Brazil’s Bot Wars: How Automation Amplified Voices in the 2018 Election

by

Author: Denis Avetisyan

A new study reveals how social bots functioned as echo chambers during Brazil’s 2018 elections, primarily boosting specific candidates and narratives rather than driving independent conversation.

Analysis of social media sentiment surrounding the 2018 Brazilian presidential election reveals distinct trends between accounts classified as likely bots-using BotometerLite thresholds of 0.5, 0.7, and 0.9-and those identified as human-generated, suggesting potential manipulation or amplification of specific viewpoints through automated accounts, as evidenced by differing fourteen-day moving average sentiment scores.
Analysis of social media sentiment surrounding the 2018 Brazilian presidential election reveals distinct trends between accounts classified as likely bots-using BotometerLite thresholds of 0.5, 0.7, and 0.9-and those identified as human-generated, suggesting potential manipulation or amplification of specific viewpoints through automated accounts, as evidenced by differing fourteen-day moving average sentiment scores.

Network and sentiment analysis demonstrate that bots exhibited limited thematic diversity and emotional range compared to human users during the election period.

While social media promises broadened democratic participation, the proliferation of automated accounts raises concerns about manipulated online discourse. This study, ‘Echoes of Automation: How Bots Shaped Political Discourse in Brazil’, analyzes over 315 million tweets from the 2018 and 2022 Brazilian elections, revealing that bots functioned primarily as narrow amplifiers of specific candidates and narratives. Unlike human users whose sentiment and thematic engagement varied dynamically, bots exhibited limited emotional range and repetitive messaging, particularly favoring Bolsonaro-centric content. To what extent do these ‘echoes of automation’ distort public opinion and undermine the integrity of online political debate?

The Shifting Sands of Digital Discourse

The advent of social media, and Twitter in particular, represents a seismic shift in how political discourse unfolds. Prior to these platforms, traditional media outlets largely controlled the flow of information and shaped public conversation. Now, citizens can directly engage with politicians, share opinions with vast networks, and mobilize around specific issues, bypassing conventional gatekeepers. This democratization of information, however, is not without its complexities. While offering unprecedented avenues for participation and influence, the speed and reach of social media also facilitate the rapid spread of misinformation, echo chambers, and polarized viewpoints. The immediacy of platforms like Twitter incentivizes concise, often emotionally charged communication, potentially prioritizing engagement over nuanced discussion, and fundamentally altering the character of political debate.

Analyzing contemporary political conversation online presents significant challenges beyond simply tracking opinions; the digital sphere is populated by automated accounts – often referred to as bots – designed to amplify specific messages and artificially inflate perceptions of support. These accounts can skew public sentiment and distort the organic flow of discussion, making it difficult to discern genuine public opinion. Further complicating matters is the nuanced expression of emotion in online communication; sarcasm, irony, and coded language are prevalent, and algorithms struggle to accurately interpret the emotional intent behind text-based communication. Consequently, researchers must develop sophisticated analytical tools to filter out inauthentic activity and accurately assess the emotional landscape of digital political discourse, accounting for both the volume of messages and the subtleties of human expression.

The 2018 Brazilian presidential election offered a compelling microcosm of how digital platforms can reshape political realities. Intense online activity surrounding the election wasn’t simply a reflection of existing political divides, but actively amplified and restructured them. Jair Bolsonaro’s campaign, in particular, demonstrated a mastery of social media, circumventing traditional media outlets and directly engaging with voters through platforms like WhatsApp and Facebook. This direct communication, often characterized by emotionally charged messaging and the rapid dissemination of information-and misinformation-proved highly effective in mobilizing support. The election highlighted not only the potential for online platforms to empower previously marginalized voices, but also the vulnerability of democratic processes to manipulation and the spread of polarized content, making it a crucial case study for understanding contemporary political communication.

Analysis of tweet sentiment before and after the 2018 Brazilian presidential election reveals that, regardless of the bot detection threshold used, tweets classified as originating from bots (red) consistently exhibit more negative sentiment than those from humans (blue).
Analysis of tweet sentiment before and after the 2018 Brazilian presidential election reveals that, regardless of the bot detection threshold used, tweets classified as originating from bots (red) consistently exhibit more negative sentiment than those from humans (blue).

Dissecting the Digital Landscape: A Methodological Approach

Latent Dirichlet Allocation (LDA) was utilized as the topic modelling technique to discern prevalent themes within the Twitter dataset. LDA operates on the principle that each document – in this case, a tweet – is a mixture of topics, and each topic is a distribution of words. The algorithm iteratively assigns probabilities to both document-topic and topic-word distributions, identifying underlying thematic structures based on word co-occurrence patterns. Specifically, the number of topics was determined through coherence scoring, evaluating the semantic similarity between the most frequent words within each identified topic. This approach allowed for the quantification of the relative prominence of different political themes within the observed online discourse, providing insights into the key subjects driving conversation on the platform.

Text preprocessing involved the elimination of frequently occurring Portuguese stopwords – words like “e”, “de”, and “a” – which contribute little to thematic analysis. Following stopword removal, the CountVectorizer algorithm was utilized to transform the textual data into a numerical representation. This process creates a document-term matrix, where each row represents a tweet and each column represents a unique term (word) from the corpus; the values within the matrix indicate the frequency of each term within each tweet. This numerical format is essential for compatibility with the Latent Dirichlet Allocation (LDA) topic modelling algorithm and enables quantitative analysis of the textual data.

Sentiment Analysis was performed on the corpus of tweets to determine the emotional polarity expressed within the online political discourse. This was achieved using the SentiLex-PT02 lexicon, a resource specifically designed for sentiment analysis of Portuguese text. Each tweet was analyzed by matching words against the lexicon, which assigns a polarity score – positive, negative, or neutral – to individual terms. The aggregate scores for each tweet were then used to categorize the overall sentiment expressed, allowing for the quantification of public opinion and the identification of trends in emotional response to political topics. This process yielded data on the distribution of positive, negative, and neutral sentiments across the dataset, providing insights into the prevailing emotional tone of the conversation.

Bot detection was a critical component of this study’s methodology, implemented to mitigate the influence of automated accounts on the analysis of public opinion. The BotometerLite tool, developed by Indiana University, was utilized to assess the probability of a Twitter account being a bot based on its profile characteristics and activity patterns. Accounts were assigned a bot score ranging from 0 to 5, with higher scores indicating a greater likelihood of automation. A threshold was established to filter out accounts with scores exceeding a predetermined value, thereby focusing the analysis on content originating from likely human users and enhancing the reliability of the identified discourse patterns. This filtering process aimed to minimize the skewing of results caused by artificially amplified or fabricated narratives commonly associated with bot activity.

Mapping the Network: Automated Influence and Conversational Patterns

The observed Twitter data exhibited a complex network topology characterized by dense interconnectivity among users. Analysis revealed both broad-reach accounts functioning as information hubs and numerous tightly-knit clusters of users engaging primarily within their groups. Degree distribution analysis indicated a power-law relationship, suggesting the presence of a small number of highly connected nodes influencing a large proportion of the network. Further investigation identified distinct communities based on retweet and mention patterns, indicative of echo chambers and polarized information dissemination. The network’s structure facilitated rapid propagation of information, but also contributed to the amplification of biased content and limited exposure to diverse perspectives.

Analysis of the Twitter dataset revealed a substantial volume of activity originating from automated accounts. These accounts participated in multiple forms of engagement, including the dissemination of information through retweets, direct interaction via replies, and the creation of original content. The prevalence of automated activity indicates a potentially significant influence on the observed information landscape, necessitating further investigation to determine the scope and impact of these accounts on user perceptions and discussions. Quantifiable metrics related to bot activity, such as the frequency of posts and engagement rates, were key to identifying and characterizing these accounts within the broader network.

Sentiment analysis of the Twitter data revealed a marked polarization in emotional expression related to key political topics. This polarization manifested as distinct and separable clusters of positive and negative sentiment. Quantitative analysis indicated a non-normal distribution of sentiment scores, suggesting that opinions were not centered around a neutral position but rather concentrated at opposing extremes. The observed clusters were consistently associated with specific keywords and hashtags relating to the election, and exhibited limited overlap, indicating minimal cross-sentiment engagement between opposing viewpoints. Further investigation revealed that the intensity of polarized sentiment fluctuated in correlation with major campaign events and media coverage.

Analysis of Twitter data from the Brazilian electorate revealed a substantial disparity in reply behavior between automated accounts and genuine users. While human users exhibited reply rates between 20% and 40% of their overall activity, bot accounts demonstrated significantly higher reply rates, ranging from approximately 25% to nearly 80%. Critically, this bot-driven reply activity increased markedly after the election period, suggesting a coordinated effort to amplify specific messages or engage in disproportionate commentary following the results.

The Broader Implications: Deciphering the Digital Political Landscape

Empirical analysis reveals a substantial and pervasive influence of automated accounts – often referred to as bots – on the shaping of online political discourse. This study demonstrates that these accounts are not simply noise within the digital sphere, but actively participate in, and potentially manipulate, conversations surrounding political candidates and issues. By analyzing large datasets of online activity, researchers found that bots consistently amplified specific narratives, potentially swaying public opinion and even impacting electoral outcomes. The sheer volume of content generated by these automated accounts, coupled with their strategic targeting of key terms and users, suggests a capacity to artificially inflate the prominence of certain viewpoints and suppress others, raising critical questions about the integrity of online political spaces and the authenticity of public debate.

Topic modelling and sentiment analysis offer powerful tools for charting the dynamic shifts within online political discourse. By computationally dissecting large volumes of text, researchers can discern the prevailing themes – from specific policy debates to overarching ideological stances – and trace how these narratives gain traction, evolve, or fade over time. Sentiment analysis, in particular, allows for the quantification of emotional responses to political events and figures, revealing patterns in public opinion that might otherwise remain hidden. This combination enables the identification of emerging trends before they become mainstream, providing valuable insights into the shaping of public perception and potentially forecasting shifts in the political landscape. The ability to monitor these changes in real-time offers an unprecedented opportunity to understand how information spreads, how narratives are constructed, and how political conversations unfold in the digital age, ultimately enhancing the capacity to analyze and interpret the complex interplay between online communication and political outcomes.

Analyzing the structure of online political networks reveals how information disseminates and who drives the conversation. Researchers are increasingly focused on network topology – the mapping of connections between users – to pinpoint influential actors beyond simple follower counts. By identifying densely connected clusters and central nodes within these communities, it becomes possible to trace the origins and pathways of political narratives. This approach moves beyond examining what is being said to understand who is saying it and how it spreads, revealing potential manipulation attempts or the organic amplification of specific viewpoints. Understanding these network dynamics is vital for discerning authentic grassroots movements from artificially inflated trends and for evaluating the overall health of online political discourse.

The study’s findings illuminate a critical dynamic within contemporary democratic processes: the disproportionate influence of automated accounts on shaping online political discourse. Analysis reveals these accounts frequently concentrate on specific candidates – notably, Bolsonaro in this instance – limiting the breadth of topical coverage when compared to human-generated conversations. Beyond topic selection, the emotional tenor of bot activity appears constrained, exhibiting a relative stability absent in the more volatile and nuanced sentiment expressed by human users responding to unfolding political events. This suggests automated influence isn’t merely about volume, but also about a focused, consistent messaging strategy that potentially circumvents the organic ebb and flow of public opinion, underscoring the urgent need for greater transparency and accountability within online political communication channels to safeguard democratic integrity.

The study of bot activity during the 2018 Brazilian elections reveals a concerning trend: amplification rather than nuanced discussion. These automated accounts, while numerous, largely lacked the thematic diversity observed in human political discourse. This echoes G.H. Hardy’s sentiment: “Mathematics may not predict the future, but it can predict the past.” While this research focuses on digital behavior, the predictable pattern of bot amplification – favoring specific narratives – demonstrates a lack of genuine engagement with complex political issues. The algorithms driving these bots, devoid of critical thought, operate on a principle of repetition, reinforcing existing biases rather than fostering informed debate, a concerning demonstration of predictable, albeit undesirable, behavior.

The Algorithm’s Horizon

The study of automated influence, as demonstrated by the analysis of the 2018 Brazilian elections, reveals a curious symmetry. These bots, while demonstrably effective at amplification, exhibit a fundamental poverty of expression. Their thematic limitations are not merely a technological constraint, but a logical consequence of their design. A perfect amplifier requires a singular focus, a distillation of message devoid of nuance. The question is not whether these systems can mimic human discourse, but whether such mimicry is even desirable from the perspective of efficient information propagation.

Future inquiry must move beyond simple detection. The identification of bot accounts, while necessary, addresses only the symptom. The deeper problem lies in the susceptibility of the network itself. A truly robust analysis demands a formalization of persuasive rhetoric – a mathematical description of how narratives gain traction, independent of the agent transmitting them. This is not a problem of ‘fake news,’ but a question of systemic vulnerability.

Ultimately, the pursuit of detecting ‘inauthentic’ behavior feels… quaint. If a system achieves the desired outcome – the propagation of a narrative – does the origin of that narrative truly matter? The focus should shift to understanding the inherent fragility of belief systems, and the mathematical principles governing their manipulation. The algorithm is not the problem; it is merely a precise instrument revealing the imperfections of the system it interacts with.

Original article: https://arxiv.org/pdf/2512.10749.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-12 22:30