently a professor at the Technical University of Madrid and a member of the NLP-DL research group. His main research areas include Science and Technology Studies (STS), the interrelationships between information technologies and the social sciences, and Social Network Analysis. Among other initiatives, he has participated in several national and European projects related to disinformation and its consequences.

Leer más