Shaping climate change discourse: the nexus between political media landscape and recommendation systems in social networks

Recommendation systems have become ubiquitous. In the age of artificial intelligence and social media, the contemporary humans live in a digital matrix, where recommendation systems have the authority to tell them what movie to watch, what person to meet, what restaurant to go and what news to read and they have been integrated in almost every aspect our lives. Companies and digital businesses rely on recommendation systems to increase their profit, and we as users rely on them as they reduce information overload and make our lives easier. They boost potential online engagement, reduce customers churn, maximize user activity on different platforms and increase sales and market share. Recommendation systems offer mutual benefits to producers and consumers. They optimize micro- and macro-customer targeting and improve key performance indicators on customers. They actively participate in creating our individual and collective tastes, and in the long run, they participate in shaping our identity with important ethical implications (Milano et al. 2020, 2021).

Recommendation algorithms have far-reaching impacts on society, and understanding these effects is crucial for various reasons. Firstly, they can inadvertently contribute to the creation of information bubbles and filter bubbles. These algorithms prioritize content similar to users’ previous engagements, potentially reinforcing biases and limiting exposure to diverse viewpoints. Additionally, recommendation algorithms filter out content that does not align with a user’s interests, narrowing perspectives and hindering exposure to dissenting viewpoints, which is vital for fostering a more informed and open society.

Secondly, privacy concerns arise as recommendation algorithms rely on user data for personalization. The collection and analysis of personal information can lead to privacy breaches. The study of recommendation algorithms is essential to assess the ethical implications of data collection and usage, ensuring that user privacy is respected. Moreover, these algorithms significantly impact monetization and advertising on digital platforms, affecting advertising effectiveness, user behavior and the revenue models of digital companies. Understanding these impacts is vital for both businesses and consumers. Furthermore, recommendation algorithms have behavioral influences, suggesting products, content or actions to users. These influences can be used for positive purposes but also for manipulation or harm, making it essential to study the mechanisms behind these behavioral effects and promote responsible technology use. Lastly, recommendation algorithms play a critical role in promoting or hindering fairness and equity by shaping the content users see. Biases in these algorithms can perpetuate societal inequalities, necessitating research and policymaking efforts to address algorithmic bias and ensure equitable access to opportunities and resources. Additionally, studying recommendation algorithms enhances transparency and accountability, shedding light on the often opaque and proprietary algorithms used by tech companies, allowing users to make more informed choices. Policymakers require a deep understanding of these algorithms to craft effective regulations that balance innovation with societal well-being, highlighting the importance of comprehensive research in this field.

Recommendation systems use and process the digital footprint of millions of users mainly to retrieve and exploit their cognitive biases, and therefore, they can personalize their recommendations and deeply hit their users. For example, Youyou et al. (2015) demonstrated that Facebook likes can accurately predict users’ sensitive information such as sexual orientation, political views, personality traits, use of addictive substances, personal separation and religious belief. In this case, they reported with 200–300 likes on Facebook, and the artificial intelligence behind the recommendation algorithm in this platform can predict users’ personality traits better than their family members or even better than themselves. As algorithms become smarter and computationally more complex, they need even less input data to predict users’ information. For example, Kosinski (2021) demonstrated that artificial intelligence can predict a person’s political orientation only by their facial images with accuracy up to 72%, which was found to be higher than accuracy of judgments made by humans.

There are many societal implications with such smart and powerful algorithms. For example, recommendation systems can affect and control public participation in political affairs (Feezell et al. 2021). In this context, digital manipulation of democratic elections is a very important implication of recommendation systems in the realm of politics. Epstein and Robertson (2015) studied that voters’ choices are significantly influenced by internet search rankings, mostly because they trust and favor higher-ranked results over those that are rated lower. In this study, researchers investigated whether search results could be manipulated to change the preferences of unsure voters in democratic elections given their apparent authority. According to their findings, biased search rankings can change the voting preferences of unsure voters from 20 to 80%, while a democratic election is won on average with a margin of 7%. Due to such serious and important impacts, digital companies have always been an attractive target for lobbying groups (Kreiss et al. 2018).

Based on such social implications, recommendation systems have a critical role in propagation of information in society, which can lead to indoctrination and mass customization of people. There are different underlying variables that affect the mechanism through which a message can propagate in a network. For example, the global structure of the network itself has been cited as a decisive factor shaping the diffusion of a message. In this case, Watts and Dodds (2007) used a series of computer simulations and interpersonal influence processes and they found that diffusion of a message over large cascades is not driven by influential people but by a large body of easily influenced individuals. Bakshy et al. (2011) further tested this hypothesis by investigating the attributes and relative influence of 1.6 million twitter users and tracking 74 million diffusion patterns that happened during a two-month time interval. They challenged the conventional assumption about celebrities’ central role in the formation of public opinion. They found that only two variables of celebrities, namely number of followers and past influence, are significant to predict whether a celebrity’s tweet can go viral, and the prediction accuracy was only 0.34, which is not a very good fit for a predictive model. They found that the density and structure of a network itself can properly explain whether there is possibility for a message to become viral. In other words, an intensively connected network among its nodes can properly predict the diffusion of a message no matter an influential or a regular person starts the cascade, and in a poorly connected network with many isolated nodes, there is no space for a message to propagate and again no matter who first starts it.

In addition to the global structure of the network, the content and attractiveness of the message itself is also an important factor affecting its successful spread across a network. Generally, those messages that can better stimulate and exploit users’ cognitive biases have a better chance of propagation. For example, Vosoughi et al. (2018) modeled differential diffusion of true and false news through ~ 126,000 stories tweeted by ~ 3 million people more than 4.5 million times and they concluded that false news diffuses significantly farther, faster, deeper and more broadly than truth. They also found out that false stories are associated with fear, disgust and surprise in replies, and they showed a higher level of novelty compared to true news and people were more likely to share false news with such elements than true stories. In contrast, true stories were correlated with features such as anticipation, sadness, joy and trust. They also found that robots are indifferent in spreading false and true stories and the reason why false news spreads more than truth is because humans accelerate their diffusion and not robots. Like fake news, it seems conspiracy theories can also attract a fair amount of attention from users. In this case, Faddoul et al. (2020) presented a longitudinal analysis of YouTube’s promotion of conspiracy videos. They found that during 2018–2019 and based on a monthly time interval, up to 10% of recommended videos were outrageously fake news or conspiracy theories. Interestingly, the name of YouTube’s recommender engine, which drives 70% of YouTube views, is called Watch Time Optimization Algorithm and it acts like a “puppet master” over what users watch to keep their attention as much as it can by exploiting their cognitive biases.

Ethnic slurs can also trigger a chain of reactions from users, and therefore, they can accelerate propagation of a message or behavior in a network. In this case, Spörlein and Schlueter (2021) studied popular German political talk shows and they showed how ethnic insults propagate through network of comments in YouTube videos. They argue that ethnic insults imply social norms, they can trigger offensive behavior, and therefore, they can accelerate their contagiousness to other insulting commenting, and finally, it can lead to propagation of the behavior over the network. The spread of such comments is even faster and more intensive at the time of radical conflicts such as sexual assaults or terrorist attacks. Likewise, Andersson (2021) studied the relationship between impoliteness and homophily based on discussions underneath YouTube videos about Swedish climate activist, Greta Thunberg. She found out that recurrent impoliteness in the comment section of the videos about Greta Thunberg is not necessarily due to the design of social media that allows such behavior behind anonymity and social detachment, but rather this behavior can signal other like-minded users to form a front and build homophilous online communities.

In addition to network structure and enticing content, there are also a series of non-content variables such as duration, like count, positive comments, title, dislike count, view and negative comments that can predict the chances of a message or video going viral. In this regard, Halim et al. (2022) developed an AI-based framework to study non-content variables that affect the chances of a video to be landed in the trending category of YouTube.

There have been different solutions implemented to reduce the negative impacts of recommendation systems on society. For example, the design of a network itself can polarize or depolarize a context. In this case, Guilbeault et al. (2018) studied how political polarization can be solved between Democrats and Republicans about climate change. They studied whether it is the case that people not interacting with each other causes further polarization or it is the case that polarization is caused by putting people together. They designed an experiment in which they studied 2400 Republicans and Democrats and they found huge polarization between them in terms of predicting future climate change patterns. Then, they created various versions of a social media network where individuals from different political affiliations (Democrats and Republicans) had the option to either see or not see the political orientation of their counterparts. Their findings revealed that when people interacted without the ability to observe or anticipate the political views of the other party, it led to a significant increase in interaction and social learning. Furthermore, this depoliticized the context, suggesting that removing visibility of political affiliations can foster greater engagement and reduce polarization. Structural attributes of the network can also dictate depolarization to its users. For example, Santos et al. (2021) studied the effect of different social network typologies on the dynamics of polarization in social networks. They found that preferential link establishment between nodes with similar structures (i.e., nodes that share many neighbors between each other are structurally similar) encourages opinion polarization and connecting nodes with dissimilar structures can reduce polarization. In contrast to insulting, personal emotions were shown to be effective in bridging the gap between polarized groups. For instance, Kubin et al. (2021) studied YouTube comments about political topics such as abortion in Fox News and CNN channels and they found that personal experiences involving harm can bridge moral and political divides better than facts. In other words, people become more tolerant toward different narratives of truth when they are accomplished with subjective experiences involving harm than providing facts.

There are different sources that rate media bias in news outlets. In this study, the bias rating result from (AllSides Media Bias Ratings 2022) has been adopted (Fig. 1), which has repeatedly been used as a source of data for media bias analysis in scientific papers. In this portal, the bias ratings are based on scientific and multi-partisan analysis. AllSides combines different crowdsourcing methods such as blind bias survey, editorial review, third-party analysis, independent review and community feedback for their ratings and in this study, their latest media bias results have been used (published in 2022). The media bias is categorized into five groups of Left, Lean Left, Center, Lean Right and Right. In this case, 67 news outlets were considered among which 66 had active YouTube channels (Alternet was the only outlet with no YouTube channel). Then, the most popular channels were selected from Left (including Lean Left), Center and Right (including Lean Right) parties to feed into web scraping and social network analysis for measuring the bias in YouTube’s recommendation algorithm. The total number of subscribers in each category (Left, Center and Right) was also used as a basis to simulate the competition between different political perspectives in society for pushing their own narrative of climate change. The competition was simulated by growing a hypothetical society using an agent-based modeling approach.

In this study, the web data was extracted from YouTube using a visual web scraping tool offered by webscraper.io, which is a browser extension for Google Chrome. This extension uses a modular structure made of selectors and it navigates the scraper on how to traverse through the target website and what data to extract. The URL, title and the hosting channel of the videos were extracted. Accordingly, the data collection process for YouTube videos in this study involved several steps:

The data collection process is designed to minimize bias and provide a structured dataset for subsequent analysis and modeling. Figure 2 shows the schematic representation of the network typology used in this study to scrap YouTube data.

Social network analysis is a mathematical framework for studying social structures using networks and graph theory. The structure of a social network is characterized by nodes (i.e., videos and hosting channels) and links (i.e., hyperlink URLs to videos and channels) that connect nodes to each other. In this study, Gephi version 0.10 (Bastian et al. 2009) was used to study the network structure of recommended videos and their hosting channels. In this case, a directed network was used to measure the frequency and distribution pattern of recommended videos and their hosting channels based on web data derived from the last step and check whether there is a specific biased pattern fitting into political agenda of each channel. The descriptive statistics of distribution pattern of recommended videos and their hosting channels were then calculated, and the Q–Q plots were drawn to visualize the pattern of the distribution.

An agent-based model is a computational tool for simulating the behavior, actions and interactions of autonomous agents with a system of values that interact locally, and the result of their collective interactions can lead to emergence of a global pattern. Agent-based models combine different elements from game theory, complex systems, computational sociology and evolutionary programming. They have a wide variety of applications in biology, ecology and social science. In case of this study, this tool was used to simulate a theoretical society based on two-step flow of communication model (adopted from Hilbert 2023) to study the competition between Left, Center and Right media outlets in propagating climate change information in society from their own perspective. The relative size of the media for each party (Left, Center and Right) was configured based on the total number of subscribers to all their respective channels in YouTube. Three media outlets were introduced to the model as patch-based agents on which moving agents can walk onto and collect information. Moving agents represent humans in a society that can collect and store information in their memory whenever they interact with a media patch. The agents were programmed to wander through agent-based environment in a random pattern such that they take a 60-degree turn to a random direction during each model run. Put simply, every time an agent walks onto a media patch of a specific party, they consume one unit of information from that media and add it to their memory and then keep track of the same process through the next model runs. Based on the concept of two-step flow of communication (Baran and Davis 2014), the agents were able to communicate with each other and affect each other’s performance.

According to two-step flow of communication model, most people are not directly influenced by mass media, and instead, their beliefs and opinions are shaped by opinion leaders who receive and interpret the message from the mass media and put them into context. In other words, opinion leaders or influentials are those who first get exposed to a specific media content and then interpret it according to their own opinion. The influentials then infiltrate the general public to indoctrinate the opinion into their followers. This process was internalized into the model through the interaction and information exchange between agents such that when two agents with information from different media outlets meet each other at the same point in the environment, the one who has accumulated less information from a media patch, sets its memory back to 0. In other words, it is influenced by a more influential agent, who has consumed more information from a different media outlet. Since the relative size of the media patches correspond to the number of subscribers from all their respective channels in YouTube, an additional element was added to the model to configure the results under different scenarios. In this regard, since the size of Left YouTube channels vastly dominate the media landscape, a series of weights were assigned to the message from the Right channels to study and compare the effect of media size and the weight of a message together on propagation of information in a society. In the context of this research, the weight of the message refers to the level of effort that a media outlet invests in crafting their message to make it more appealing, convincing and impactful for their target audience. In real-life terms, the weight of the message reflects the attractiveness and persuasive power of the content in capturing the attention of viewers and encouraging them to embrace the new message being conveyed. By assigning a weight to the message, we aimed to capture the potential influence and reach of the media’s narrative, highlighting its ability to displace prior messages and effectively engage a wider audience. In other words, although Right media are smaller in size, what would happen if they allocated their resources to tailor their message in a very engaging and convincing manner compared to other media with bigger pool of subscribers? This question was explored by assigning weights to those agents carrying information from Right media in the environment. The set of weights was 100, 10,000 and 1,000,000, respectively. Put simply, agents from Left and Center media can beat Right agents in information exchange only if they have already accumulated 100, 10,000 and 1,000,000 times more information than Right media agents; otherwise, the Right agents will win and set the information value of their counterparts back to 0. By comparing the results under such scenarios, one can study whether it is the size of the media or the weight of the message that affects the success of a political party to push their own narrative of climate change in society. As a comparative basis to the weighted scenarios, one more scenario was added in which only agents of Left and Right media compete but the conditions for the size of their media and the weight of their messages were absolutely equal.

The model parameters included three counters, measuring the real-time count of messages from various parties being carried by agents. Additionally, a slider was used to set the initial population size at the beginning of the model run. An integral component was the dynamic graph, which tracked the total number of messages associated with each party as they were being processed. This dynamic graph was used to monitor and analyze the flow of messages throughout the model’s operation, and the result of the graph was exported in the form of an excel file to draw Fig. 5. Netlogo version 6.3.0 (Tisue and Wilensky 2004) was used to configure the agent-based model.

Figure 3 demonstrates the overall flowchart of this study.

The number of subscribers for each media channel was extracted from YouTube (Table 1). Accordingly, Left and Lean Left channels together totally dominate the media landscape in terms of both the frequency of channels and the total number of subscribers. The total number of subscribers to Left, Center and Right channels was 80,965,200, 23,220,700 and 26,296,350, respectively. The most popular channels across Left, Center and Right showed less difference in terms of the size of subscribers. CNN, BBC News and Fox News were the most subscribed channels for Left, Center and Right parties with 14,600,000, 13,700,000 and 10,200,000 subscribers, respectively. In addition to having a smaller pool of subscribers, the drop rate in the number of subscribers across Center and Right channels is much bigger than that of the Left media, which means the majority of followers are only subscribed to first few channels. In other words, Left channels have a much bigger network of audiences from society, higher diversity in their media landscape and less inequality across their media outlets. CNN, BBC News and Fox News were selected for web scarping and analyzing the bias in YouTube’s recommendation system, and the total number of subscribers was used as an input to configure the agent-based model.

CNN, BBC News and Fox News are not only impactful news organizations but also the most subscribed YouTube news channels among all other media outlets from their respective political parties. This distinction amplifies their influence on the digital landscape, where millions of viewers from around the world actively engage with their content and contribute to the ongoing discourse on climate change. With millions of viewers worldwide, these news outlets have the power to inform, educate and sway public opinion on climate-related issues. They provide a spectrum of perspectives, from progressive to conservative, which resembles the diversity of opinions on this matter.

The web scraping process for these channels was undertaken separately, and it resulted in structured data files such that a user could feed them into social network analysis with minor data wrangling efforts. The extracted data included the title and URLs of the recommended videos and their hosting channels. The data for recommended videos and channels was first divided and then fed into Gephi to separately analyze the bias in YouTube’s recommendation algorithm for both recommended videos and YouTube channels.

The results of social network analysis of YouTube’s recommendation algorithm are presented in Table 2, 3 and 4. Table 2 shows the top 10 climate change-related recommended videos and YouTube channels to a Fox News viewer. As the titles of the videos indicate, the content of all these videos has a skeptical stance toward climate change. Fox News itself, by a big margin to the second-most suggested channel, is the most recommended channel to its own users, and most of the other top recommended channels are labeled under Right category in media landscape. Forbes Breaking News from Center and The Young Turks and MSNBC from the Left media are also top recommended channels to Fox News users but with a big difference to Fox News self-recommendations. The recommendations accounting for these top 10 videos and channels cover 15% and 31% of all recommendations for recommended videos and channels, respectively, which shows they have a very big impact on overall performance of YouTube’s recommendation algorithm.

Table 3 shows top 10 recommended videos and channels to BBC News users. All top recommended videos are from BBC News itself and it seems the content of all videos acknowledge climate change as a global concern except one video which criticizes the concept of climate change in a political context (video title: Vladimir Putin criticizes Greta Thunberg’s UN speech on climate change—BBC News). All top recommended channels are labeled as either Center or Left (only WION is labeled as slightly Right), which shows BBC News is more on the Left side when it comes to climate change and this matter is supported by YouTube’s recommendation algorithm as well. The top 10 recommended videos and channels, respectively, cover 16% and 24% of all recommendations made to BBC News users, and like Fox News, BBC News self-recommendations to its channel totally dominate the performance of the recommendation algorithm.

Table 4 shows the top 10 recommended videos and channels related to climate change to CNN users. All the top recommended videos predominantly originate from CNN itself, while the highest recommended channels consist of TEDx Talks, StarTalk and CNN, with minimal variations between them. All recommended videos confirm climate change as a global concern (the second top recommended video is a debate between a democrat and republican on climate change) and top recommended channels are either Left or Center which has a similar stance toward climate change. These top 10 recommendations cover 16% and 27% of all suggestions made to CNN users.

It seems YouTube’s recommendation algorithm is biased toward Left and Center media as they have a chance to be among top recommended channels to Fox News users, but the opposite is hardly true. Center media, TEDx Talks channel, appears to be among top recommended channels to all users of CNN, BBC News and Fox News followed by another Center media, DW Documentary, which is located among top recommended channels to CNN and BBC News users. In general, there are four YouTube channels commonly found among the top recommendations for both CNN and BBC News viewers. However, the overlap with Fox News is limited to only one channel. In other words, the network of Fox News users is less connected to the network of users from Left and Center, which means users of CNN and BBC News are less likely to end up in Right media, but Fox News users have a fair chance of getting recommendations from Center and Right channels.

Table 5 presents a series of global statistics for networks of recommended videos and channels to CNN, BBC News and Fox News users. For example, all average values are greater than medians for recommended videos and channels and it means the data is heavily skewed to the right and the skewness values also confirm the same issue. (Skewness > 0.5 means data distribution is not symmetric.) This issue is also visually presented in Fig. 4. In other words, the distribution patterns of links to recommended videos and channels suggest that an extremely small number of videos and channels receive a disproportionately high number of recommendations, while a vast number of unique videos and channels receive very few recommendations through YouTube’s recommendation algorithm. This biased distribution indicates a preference for specific content within the algorithm.

The results of agent-based modeling based on two-step flow of communication model are presented in Fig. 5. The results of agent-based modeling were exported in the form of excel files, and the graphs were visualized in R environment. This figure visualizes the results of the competition between the media from different political parties in spreading climate change information in society from their own perspective. Figure 5a shows the result of competition between only Left and Right media outlets under equal conditions. In other words, equal number of subscribers was assumed for both media with equal weights to their corresponding message. The figure clearly illustrates a cyclical pattern in which Left-leaning media sources and Right-leaning media sources take turns in dominating the information landscape. Initially, Fig. 5a portrays a scenario where Right-leaning media outlets wield a significant influence over public discourse. However, as time progresses, the Left-leaning media outlets begin to gain momentum, eventually surpassing their Right-leaning counterparts in terms of the volume and impact of their messaging, and the same cyclic pattern is constantly repeated over time. In essence, according to the figure, there is a recurrent pattern where Left-leaning and Right-leaning media sources take turns in dominating the societal conversation on climate change and there is no single absolute winner and both sides have equal chance in preaching their ecological agenda. In Figs. 5b–d, the model is configured for the size of the media according to the total number of subscribers from all their respective channels in YouTube and the competition was simulated with higher weights assigned to the message of Right media. The weights were 100, 10,000 and 1,000,000 for Figs. 5b, c and d, respectively. Accordingly, the Left media under all scenarios dominate Right and Center media in spreading their own narrative of climate change in society. In other words, the size of the audience has a bigger impact on the success of a respective media than the weight of its message. If a media has a smaller network of subscribers, their message will not reach beyond a certain limit no matter how big the weight of their message is, and in contrast, if a media has a bigger pool of followers, their message will reach to broader network of people in society no matter how small the weight of their message is. Accordingly, owing to their relatively equal network sizes, the graphs for Right and Center media outlets illustrate similar patterns (Figs. 5b–d), with competition between these two parties being more balanced compared to the Left media outlets, which completely dominate the information landscape regarding climate change discourse.