The press is commonly referred to as the fourth estate. The term can be traced back to the traditional division of state power into three separate branches. This designation refers to the significant role that the media plays in the functioning of modern states. At first glance, it may seem peculiar to equate it with the executive, legislative, and judicial powers, but it is worth considering that except for the least developed countries, most citizens everywhere obtain their information about politics and state affairs from the press in one way or another. Considering this, it is easy to see the crucial role that the media plays in the functioning of the state, as it determines when, how, and what information people can receive. Consequently, accurate and fact-based reporting is in everyone’s interest. More…

Measuring media bias has long been an unresolved problem in quantitative research. Given the highly politically nuanced nature of the problem, it is worth briefly reviewing the main research directions, problems, and possible solutions to the issue of how bias can be measured. What do we call bias? How can we determine whether a given medium is biased or objective? These are difficult questions since determining bias is inherently subjective, but in this post, we will try to go around the topic and give you some ideas. 

In our highly mediated world, it is more important than ever before to get an accurate picture of the reliability of our news sources, not only for us but also for researchers in the field. Measuring media bias is a decisive area of research in many respects. It helps us understand the fault lines present in our societies, the impact of economic interests on information flow, and the state of our democracies. However, measuring bias is far from trivial. In the following, we will review the challenges researchers face on the subject and the methods they use to measure media bias. This not only gives us insight into the activities of researchers but also helps us better understand the characteristics of bias within the media.

If we want to assess the bias of individual media outlets in any way, we must face two significant challenges, which can be called subjectivity and the unobservable population (Groeling, 2013). The problem of subjectivity is relatively simple. It refers to the fact that different news consumers will have different opinions about whether a particular article or report is biased, or even which direction it is biased in. One reason for this is that individuals themselves have preferences and preconceptions. They have ideological and political preferences, which may make them more inclined to view news as biased against their own positions. This is particularly common in cases where politicians refer to a hostile media environment; it is common for both sides of a debate to simultaneously claim that the press is biased and hostile towards them.

Additionally, news consumers also have preconceptions about individual outlets. Whether a piece of news is biased and in what sense it is biased can be judged differently depending on the source, as individuals already have some idea before consuming the news about whether the source is biased. To avoid this problem, researchers use tools that do not require subjective, individual judgment.

The second significant problem is the unobservable population. First, here is a little context: in statistics, we distinguish between the population and the sample when conducting a survey. The population is the group from which we want to learn something. From it,  we should take a sample. For instance, if our population is the voting-age citizens of a country, then our sample would be the people we contacted and surveyed during our research. Media outlets operate similarly. Their population is all the news they receive, much larger than the information they report. This is a problem for researchers because they only encounter the sample, i.e., the information that made it into the news; they can never see the whole population. Consequently, they cannot know whether a particular media outlet was biased in any way in selecting the news. Often, when accused of bias, media outlets refer to only providing unbiased information about relevant news and cannot control whether these are favorable or unfavorable to anyone. Without the invisible population, the only reference point available to researchers is the reporting of other media outlets. Therefore, studies often do not attempt to explore bias in absolute terms but only comparatively examine it by comparing two or more different news sources.

When defining media bias, we discussed how bias can appear in diverse ways. Similarly, the methods of measuring it are also eclectic, as we will see. Most studies deal with print media because it has been the dominant news source in the past, and it is relatively easier to obtain the necessary data for analysis than television or radio broadcasts. Another characteristic of these studies is that they primarily focus on the United States, but to a lesser extent, Western Europe is also overrepresented.

We have detailed the difficulties researchers face so far; to nuance this, let’s start with the simplest method possible. In the USA, it was common for some media outlets to explicitly declare which party they supported, but this is no longer typical today; only analyses of the past rely on such explicit endorsements (D. Ho & Quinn, 2008). A modern equivalent is when the editorial staff of individual media outlets takes a stand on specific issues (Ansolabehere, 2006; D. E. Ho & Quinn, 2007; Puglisi & Snyder, 2015). This method based on “self-declaration” not only involves less work but is as objective as measuring bias can be, and it does not require the relative comparison of outlets. It allows us to assess how much of the press segment takes a stand in one way or another. The disadvantages are that we do not get information about the extent of bias, and only statements related to parties or politicians can be examined this way.

Another approach examines the sources cited by the media outlet, such as opinion pollsters, who also have biases. If we consider this bias known, then we can infer the bias of individual news sources based on how often they refer to different opinion pollsters; if they refer more often to those with certain biases, then they are likely biased (Groeling, 2008). In this example, we talked about opinion pollsters, but research of this nature has also been conducted with think tanks and other opinion leaders (Gans & Leigh, 2012; Groseclose & Milyo, 2005). It has also been examined what kind of images are displayed of certain politicians, although a challenge here is to judge which images are favorable and which are unfavorable (Hehman et al., 2012). In television broadcasts, an important question is how often certain politicians appear, and it is a significant advantage if a politician can convey their position directly to viewers in their own words.

Finally, the most innovative studies use a method called text mining. The essence of this approach is to convert texts into numerical data (using complex procedures not worth detailing here) for subsequent measurement. With these procedures, we can examine, for example, the sentiment of texts, whether they are positive or negative (Niven, 2003; Schiffer, 2006). While this approach has the allure of novelty, it still does not provide a perfect solution. Take sentiment analysis, for example; if a party typically appears in negative-toned articles, it does not necessarily mean that the negative tone refers to the party itself. Of course, with a sufficiently large sample, the correlation exists, but it is not perfect, so the result obtained also contains some uncertainty. There are more sophisticated methods than this, and sentiment analysis itself can be executed in a thousand different ways (which greatly affects the accuracy of the measurement), but there is no perfect text-mining procedure.

From the number of approaches mentioned, it can be seen that media bias takes many forms, just as researchers are forced to employ various measurement methods, but none of them can be perfect. There is no objective methodology that provides an accurate and objective measurement of bias for articles, news reports, or media outlets. These investigations can provide snapshots and point out correlations, causes, and potential effects of bias. The reason for this limitation is primarily that the subject of the investigation is constantly changing and evolving, and to select the appropriate approach, the researcher must have a thorough understanding of the social and political context of the media being studied.

In earlier blog posts, we detailed how, in today’s mediatised societies, the media not only informs but also often shapes public opinion and, as such, is a crucial platform for every political actor. Since we get information about governmental activities through the media, influencing the media can be an attractive opportunity for any government. In this context, the concept of media capture arises, which is one particularly problematic manifestation of media bias. As the term suggests, media capture refers to external influence on the media; it occurs when a particular government influences media outlets, thereby manipulating the content and nature of their reporting. Such a practice raises numerous ethical and democratic concerns as it threatens media independence and pluralism, which are fundamental to democratic societies.

Media capture occurs when a particular media outlet or outlets shape their reporting based on government incentives or pressure. However, for capture to occur, the government must have tools to influence individual media outlets rather than just the entire media market. Moving forward, we will go through various manifestations and types of media bias, categorised based on the means of exerting influence. When examining media capture in scholarly literature, it is always a prerequisite to demonstrate that a particular media outlet genuinely distorts its reporting due to government intervention.

One of the most effective tools available to any government is the budget. The most practical way to selectively allocate state resources is by directing advertising spending only to certain “privileged” media outlets. In a Hungarian analysis, Ádám Szeidl and Ferenc Szűcs examined the correlation between state advertising and the content of media coverage (Szeidl & Szucs, 2021). Their analysis concluded that state advertisements significantly influence news coverage; media outlets receiving them were less likely to report on scandals examined during the research. In a previous analysis, Attila Bátorfy and Ágnes Urbán detailed how successive Hungarian governments used state advertising to shape the media environment (Bátorfy & Urbán, 2020). State advertisements influence media outlets in two ways. Firstly, they provide a financial incentive to modify their reporting. Secondly, they make it more profitable in the long term for entities aligned with the ruling party or parties to operate media outlets, thus exerting a longer-term distorting effect.

The case of public media deserves special attention. On the one hand, through positive examples, we can see that it can provide impartial and high-quality news coverage. However, all too often, public media functions as a propaganda tool for the government. This is particularly problematic considering that in most countries, the largest press organisations are some form of public media (Djankov et al., 2003). The case of Poland is illustrative in terms of public media and media capture. While the Law and Justice Party (PiS) was in power, criticism was regularly directed at the public media’s coverage (Kerpel, 2017). Subsequently, after a multi-party coalition led by the Civic Coalition (KO) came to power, significant changes in the leadership of state media led to prolonged conflicts (Politico, 2023). The situation with Hungarian public media is similar; in 2020, a leaked recording revealed that Balázs Bende explicitly instructed several public media employees that it was their job to strengthen and represent the government’s position before the European Parliament elections in March.

“Bende Balázs: The situation is as follows: everyone is aware that at the end of May, there are European parliamentary elections, and I am sure that no one will be surprised if I say that in this institution… we do not support the opposition coalition. If this statement surprises anyone, they can leave now. For those not surprised, it is not surprising that if the institution supports it, we all work accordingly. There’s no question. There will be no question about what material I want in the future. There will be no question about how to write that material. Those who don’t know can also go home and need not return.”

(Telex, 2020)

Bribery, as a tool, initially appears significantly similar to state advertising; a government expects certain types of coverage in exchange for financial compensation. However, two crucial differences distinguish the two cases. Firstly, bribery is not legal, unlike advertising. Secondly, the first point leads to the second: state advertising is documented, allowing citizens at least the opportunity to understand how the government influences the media, thus making the government more accountable to unoccupied press organs and civil organisations.

In countries where judicial institutions have partially or entirely fallen under government control, courts often turn against journalists and media outlets to hinder their operations. Finally, it is important to mention the most radical tool of media capture: violence against journalists, which is also frequently used for intimidation purposes and to exert control over individual media outlets.

Thus far, we have examined what media capture is and provided examples of its various manifestations and types. It is important to bear in mind that the concept of media capture is an ideal type; it generally cannot be fully matched with real cases. Governments cannot exert absolute control, and there are no governments that do not influence their countries’ media environment. Reality typically lies somewhere between the two extremes. As seen in the tools of media capture listed, examples escalate in terms of their consequences and violations of democratic norms. Generally, it can be said that the worse a country’s democratic institutions’ checks and balances function, the greater the impact the government has on the media.

Péter Gelányi

In this blog post, we will review a text-mining mini-research examining media bias, giving us insight into the steps of analysis and the decisions made by researchers, which significantly influence the results obtained. The vast majority of people encounter scientific results only through the media. Typically, they only have the opportunity to familiarise themselves with a summary of the research conclusion (often presented misleadingly). In contrast, the process leading to that conclusion remains a black box. Consequently, we may have too much or too little trust in these results.

We might have too little trust because we cannot see the research process and find it difficult to trust it. Alternatively, we might be overly confident in a particular result because, lacking knowledge of the research process, we are unaware of its limitations and the research decisions that influenced the obtained result. In reality, individual studies almost always contribute to a deeper and more comprehensive understanding of their topic, but their conclusions are not absolute; they are the result of the work of many researchers over the years. In other words, science is an iterative process. This is particularly true for social sciences, where the approach to the topic under investigation is crucial.

The subject of our mini-research is the change of ownership of two Hungarian online media products, Origo and Index, and the perceived changes in their reporting. Origo and Index have many similarities in several aspects. Origo’s news portal was sold in February 2016, with the new owners indirectly linked to the Hungarian government (24.hu, 2016). Following the change in ownership, numerous employees were laid off, and significant changes occurred in Origo’s reporting. A similar case occurred with Index in 2020 (24.hu, 2020), followed by protests (hvg.hu, 2020). Generally, the change in ownership of these two mentioned newspapers is perceived in public discourse as affecting their reporting. To investigate this question, we collected data from online Hungarian media sources. In addition to articles surrounding Origo and Index ownership changes, we collected articles from seven other online media outlets for comparison: 168.hu, 24.hu, hvg.hu, magyarnarancs.hu, telex.hu, hiradó.hu, and finally magyarnemzet.hu.

Every research study has a hypothesis, which is a preliminary assumption regarding the results of the experiment or analysis. In this case, we hypothesise that after the change of ownership, the articles of Origo and Index will be more in line with the communication of the Hungarian government than before and less in line with the communication of the opposition in Hungary. After formulating our hypothesis, the next step is operationalisation. During operationalisation, we must turn our preliminary abstract ideas into measurable, concrete phenomena. In practice, we must define an exact method for determining whether a media report aligns with government or opposition communication.

In this case, we use text mining tools (which is already a significant research decision). Among these, a relatively simple approach is examining word frequencies. Suppose we have a list of words that we believe are associated with government or opposition communication. In that case, we can look at the frequency of these words in the articles we have collected. The approach of this mini-research bears significant resemblance to the analysis of Matthew Gentzkow and Jesse M. Shapiro, but with some simplifications (Gentzkow & Shapiro, 2010). First, we need to determine the list of words that we believe are associated with government and opposition communication. We can use parliamentary speeches to select two-word lists through relative frequency analysis. Thus, we select words that government representatives use more frequently than opposition representatives and vice versa. Speeches by independent representatives and the house speaker are excluded from the data in advance. This step involves two important research decisions: firstly, we used parliamentary speeches instead of other sources, such as politicians’ Facebook posts, party programs, or news content appearing on the parties’ official websites, all of which could have been potential alternatives. The second important decision we made was to collect two-word lists instead of creating a separate list for each party present in parliament.

Following this, we can compare our list of words with the vocabulary of the collected media to see the relationship between the two. This relationship is visible in the first figure.

It is important to emphasise that these articles were created following the change of ownership, which is why Index and Origo were categorised based on expert assessment as close to the government. The chart’s columns show the difference in occurrences between expressions associated with the governing party and those associated with the opposition, subtracting the latter from the former. Thus, in the case of media categorised as close to the government, the ratio of expressions related to the governing party per article is much higher, and both Index and Origo seem to be closer to these media outlets. Among them, Magyar Nemzet stands out with particularly high values. Another important observation is that even in the case of media marked as critical of the government, expressions used more frequently by representatives of the governing parties are in the majority. This is the fact that being in the governing position generally allows for more opportunities to shape public discourse. Parties in the governing position take action and pass laws, while parties in the opposition are relegated to reacting and formulating alternatives. It is significantly easier for the former to get into the news.

After comparing the occurrence of words over a given time interval, we can now examine the temporal changes in the case of Index and Origo during the period surrounding the change of ownership. This relationship is visible in the second and third figures.

There is understandably significant fluctuation in the daily occurrences of the examined words, so we smoothed out the trend lines in the above charts using moving averages. It is important to highlight two things: firstly, we can observe that in both cases, the proportion of expressions associated with the government coalition has increased. However, based on our hypothesis, the relationship is not as straightforward as we might assume. There are several outliers in the data, and the period before the change of ownership cannot be considered stable at all, with the most notable observation occurring for Index in 2020. If we delve into the occurrences of words, we can see that the outliers are related to the coronavirus pandemic, reflecting the extra attention paid to government communication. This case highlights the limitations of our approach. While our results show a correlation between our word lists and the orientation of the media, and to some extent, we can detect changes in Index and Origo using this method, the orientation of the media is not the sole determinant of how much they reflect government or opposition communication.

In this example, we certainly did not see definitive results; the model would require significant fine-tuning. However, our goal was not to present definitive results but to highlight how research decisions influence the results of studies. Of course, the fact that a study could have been conducted in various ways does not mean that all results are equally valid. The validity of the present results is also called into question by the imperfections of our approach; we collected relatively little data, could have weighted the expressions according to their relevance, filtered them more effectively, could have taken into account their semantic similarities, etc. And we haven’t even mentioned the preparation of the texts, a process that could warrant a whole separate blog post. There are no perfect research methods, but better and worse ones. If we occasionally delve into the methodology of a study, it helps us understand how reliable its results are and what limitations the study’s conclusions have.

Péter Gelányi

In 2016, Ahmed Mansoor, an internationally recognised human rights activist, received an SMS from an unknown sender. The message promised information about torture committed by authorities in the United Arab Emirates, accessible through a link. Mansoor, who had previously been targeted with spyware, forwarded the message to Citizen’s Lab, an interdisciplinary laboratory specialising in phone spyware—in collaboration with Lookout Security, Citizen’s Lab determined that the link was associated with the Israeli NSO Group’s infrastructure. Clicking it would have compromised Mansoor’s device, uploading spyware onto his mobile phone (Marczak & Scott-Railton, 2016). In this blog post, we will briefly review the Pegasus scandal and then, as in the past, conduct a mini-analysis to examine the coverage of 4 Hungarian media outlets. The Pegasus case in the Hungarian press is a good example of how our news diet shapes what news we learn about, the aspects of it we’re informed about, and the significance we attribute to them.

The Israeli NSO (named after the company’s three founders, Niv, Shalev, and Omri) developed the Pegasus spyware in 2011. Pegasus can surreptitiously hack targeted phones, accessing all their communications, such as SMS, emails, calls, and other messages. Additionally, it can activate the camera or microphone on the device without the user’s knowledge, effectively turning the phone into a quasi-24/7 surveillance tool. The software is officially sold only to governments by NSO with approval from the Israeli Ministry of Defense, exclusively for monitoring targets suspected of terrorism or organised crime (Telex, 2021). Pegasus was used by Mexican authorities to capture the drug lord Joaquín “El Chapo” Guzmán Loera, and European authorities used it to prevent several terrorist attacks (Bergman & Mazzetti, 2022). However, the software became notorious for its misuse and abuse; it has been used all over the globe against human rights activists, collecting information on opposition politicians in Poland, which was later leaked to the Polish press (AFP, n.d. 2021), and the Saudi state also used it against women’s rights activists and journalist Jamal Khashoggi before his murder by the hands of Saudi operatives.

The software also targeted several Hungarian entities, including journalists, businessmen, media owners, the president of the Bar Association, and Attila Aszódi, a state secretary. State leaders and agencies questioned in the case did not respond whether the Hungarian government procured or used the software until November 5, 2021, when Lajos Kósa, chairman of the Defense and Law Enforcement Committee, unexpectedly claimed in an unscripted interview with RTL that the Ministry of Interior had procured the Pegasus software (Telex, 2021).

In our mini-analysis, we will examine four online media outlets: Telex and 24.hu, considered critical of the Hungarian government, and Origo and Magyar Nemzet, which are considered close to and supportive of the Hungarian government. Using the keyword “Pegasus,” we will compile a list of relevant articles from these four sources, except for Telex, where the „Pegasus” tag can be used to identify relevant articles, making our task more manageable. However, we need to manually filter out false positives for the other three media outlets. Examples include record-breaking racehorse sales, a Turkish airline, and the Pegasus constellation. Once we are sure that our collection contains only articles related to spyware, the most straightforward approach is to compare the number of articles produced by each outlet.

The results show that, as our preliminary assumptions suggested, significantly more articles were published in media outlets critical of the government regarding a scandal that was embarrassing for the Hungarian government in several aspects. Origo’s 25 articles are particularly low. In the next step, we can examine how the number of articles changes during the relevant period from July 18, 2021, to February 22, 2024 (the date of article collection).

In the second figure, we can see that in the reports of Telex and 24.hu, especially in the case of the former, a large number of articles appeared when the scandal broke, following the topic has periodically reemerged in the form of 1-2 articles per day. In contrast, the distribution of article numbers over time is much flatter for Origo and Magyar Nemzet.

Following this, we can also look into the context in which Pegasus appears in the news of the four examined media outlets. For this, we use a slightly more complicated method called named entity recognition. With the help of a large pre-trained language model, we can identify the named entities present in our texts. A named entity is any noun that denotes something specific or unique. For example, the word “capital” is not a named entity, but “Warsaw” and “Prague” are. Furthermore, named entities include names of various institutions, organisations, personal names, email addresses, etc. From these examples, it is already apparent why named entities are useful for examining the content of articles; through the named entities of an article, we can see its most important elements in the outline. First, we separate the collected texts into those from a government-critical perspective (Telex and 24.hu) and those from a government-friendly perspective (Magyar Nemzet and Origo). Then, we gather all the named entities and count how often each occurs. Finally, we can depict the most common ones in a word cloud, where the size of each expression represents the frequency of its occurrence. The first figure shows the word cloud generated from the articles of Telex and 24.hu.

The figure shows, on the one hand, organisations related to Pegasus, such as the NSO company developing the software, and media companies that have participated in international investigative work: Direkt36, Le Monde, and Citizen Lab. Furthermore, we can see the names of politicians important in the Hungarian context of the case, such as Sándor Pintér, Lajos Kósa, Judit Varga, Péter Szijjártó, Gergely Gulyás, as well as the names of several relevant Hungarian institutions, such as the Ministry of Interior, the National Assembly, and the Information Office. Finally, we can observe echoes of the case’s international aspects. Israel appears, not surprisingly. Saudi Arabia and Morocco have also abused the software on multiple occasions, and it has been revealed that they spied on high-ranking Algerian military and political leaders using the software. Overall, the word cloud thus obtained reflects a news stream that primarily focuses on the Hungarian aspect of the case. In the fourth figure, we can view the word cloud generated from the Origo and Magyar Nemzet articles using the same method.

A technical detail regarding the fourth figure is that the words are larger; the reason for this is that they are based on fewer articles, resulting in less variance in the occurrence of individual named entities, and therefore, the sizes of the words are less differentiated. Here, the reaction to international news is less evident; some relevant Hungarian politicians’ names appear, such as Judit Varga, but the most frequently occurring names are those of Hungarian politicians who were in opposition and did not hold relevant positions of power during the surveillance of Hungarian citizens with Pegasus, such as Péter Márki-Zay, Ferenc Gyurcsány, and Gergely Karácsony. Although we also saw expressions related to the European Union in the previous word cloud, their weight is noticeably greater here. Based on this, we can infer that these two outlets placed greater emphasis on discussing reactions to the scandal in their articles.

The mini-analysis presented in the blog post is simplified in several respects, but it provides insight into the reporting on Pegasus by the four media outlets. The results show that Hungarian consumers may reach different conclusions about the significance of the Pegasus scandal depending on the sources from which they obtain their news. Additionally, they may become acquainted with different aspects of the scandal’s Hungarian dimension. The Pegasus case was chosen to illustrate this phenomenon; therefore, we may not expect similar discrepancies in politically less sensitive topics.

Péter Gelányi

Introduction

The Russian-Ukrainian war is not just a battle fought with weapons; it is also a war of stories. Around the world, media outlets have shaped how people understand the conflict. Western media often portrays it as Russia’s unprovoked aggression, while Russian media frames it as a justified response to NATO’s expansion. But how do narratives play out in Czech media, a country with a very diverse media landscape?

Our pilot study used advanced natural language processing (NLP) techniques to uncover how Czech news outlets cover the Russian-Ukrainian war. By analyzing media bias and tracking the narratives in play, this research reveals how news platforms can add to the formation of public perception of the war.

What Are Strategic Narratives?

Strategic narratives are stories crafted by states and organizations to explain their actions and policies. These narratives frame conflicts, shape public opinion and justify political decisions. In the context of the Russian-Ukrainian war, two key narratives dominate:

1. Western Narrative: The war is described as an illegal invasion by Russia, violating Ukraine’s sovereignty and endangering Europe.
2. Russian Narrative: The conflict is cast as a defensive measure against NATO’s expansion and Western interference.

Understanding media bias means examining how these narratives are presented, whether they are simply reported, selectively highlighted, or actively endorsed by news outlets.

How the Study Was Conducted

This research used cutting-edge AI tools to study bias in Czech media coverage. Here’s how it worked:

1. Collecting News Articles The study analyzed articles from a range of Czech news outlets, including:
   • iRozhlas: Public broadcaster
   • Novinky: Mainstream, center-left
   • iDnes: Mainstream, center-right
   • Parlamentni Listy: Disinformation outlet
2. Classifying Narratives A powerful AI model called DeBERTa-v3 analyzed over 6,400 labeled text samples to detect whether articles aligned with Western or Russian narratives. Tools like ChatGPT were used to validate the findings, achieving high accuracy in identifying narrative bias.
3. Measuring Bias The study examined how closely sentences in each article aligned with either the Western or Russian narrative. By aggregating this data, researchers tracked bias patterns over time.

Key Takeaways

1. Different News, Different Biases Czech media showed clear differences in how they reported the war:

• Public broadcasters and mainstream outlets leaned toward the Western narrative.
• Disinformation platforms like Parlamentni Listy often echoed Russian propaganda.

Figure 1: Difference between Western and Russian Narrative over time (higher means more Russian Narrative reported)

2. AI as a Research Tool Generative AI tools proved invaluable for analyzing media bias. By fine-tuning models and labeling data, the study achieved unprecedented accuracy in detecting biases. This approach highlights how technology can enhance our understanding of complex social issues.
3. Reporting vs. Endorsing Narratives: The study emphasized the importance of distinguishing between merely reporting a narrative and actively endorsing it. This subtle difference is key to understanding the intent behind media coverage.
4. Bias Changes Over Time Media bias evolved throughout the conflict, influenced by major events like battlefield developments or international negotiations. This dynamic nature underscores the importance of ongoing analysis.

Figure 2: The Prevalence of a “Bias” Effect after the Initial Spike

Why It Matters

Being aware of media bias helps people critically evaluate the news they consume. The study underscores the importance of consulting diverse sources to get a balanced perspective on global events. Insights from this research can help governments combat disinformation and promote accurate reporting. The study’s methods could also be applied to other media landscapes to track and address bias. This work demonstrates how combining AI tools with social science theories can tackle complex questions. It prepares the way for new studies in political communication and the role of media in shaping public opinion.

Conclusion

The Russian-Ukrainian war is a stark reminder of how powerful narratives can influence public opinion. By using innovative AI techniques, this research offers a detailed look at media bias in Czech news coverage. It highlights the critical role of journalism in democracies and the need to stay vigilant against disinformation.

Jakub Stauber

The literature typically uses three approaches to media bias: coverage bias, gatekeeping bias, and statement bias. Coverage bias deals with the visibility of subjects, actors, and entities and the question of whether different events and persons appear in the media at all. Selection bias or gatekeeper bias concerns the selection or rejection of news items, data, etc., related to particular events, and statement bias concerns how events, processes, and persons are reported in particular media.

How does media bias manifest itself on two online portals representing opposing ideologies and linked to opposing political sides? The question will be examined through a series of articles published on origo.hu and telex.hu, analysing the media discourse of the teachers’ protests and related events in Hungary. After reviewing about three years of articles on public education on the two portals, it can be concluded that, although there are common themes, the emphases, framing and stylistic elements are fundamentally very different. To take a closer look at media bias, we have selected an event that appeared in both portals, but the readings of the two portals may have led the exclusive readers to completely opposite conclusions and opinions.

The event and its readings

On the morning of 24 April 2023, a sit-in took place in front of the Parliament building, with the participation of some opposition MPs and others. This was followed by a demonstration in the afternoon, which marched through the city to the Carmelite Monastery. Here, protesters attempted to dismantle the cordon around the building, leading to a scuffle with police. The police fired tear gas, and several demonstrators, including MPs, were arrested.

First, let us look at the headlines of the news reports on the demonstration.

Origo.hu:

– Left-wing activist linked to the Soros network organised this morning’s unannounced action at the BM 2023.04.24

– Protest in Budapest 2023.04.24

– ORFK: Several participants of the demonstration acted aggressively against the police 2023.04.24

– Leftists provoked with violent demonstration – with photos 2023.04.25

– The dollar left is not afraid of violence when it comes to political gain 2023.04.27

– What are left-wing politicians trying to achieve with their violent behaviour? 2023.04.28

Telex.hu:

– Protesting teachers and students march to the Ministry of Interior 2023.04.24

– Teachers and students protesting in front of the Ministry of the Interior start marching to the Prime Minister’s office, Carmelita 2023.04.24

– “Down with the cordon!” – protesting students and MPs scuffle with police at the Carmelite, tear gas fired 2023.04.24

– MPs and policemen scuffled again at Karmelita, Hadházy and Gelencs were also dragged away by hand 2023.04.24

– Demonstration for education ended in scuffle and tear gas 2023.04.24

(Potential) consequences

The demonstration reports clearly show the phenomenon of statement bias or, in another analytical approach, of different framing. The phrases “a few hundred” and “at least a thousand” may be used to describe the same quantity, but their effect is completely different. The situation is similar in the narratives of violence – the “scuffle” following the cordon-breaking is presented as demonstrator violence on the government side and police violence on the opposition portal. Police action against MPs is labelled on one side as ‘hand-pulling’ from the crowd, on the other as ‘dragging’.

Although the ability to influence the framing of news content – especially in relation to “divisive” processes and phenomena – is not significant in Hungarian research, it may, together with the selectivity of news and news elements, contribute to the development of different perceptions of reality, resulting in two unreflected “parallel information and orientation universes”, two “parallel realities”.

László Kiss

References

D’Alessio, D; Allen, M. 2000: Media bias in presidential elections: a meta-analysis. Journal of Communication. 50 (4): 133–156. doi:10.1111/j.1460-2466.2000.tb02866

Entman, Robert M. 1993: Framing: Toward a Clarification of a Fractured Paradigm. Journal of Communication, 43 (4), 51–58, https://doi.org/10.1111/j.1460-2466.1993.tb01304.x

Gentzkow, M.; Shapiro, J. M. 2008: Competition and Truth in the Market for News. Journal of Economic perspectives, 22(2), 133-154

Janky B. 2016: A téma-keretezés hatása a menekült-politikával kapcsolatos véleményekre. Regio (4). 121-139.

Janky B.; Kmetty Z.; és Naszályi N.; Tamássy R. 2018: Médiakeretezés és LMBTQ+ emberek iránti attitűdök : Egy survey- és természetes kísérlet eredményei = Media framing and attitudes toward LMBTQ people : The results of a survey- and natural experiment. Socio.hu – Társadalomtudományi Szemle (3). pp. 1-24.

Numerous studies have shown that mass media channels influence public and individual opinions on social problems, the economic situation, politics, and current events. The online press also plays a crucial role in this, as it can influence readers’ opinions not only by selecting the content to be published but also by contextualizing the news and articles and how they are presented.

The influence of the media

The media create a media representation of events by presenting different content, but this “reality” is to some extent constructed (Reifová, 2004; Uhaľová, 2017) based on criteria and editorial values (Kunczik, 1995 in Uhaľová, 2017). The editorial staff determines which events to cover, to what extent, which personnel to show, which to give less or more attention to, and what message to transmit to the reader, according to its values and goals.

According to Haselmayer and his associates (2017), the selection of news to be displayed, the so-called “gatekeeper activity”, is determined by both the supply and the demand side. On the supply side, there are individuals, events and their messages that are close to the editorial viewpoint and ideology, while on the demand side stands the needs and capacity of the readership. Another important factor is the news value, which can be determined by the presence or absence of certain news factors (Staab, 1990 In: Haselmayer, Wagner & Meyer, 2017), including “the role of power elites, conflict, surprise, relevance and continuity of the topic on the media agenda” (Haselmayer, Wagner & Meyer, 2017). Wardle (2020) distinguishes, among others, the following types of content: clickbait content, misleading content, real content reframed with false context, deceptive content, manipulated content and fabricated content.

The topic of the research

Our research investigated the critical situation of the Slovak health care system in the online press in Slovakia, on the six most visited Slovakian and two other online Hungarian news portals. We used qualitative and quantitative methods to investigate how the articles published by each news portal portray the critical situation in the health sector and how they influence the reader.

Two critical situations are addressed in the research:

1. The announcement of doctors’ readiness to strike (2022)
2. The mass resignation of doctors, the possible collapse of the health care system (2022)

The doctors’ strike preparedness was reported in the daily press on the same day, 5 April 2022, the day of the announcement of the strike preparedness. On the same day, the Slovak Ministry of Health reacted to the doctors’ union’s announcement, and several daily press reports also reported on this. The government had until the end of November 2022 to take the necessary steps so the doctors would withdraw their resignation. Before the deadline, the most widely read daily newspapers covered the situation, with articles on the doctors’ mass resignations and the actions of the Ministry of Health and the government concerning the case.

Results

It can be stated that the target audience of the different news portals is different; the text and style of the articles and writings are tailored to the needs of the reader, and all the news portals use the terms “crisis” and “threat”/”compromise” at least once in the title, but they are also used several times in the texts, the online press mentioned politicians in similar proportions, these are typically reports taken from the Slovak News Agency. It can also be noted that the online news portals included in the study take a significant amount of articles from news agencies.

The results show that the researched online press, to varying degrees but equally typical, often make a text taken from a news agency more interesting by using different, more or less eye-catching titles and subheadings and by highlighting paragraphs and sections of text to make it different from similar articles in other press. Overall, it can be said that the main influencing factor in the online press studied is how the news is reported on the subject under investigation.

Magdaléna Husťáková

This paper explores media freedom, bias, and manipulation in the Eurasian post-socialist space, proposing a novel machine-learning model to evaluate media freedom through sentiment analysis. The study defines media freedom as the ability of journalists to safely criticize influential politicians, measured by the relative sentiment in media coverage. Using over one million articles from five countries—Russia, Kazakhstan, Belarus, Ukraine, and Poland—published between 2017 and 2020, the paper investigates differences in media environments across democratic and autocratic regimes.

The results highlight significant disparities in media freedom. Kazakhstan and Belarus have the most constrained environments, with positive sentiment about influential politicians dominating the state-controlled media. In Belarus, independent or niche outlets occasionally publish critical views, but their reach is limited compared to mass-circulation newspapers. Kazakhstan’s media is highly supportive of influential elites, reflecting significant government control. In contrast, Poland emerges as the most media-free environment among the studied countries. Despite increasing polarization and government influence in some outlets, Polish media retains the ability to critically assess political leaders, reflecting robust democratic norms and institutional safeguards. Ukraine’s media landscape is transitional, with characteristics of both free democracies and more controlled systems. Some outlets owned by oligarchs exhibit a mix of positive and critical coverage, depending on political alignments.

Russia presents a complex case. While online media appears more balanced than state-controlled television, evidence of manipulation distorts the perception of freedom. For example, mentions of influential politicians decline during politically or economically unfavorable periods. This suggests sophisticated media control strategies that blend censorship with narrative management. The findings align with the notion of “spin autocrats” who manipulate public opinion through subtle methods rather than outright repression.

The study also examines media behavior during presidential elections, revealing sharp contrasts between democratic and autocratic contexts. In democracies like Poland and Ukraine, mentions of presidential candidates rise significantly during election periods, often accompanied by a decline in sentiment scores, reflecting increased scrutiny by the press. Conversely, autocratic regimes like Russia, Belarus, and Kazakhstan tightly control election narratives, portraying ruling elites positively while limiting exposure to opposition candidates. For example, Belarusian media heavily praised President Lukashenko before the 2020 elections, but the sentiment shifted as public dissent grew following contested results.

To detect censorship and manipulation, the paper uses economic indicators such as currency fluctuations to analyze changes in media sentiment. Findings reveal minimal manipulation in most countries except Russia, where influential politicians are mentioned less frequently during periods of economic downturns, suggesting a deliberate strategy to suppress unfavorable narratives. Despite this, Russia’s online media still exhibits some level of freedom compared to other forms of mass media.

The paper introduces a scalable, data-driven approach to assessing media freedom using machine learning. By incorporating sentiment and topic modeling, the model reduces subjectivity and enables nuanced analyses of media environments across diverse contexts. The findings underscore the importance of free media in supporting democratic accountability and highlight the ongoing challenges of censorship and manipulation in post-socialist countries. This approach provides a valuable tool for future comparative studies of media freedom.

Krzysztof Rybinski

Related publication: Krzysztof Rybinski: Media Freedom, Bias, and Manipulation in the Eurasian Post-Socialist Space. Communist and Post-Communist Studies 1–30. https://doi.org/10.1525/cpcs.2024.2315108

In Hungary, restrictive legislation (such as Act LXXIX of 2021) and stigmatising rhetoric continue to shape public discourse on gender and sexuality. Significant gaps remain in the media analysis, particularly regarding the influence of media framing of sexual and gender minorities. Scholars agree that Hungary experiences democratic backsliding similarly to other countries in the Central and Eastern European (CEE) region (Bernhard, 2021) and warn that this illiberal transformation cannot be understood without gender politics (Grzebalska & Pető 2021). The systematic undermining of gender equality in the media is often cited as one of the key signs of backsliding.

Communication through the media plays a key role in democratic regimes; fact-based, independent news sources guarantee freedom of speech and the free flow of information (Council of Europe, 2021). In the era of digital media, fake news and disinformation have become a serious social, economic, and political problem in CEE. The media also play a key role in raising awareness and educating the public on LGBTQI+ issues, which, in a suitably supportive environment, can lead to a higher level of social acceptance (Randev, 2022). By shaping the representation of the LGBTQI+ community, the media influences the behaviour and attitudes of people (Happer & Philo, 2013).

In line with the V4 Media Slant project’s general objectives, this research aimed to explore how the representation of LGBTQI+ issues in online news portals in Hungary has evolved in the period between 2021 and 2024. The research analysed nearly one million articles from the seven most prominent Hungarian news portals (24.hu, 444.hu, HVG, Index, Magyar Nemzet, Origo, Telex) using text mining techniques from computational social science. Approximately 6% of the articles addressed LGBTQI+ topics. Based on the Media compass of 21 Research Center (21 Kutatóközpont), the portals were classified into pro-government (Magyar Nemzet, Index, Origo) and government-critical (24.hu, 444.hu, HVG, Telex). Using sentiment analysis, the emotional content of the topics concerning LGBTQI+ issues were revealed.

Results show that the language use (measured in the frequency of anti-gender terminology such as “paedophile”, “child protection”, “ideology”) varies across political orientation; government-critical news outlets use these terms more often than pro-government outlets. This may suggest that critical outlets are quoting or even challenging the anti-gender narrative of the government.

The emotional content of the LGBTQI+ news was measured using sentiment analysis. Sentiment analysis (also referred to as opinion mining) is a method within the field of computational linguistics that emerged in the early 2000s. The primary objective of sentiment analysis is to determine whether a given textual unit—such as a word, sentence, or document—conveys positive, negative, or neutral emotional content (Liu, 2010). Sentiment analysis is a computer-supported method that can be conducted using either predefined sentiment dictionaries or state-of-the-art large language models. For instance, an article may be classified as positive if most of its analytical units (e.g., words or sentences) express positive sentiment. Classification (i.e. classifying a linguistic unit as positive, negative or neutral) can be done on a document, sentence, or aspect level.

Overall, results reveal a generally balanced outlook on the sentiment of LGBTQI+-related articles. Anti-gender content is consistently more negative than value-neutral or scientific coverage in case of all portals. Differences, where they exist, point toward individual news sites rather than the political orientation.. For instance, 444hu (government-critical) and Magyar Nemzet (pro-government) both cover LGBTQI+ content in a consistently more negative way than other portals on the same or opposite political spectrum. Therefore, it can be concluded that what matters more is who discusses the topic and in what context, rather than the political orientation alone.

To conclude, this research found that anti-gender terminology is more likely to appear on government-critical news portals than on pro-government portals, suggesting that the framing based on the language depends heavily on context. The findings also point to significant variation between individual news portals that goes beyond clear political alignment, underscoring the importance of assessing each outlet on its own terms. These nuances emphasise that media analysis must remain sensitive to both linguistic choices and the political and social context.

Limitations should be considered when interpreting the findings of this research. First, the use of keyword filtering presents challenges, such as the inclusion of irrelevant content (or the exclusion of relevant articles). The lack of a timeline analysis limits insights into how specific events—such as the passage of Act LXXIX of 2021—may have influenced coverage. Lastly, the research did not employ topic modelling, which would have added valuable information on the themes and structures of the corpus. Research on the spreading of government narratives (Horváth & Ring, 2025) proves that propaganda spreads fast in Hungarian online media. It is also worth future investigation to understand how long an anti-LGBTQI+ narrative is present and which events spark an increase in this activity.

Laura Seben

Sources

21 Kutatóközpont. (n.d.) Elkészült a 21 Kutatóközpont Médiairánytűje!. 21 Kutatóközpont. https://21kutatokozpont.hu/mediairanytu.html.

Bernhard, M. (2021). Democratic Backsliding in Poland and Hungary. Slavic Review, 80(3), 585–607. doi:10.1017/slr.2021.145.

Council of Europe Commissioner for Human Rights. (2021). Memorandum on freedom of expression and media freedom in Hungary. https://rm.coe.int/memorandum-on- freedom-of-expression-and-media-freedom-in-hungary/1680a1e67e.

Grzebalska, W., Pető, A. (2018). The gendered modus operandi of the illiberal transformation in Hungary and Poland. Women’s Studies International Forum, vol. 68, pp. 164-172, https://doi.org/10.1016/j.wsif.2017.12.001.

Happer, C., & Philo, G. (2013). The Role of the Media in the Construction of Public Belief and Social Change. Journal of Social and Political Psychology, 1(1), 321-336. https://doi.org/10.5964/jspp.v1i1.96.

Horváth, K. & Ring, O. (2025). Az állami dezinformáció terjedésének mintázatai. Mérték Füzetek, 41. https://mertek.eu/wp-content/uploads/2025/05/Mertek_Fuzetek_41.pdf.

Liu B. (2010). Sentiment analysis and subjectivity. In Indurkhya N., Damerau F. J. (Eds.), Handbook of natural language processing (2nd ed., pp. 627–666).

Randev, D.J. (2022). Role Of The Media In An Unequal World: LGBTQ Gender Identity And Portrayal. Journal of Media, Culture and Communication. DOI:10.55529/jmcc.21.1.7.

Tamássy, R. (2024). The discursive construction of the relationship between Hungary and the European Union in light of anti-LGBTQ legislation. Corvinus Journal of Sociology and Social Policy, 15 (3).