Tracing Bias to Its Sources: A Word Embedding Audit of Racism in South African News Outlets

Tracing Bias to Its Sources: A Word Embedding Audit of Racism in South African News Outlets | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Tracing Bias to Its Sources: A Word Embedding Audit of Racism in South African News Outlets Nnaemeka Ohamadike, Kevin Durrheim, Mpho Primus This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9327726/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Race bias in South African news is well documented, but existing research has treated the media as a collective entity, leaving open the question of which specific outlets drive these patterns. This study addresses that gap by tracing racial bias to its institutional sources using word embeddings. We trained an ensemble of 10 Word2Vec models on resamples drawn from a corpus of 27,140 COVID-19 vaccination news articles, with each resample comprising 3,900 articles across 39 outlets. Each outlet was embedded as a vector based on its language, and that vector’s association was measured with validated racial stereotype vocabularies. We first found that socioeconomic race bias from a prior study replicates in the present study and correlates strongly with South African human judgments. The outlet-level analysis in this study reveals that business and finance outlets are most strongly associated with White stereotype language, while metropolitan, community newspapers, and government media are most associated with Black stereotype language. All outlets showed meaningful associations with both stereotype sets, suggesting that racial bias is a feature of South African news language rather than the product of a few outliers. The method provides a scalable, replicable framework for auditing racial bias at the institutional level and informing targeted interventions in newsroom practice and media policy. word embeddings racial bias news outlets South Africa outlet accountability computational social science Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction South Africa has struggled with racism in its news media, a struggle that predates democracy and continues to shape how different racial groups are represented in public discourse today. Apartheid-era news platforms were instruments of state-sanctioned racism – controlled largely by white-owned corporations, they legitimised white supremacy, spread propaganda, actively suppressed or marginalised Black voices, and were a tool for the apartheid government to control information (Braude, 1999 ; Berger, 2001 ; Durrheim et al., 2005 ). The South African Truth and Reconciliation Commission later documented the media’s complicity in gross human rights violations during this period (TRC, 1998). The transition to democracy in 1994 brought policy reforms and efforts to diversify newsrooms, but the underlying patterns of racism were harder to dislodge. Investigations by the South African Human Rights Commission (HRC, 1999) and the Media Monitoring Project (MMP, 1999, 2006) found racism in post-apartheid news coverage, with Black communities frequently framed through stereotypes of criminality, poverty, and social disorder. More recently, Govenden ( 2022 ) showed that despite significant changes in media ownership since 1994, “inferential racism” continued to shape representations of Black South Africans in print media through to 2014. For most of this history, evidence of media racism came from manual qualitative analyses. These approaches have produced valuable insights but are limited by relying on individual interpretation, are difficult to replicate, and are poorly suited to the scale of analysis needed to draw representative conclusions across dozens of outlets and thousands of articles (Kozlowski et al., 2019 ). They are also better at identifying explicit bias than the recent subtler forms of racism embedded in news language (see Govenden, 2022 ). Word embeddings provide a way around these limitations. They represent words as vectors in a high-dimensional space, where the distance between vectors encodes semantic association based on how words co-occur in a large corpus. Crucially, embedding models are not programmed with any assumptions about race – they learn associations from the text they are trained on. This makes them well-suited to detecting implicit bias: if a news corpus consistently places “Black” near words like “crime” and “poverty,” the model will learn and reflect that from its training data. Researchers have used this property to study ethnic and gender bias in historical corpora (Garg et al., 2018), human-like biases encoded in language (Caliskan et al., 2017 ), gender stereotypes and debiasing (Bolukbasi et al., 2016 ), broad cultural dimensions such as class and race (Kozlowski et al., 2019 ), obesity-related schemata in news (Arseniev-Koehler & Foster, 2022 ), and the sentiment structures underlying natural language (Van Loon & Freese, 2023 ). Ohamadike et al. ( 2025a ) used this computational approach to examine racism in 27,140 South African news articles from 76 outlets. They found racism in the coverage of socioeconomic phenomena across corpora, with “Black” frequently associated with crime, poverty, and poor socioeconomic conditions, and “White” with safety, wealth, civic virtue, and better socioeconomic conditions. These results were validated in many ways, including through human judgment, but they stop short of identifying which specific news platforms are responsible for driving it. This is the gap the present research addresses. Knowing that bias exists in aggregate tells us that something is wrong; knowing which platforms produce it tells us where to intervene. Using the same news corpus and embedding approach from Ohamadike et al. ( 2025a ), we adapt the speaker landscape method (Schuld et al., 2024 ) to embed each news outlet as a vector based on its language and measure that vector’s association with racial stereotypes. This allows us to identify those whose language most closely mirrors the stereotypes previously found in the aggregate corpus, providing, for the first time, a computational outlet-level account of institutional responsibility for racial bias. In what follows, we review the literature on racism in South African news and on the use of word embeddings to study racial bias, before describing the methodology, results, and implications of the analysis. 2. Literature Review 2.1 Racism in South African News The story of racism in South African news cannot be told without starting with apartheid, because the structures, practices, and assumptions of that era did not disappear when the regime ended. During apartheid, the mainstream press was predominantly owned and controlled by white corporations aligned with the ruling government. News coverage systematically dehumanised Black South Africans – depicting resistance movements as threats to order, portraying Black communities through imagery of violence and primitivism, and framing state violence against protesters as a legitimate response to chaos (Posel, 1990 ; Braude, 1999 ). Posel ( 1990 ) documented how the South African Broadcasting Corporation (SABC) news in the 1980s described township uprisings in language that cast Black protesters as mindless mobs, effectively justifying the apartheid state’s violent suppression of dissent. The Truth and Reconciliation Commission later confirmed the extent of the media’s complicity in apartheid-era human rights violations (TRC, 1998). These patterns persisted after apartheid. Just five years into democracy, the Human Rights Commission launched a formal investigation into racism in the South African media following complaints that outlets continued to produce racially biased coverage (HRC, 1999; Pityana, 2000 ). The Media Monitoring Project’s contribution to that inquiry documented recurring racial themes in news content, such as ‘Blacks are criminals’, and the investigation established that racism had not been resolved by political transition (MMP, 1999; Durrheim et al., 2005 ; Pityana, 2000 ). A follow-up MMP report years later found much the same: coverage of race and xenophobia was characterised by dramatic, stereotype-reinforcing framing that lacked context, and marginalised groups continued to be depicted in ways that reproduced their social exclusion (MMP, 2006; Wasserman, 2007 ). Govenden ( 2022 ) examined the longer arc of this problem, tracking print media content from 1994 to 2014 and finding that, despite the growth of Black media ownership during this period, “inferential racism” remained the dominant mode – bias embedded in framing choices, source selections, and narrative structures of everyday journalism. The persistence of these patterns is partly a product of structural factors that outlast formal policy change. Commercialised, market-driven press tends to reproduce the assumptions and interests of dominant economic groups (Duncan, 2014 ; Radebe, 2022 ). Wasserman ( 2010 ) points to the influence of Western journalistic models imported through South Africa’s colonial and apartheid-era ties to Britain, which carry their own assumptions about whose voices are authoritative and which events are newsworthy. Newsroom diversification in terms of race and gender has improved, but has not translated into transformed content, often prioritising the worldview of elites in a capitalist framework (Duncan, 2014 ; Govenden, 2022 ; Radebe, 2022 , 2023 ; Wasserman, 2010 ). Testing this computationally, Ohamadike et al. ( 2025b ) found that South African news journalists often quoted White speakers on global and technical vaccination topics, while confining Black speakers, especially South African ones, to localised coverage. What is missing from this body of research is an account of institutional responsibility. Studies have established that racial bias exists in South African news. But they have largely treated the media as a collective entity, leaving unanswered the question of which specific platforms contribute most to these patterns. This matters for two reasons. First, South Africa’s news landscape is not homogeneous: it includes public broadcasters, commercial print and digital outlets, specialised business publications, and community media, with different ownership structures, editorial cultures, and audience bases. Second, targeted accountability requires knowing who is responsible. Aggregate findings establish that a problem exists; they cannot, on their own, inform the institutional interventions – in editorial policy, journalist training, and media regulation – needed to address it. 2.2 Word Embeddings as a Tool for Detecting Racism Word embeddings have become one of the more productive tools for studying implicit bias in large text corpora. The core idea is that words that appear in similar contexts will have similar vector representations, and the geometric relationships between these vectors encode semantic associations based on how words co-occur in the text. A word’s position in the embedding space reflects the company it keeps in the corpus, making the space a kind of map of the cultural associations encoded in the language. Caliskan et al. ( 2017 ) provided an influential early demonstration. Using GloVe embeddings trained on web text, they showed that European-American names were more closely associated with positive attributes than African-American names in the embedding space, mirroring the results of Implicit Association Test studies in social psychology. Garg et al. (2018) extended this work by tracking how ethnic stereotypes in American English shifted across the twentieth century, training GloVe embeddings on historical text corpora. They found that negative traits associated with Chinese immigrants in the early 1900s gave way to more positive associations by the century’s end, a shift that correlated with Asian-American population growth. Their work demonstrated that embeddings can serve as a longitudinal instrument, not merely detecting bias but tracking how it changes over time. Kozlowski et al. ( 2019 ) improved the specific methodological approach most relevant to the present study. They showed that, for race, a semantic dimension defined by the difference between a cluster of “Black” terms and a cluster of “White” terms allows any word in the embedding to be scored for how closely it associates with each pole. They found, using Google News embeddings, that jazz and rap were more strongly associated with the Black pole, and opera and bluegrass with the White pole. In the South African context, Ohamadike et al. ( 2025a ) trained word2vec embeddings on a corpus of South African vaccination news and found that the race dimension reliably separated socioeconomic stereotypes, with poverty, crime, and welfare-associated terms clustering at the Black pole and wealth, safety, and terms representing good socioeconomic conditions at the White pole, and the results corresponded with human judgments. The speaker landscape method, introduced by Schuld et al. ( 2024 ), allows for extending this analysis from words to people and institutions. By inserting speaker or source identifiers as tokens into the training corpus alongside the text associated with each actor, the method produces a vector for each speaker or source that captures their surrounding language. This vector can then be subjected to the same cosine similarity analyses used for words – locating actors in the embedding space, measuring their proximity to ideological or thematic issues, and comparing groups. Ohamadike et al. ( 2025b ) used this method to study identity bias at the speaker level in South African news quotes, finding that ‘journalists privilege White voices as more authoritative and connected to global and technical vaccination discourse but confine black voices to primarily localised contexts’ (p.1). The present study adapts this framework to the platform level, treating each news outlet as the unit of analysis and measuring its vector’s association with racial stereotypes. Overall, embeddings remain an effective tool for studying implicit bias at scale, and when applied at the platform level, one of the few methods capable of attributing aggregate patterns of media bias to their institutional sources. 3. Methodology 3.1 Data This study builds on Ohamadike et al. ( 2025a ), using the same corpus of 27,140 South African news articles from 76 outlets covering COVID-19 vaccination. Articles span from January 2020 to May 2023 and were collected by Media Monitoring Africa (MMA). The same preprocessing pipeline from the original study was applied: all text was lowercased, punctuation and HTML tags removed, and no stopwords were excluded. To attribute bias to its institutional sources, each article was prefixed with a single-token label representing its news platform (e.g., source_news24 ), enabling word2vec to learn a vector for each outlet based on its language. This approach was adapted from Schuld et al. ( 2024 )’s speaker landscape. It allows each platform to acquire a unique vector representing its overall linguistic profile, used in coverage and can then be associated with stereotype centroids. Only platforms with at least 100 articles were included to ensure representationally stable vectors. For platforms exceeding this threshold, 100 articles were sampled per run (described below), equalising representation across outlets and ensuring that bias estimates reflect language use rather than publication volume. This yielded 39 eligible platforms and 3,900 articles in each resampled data set for training each of the 10 embeddings. 3.2 Word Embeddings Word embeddings were trained using the Skip-Gram architecture of the word2vec algorithm (Mikolov et al., 2013 ) implemented in Python’s Gensim library (Řehůřek & Sojka, 2011 ). Word2vec is a neural embedding method that learns vector representations for words by predicting the context in which they appear. When trained on large corpora, these representations encode the semantic relationships present in the text, including latent biases absorbed from human language (Caliskan et al., 2017 ; Kozlowski et al., 2019 ). We used word2vec word embeddings for simplicity and in line with the previous research. To account for stochastic variation in training and produce stable estimates, we trained an ensemble of 10 embeddings with different random seeds, resampling 100 articles per platform in each run. This resampling procedure serves two purposes: it controls for the disproportionate influence of high-volume outlets on the embedding space and allows uncertainty in our estimates to be quantified as cross-run variability. The hyperparameter settings used in training are shown in Table 1 . Compared with Ohamadike et al. ( 2025a ), the vector dimensionality was reduced from 250 to 200 and the minimum word count from 5 to 3, to accommodate the smaller per-run corpus. Table 1 Hyperparameter settings for word2vec training Parameter Name Value Minimum word count (m) 3 Window size 10 Architecture Skip-Gram (sg = 1) Negative samples 5 Vector dimension size 200 Epochs 20 Number of bootstrap embeddings 10 Number of worker threads 8 3.3 Constructing the Race Dimension We constructed a term-based race dimension using the same Black/White terms employed and validated by Ohamadike et al. ( 2025a ) (Table 2 ). The authors validated these terms through different means, including by projecting individual-specific Black and White speaker names, Euro-American and African country names, and non-individual-specific Black and White names onto the race dimension. They found that these names were racialised – the model and race dimension, in most cases, learned to associate the names with the relevant region (Black or White). Table 2 Word Pairs used to Construct Term-based Race Dimension Word Pairs black-white blacks-whites afro-anglo african-european african-american indigenous-settler zulu-afrikaner For each group in Table 2 , we computed a centroid – the average of all corresponding word vectors across all 10 embeddings. The race dimension is the vector connecting the two centroids (Black minus White), normalised to unit length. Positive cosine similarity with this direction indicates association with the Black pole; negative similarity indicates association with the White pole. 3.4 Analysis 1: Replicating Socioeconomic Bias To establish that the socioeconomic biases identified in the previous study persist in the platform-level corpus, we projected the same set of socioeconomic stereotype keywords onto the race dimension. These stereotypes were drawn from three sources: the Media Monitoring Project (MMP, 1999), Talbot and Durrheim ( 2012 ), and ChatGPT (2024a), and represent socially constructed and historically grounded associations tied to race in South Africa (see Ohamadike et al., 2025a for the full list and sources). For each stereotype that appeared in at least 5 of the 10 embeddings, we computed its cosine similarity with the race dimension in each embedding and averaged these scores across embeddings. The resulting mean cosine similarity score indicates each term’s directional association with the Black or White pole, with standard error across embeddings quantifying estimate uncertainty. To formally test whether White stereotypes cluster more strongly with the White pole and Black stereotypes with the Black pole, we applied the Word Embedding Association Test (WEAT) (Caliskan et al., 2017 ). The WEAT effect size is calculated as Cohen’s d – the standardised mean difference in cosine similarities between the two stereotype groups – divided by their pooled standard deviation (Ohamadike et al., 2025a ). Statistical significance was evaluated with a permutation test over 10,000 iterations, randomly reassigning group labels to estimate the null distribution of the test statistic. 3.5 Analysis 2: Human Validation of Bias Scores To validate whether the cosine similarity scores derived from the platform-level embeddings reflect the racial associations held by South African participants, we correlated the stereotype scores from Analysis 1 with human ratings collected in Ohamadike et al. ( 2025a ). In that study, 26 South African participants (21 Black, 4 White, 1 Indian) rated each socioeconomic stereotype on a 5-point Likert scale: 1 (Strongly Associated with White) to 5 (Strongly Associated with Black). Inter-rater reliability was evaluated using Krippendorff’s alpha with bootstrapped 95% confidence intervals and intraclass correlation coefficients (ICC), both scaled from 0 to 1. The results indicated moderate agreement among individual raters (Krippendorff α = 0.54, 95% CI [0.46, 0.60]; ICC(A,1) = 0.55, p < 0.001). For the present study, we averaged human ratings per stereotype and merged them with the corresponding mean cosine similarity scores. We then computed Pearson’s r between the two sets of scores. A significant positive correlation would indicate that the platform-level embeddings capture racial associations in ways that accord with human judgement, validating the bias scores. 3.6 Analysis 3: Platform-Level Bias Measurement The central analysis of this paper measures the degree to which each news platform’s learned vector is associated with Black and White stereotype language. We adapted the speaker landscape approach (Schuld et al., 2024 ) in which news platforms are positioned relative to their language use. 3.6.1 Stereotype Centroids For each of the 10 embeddings, we computed a White stereotype centroid as the mean vector of White stereotype terms present in that embedding that projected cleanly onto their theorised side of the race dimension, and analogously for Black stereotypes. Using embedding-specific centroids rather than a single grand centroid preserves within-embedding coherence: each platform vector is compared to centroids derived from the same vector space, avoiding cross-embedding vector incompatibility. Not all stereotypes project cleanly onto their theorised side of the race dimension and may introduce noise into centroid construction. We therefore retained only those stereotypes whose mean cosine similarity from Analysis 1 was directionally consistent with their theorised category (Black stereotypes with positive mean projection; White stereotypes with negative mean projection). Terms projecting to the opposite pole or at zero were excluded, although robustness checks including all terms did not materially change the results. Across embeddings, each platform’s cosine similarity with these centroids was averaged to produce stable estimates, with standard error quantifying cross-embedding variability. 3.6.2 Platform Association Scores For each embedding, we computed two cosine similarity scores: the similarity of the platform vector to the embedding’s White stereotype centroid and to the Black stereotype centroid. These scores capture the degree to which each platform’s language aligns with each set of stereotypes. The two platform association scores capture the degree to which each platform’s language is associated with each stereotype set (Black/White). We report both scores together to enable a full characterisation of each platform’s racial framing. Outlets in the high-White, high-Black quadrant of the resulting space exhibit strong stereotyping across both racial groups; those in the low-low quadrant show minimal stereotypical language. Asymmetric positions reveal directional framing. We also compute a difference score (White similarity minus Black similarity) to rank platforms on the net direction of their racial associations. 4. Results We present the results in four parts. We begin by establishing whether the socioeconomic race biases documented by Ohamadike et al. ( 2025a ) are present in the platform-level corpus, then validate those bias scores against human judgments, and finally report the main analysis attributing racial bias to individual news platforms. 4.1 Socioeconomic Stereotypes and the Race Dimension Figure 1 shows the projection of the socioeconomic stereotypes onto the race dimension derived from the platform-level embeddings. The x-axis represents each term’s cosine similarity with the bias direction, where positive values indicate association with the Black pole and negative values indicate association with the White pole. Error bars reflect the standard error of each term’s score across the 10 resampled embeddings. The results replicate the core finding from Ohamadike et al. ( 2025a ) (WEAT = − 1.31, p < 0.001): race bias in socioeconomic language is clearly present in our results, with most Black stereotypes projecting on the Black side and most White stereotypes projecting on the White side (WEAT = − 1.04, p < 0.001). Further confirming this replication, cosine similarity scores from the platform-level embeddings correlate strongly with those from the original study (r = 0.75, p < 0.001), showing that despite differences in corpus construction and hyperparameters, both embeddings learned similar racial associations from the same news data. The strongest associations with the Black pole in this study include terms like “marginalized,” “traditional,” “township,” “rural,” and “shack,” as well as government welfare identifiers specific to the South African context such as “SASSA” (the South African Social Security Agency), its former Social Relief of Distress Grant amount of “R350,” and “NSFAS” (South African government’s National Student Financial Aid Scheme targeted at low-income households). At the other end, the strongest associations with the White pole include White-stereotyped economic and civic terms like “advantage,” “investor,” “taxpayers,” “ambitious,” and “developed.” 4.2 Validation Against Human Judgements Figure 2 shows the relationship between the embedding-derived cosine similarity scores from Fig. 1 and the human ratings collected by Ohamadike et al. ( 2025a ), in which 26 South African participants rated each stereotype’s association with Black or White on a five-point Likert scale. The correlation between the two sets of scores is positive and highly significant ( r = 0.61, p < 0.001), indicating that the racial biases or associations encoded in the platform-level embeddings reflect the intuitions of South African human raters, validating our results. 4.3 Platform-Level Racial Bias Figures 3 and 4 present the main findings of the study – the association of each news platform’s embedding vector with the White and Black stereotype centroids. Figure 3 plots each platform’s Black stereotype association (x-axis) against its White stereotype association (y-axis), locating platforms in a two-dimensional bias space. Figure 4 ranks platforms by the difference between these two scores (White minus Black), computed per platform and averaged across the 10 resampled embeddings, providing a single directional measure of each outlet’s net racial orientation. The most immediately noteworthy feature of Fig. 3 is that all 39 outlets show meaningful association with both stereotype sets, indicating that race bias is not confined to a subset of outlets but is a feature of South African news language across the board. The clearest pattern in Fig. 3 is the separation of platforms by institutional character. Business, finance, and investment-focused outlets (such as Moneyweb, Financial Mail, Business Day, and Fin24) cluster in the upper-left quadrant, characterised by high White stereotype association and relatively low Black stereotype association, with Moneyweb positioned furthest left on the Black side. The language of these outlets captures the world of investment, corporate ownership, wealth, and economic development, precisely the vocabulary that the race dimension has absorbed as White-associated. Relatedly, PoliticsWeb is one of the platforms strongly associated with White-stereotyped language in both figures, positioned very close to the business and finance outlets in Fig. 3 . This is analytically interesting because PoliticsWeb is not a business but a political commentary platform that predominantly publishes opinion pieces from right-leaning and opposition voices in South Africa, like the Democratic Alliance (DA). Its strong White-stereotype association likely captures the political and economic worldview of its contributors rather than business coverage. This suggests that the White-stereotype language cluster goes beyond economic reporting into a particular aspect of political discourse. eNCA Online is another notable case in that region, ranking fifth on the White-stereotype difference score in Fig. 4 despite being a mainstream broadcast news outlet rather than a specialist business or political commentary platform. It is also considerably closer to the finance and business outlets compared to other general news broadcasters like SABC. Platforms in the high-Black, low-White quadrant include Daily Sun, Weekend Argus, Cape Argus, Cape Times, SA News, and City Press. Daily Sun is South Africa’s most widely read tabloid, targeting working-class Black South African readers in urban centres, and its strong Black stereotype association likely reflects the lived realities of the communities it covers. Compared with Sunday World, both are Black-owned publications historically targeting Black South African audiences, yet Sunday World is less Black-stereotyped, while Daily Sun is the most strongly Black-stereotyped platform in the dataset. Weekend Argus, Cape Argus, and Cape Times are all Cape Town-based Independent Media outlets, and their positioning likely captures the prominence of township, housing, community, and poor socioeconomic coverage in a city with one of South Africa’s most racially segregated socioeconomic and residential landscapes. SA News – the South African government’s official news agency – is in the high Black region and shows the lowest White stereotype association of any platform, likely because government communication in South Africa is heavily oriented toward service delivery, social welfare, crime, and developmental policy, the very stereotypes the race dimension identifies as Black-associated. These patterns should also be read in the context of COVID-19, which intensified existing inequalities and made poverty, vulnerability, and unequal access to services even more salient in public discourse. One of the most interesting results is The Conversation, which ranks among the highest on both stereotype dimensions simultaneously. As an academic journalism outlet commissioning pieces from university researchers, its high scores on both stereotypes likely capture the breadth of its coverage – it engages with the language of wealth and development on the White-associated side, and the language of poverty, inequality, and marginalisation on the Black-associated side. Similarly, Bhekisisa, a health journalism outlet, is in the high-Black, high-White region. In substantive terms, the language in both cases appears less one-sidedly racialised: it produces both the wealth, high-status, development-oriented register associated with whiteness and the crisis or deprivation-oriented register associated with blackness. Figure 3 also shows outlets that score relatively low on both stereotypes. SABC Online News is the clearest example. It shows a modest association with both centroids and does not lean strongly toward either side. This makes it one of the least stereotyped platforms in the sample. As the public broadcaster mandated to serve all South Africans, this likely captures editorial choices against the kind of sectoral, racially biased language that characterises more specialised outlets. The Times, The Citizen and, to a lesser extent, Daily Dispatch and The Herald are also nearer the middle of the plot than the more clearly polarised outlets. News24 is also relatively low on the White stereotype side, although it leans somewhat more toward the Black stereotype centroid than SABC. The implication is that some high-volume national outlets appear less semantically saturated with either stereotype category, even when they are not entirely neutral. The technology-oriented platforms (e.g., MyBroadband and BusinessTech) are low on the Black stereotype axis, but they are not neutral overall. Both lean toward the White stereotype centroid, especially in the difference plot (Fig. 4 ). This does not suggest an absence of stereotyping, but rather a discursive environment more aligned with business, infrastructure, consumer technology, and aspirational middle-class frames than with the cluster of Black-coded social vulnerability terms. Thus, they can be more accurately described as white-leaning with comparatively weak Black-stereotype association. Taken together, the figures reveal that all platforms reveal a meaningful association with both stereotype sets. White-leaning platforms are overwhelmingly those serving business, finance, and professional-class audiences; Black-leaning platforms are often those targeted at the Black working-class South African, focused on government services, or metropolitan newspapers in unequal urban settings where township life is especially salient; and the least stereotyped platforms are often those with a broad national and public mandate or those that tend to cover outside of socioeconomic phenomena (e.g., tech-oriented platforms). 5. Conclusion This study set out to identify which South African news platforms contribute to the racial biases documented in previous research (Ohamadike et al., 2025a ). Using word2vec embeddings trained on resamples of 3,900 articles across 39 eligible platforms drawn from a corpus of 27,140 COVID-19 vaccination articles, and adapting the speaker landscape method (Schuld et al., 2024 ) to embed each outlet as a vector, we measured how closely each platform’s language associates with the Black and White racial stereotype vocabularies identified in the previous study. The results of the present study replicated the original findings – socioeconomic stereotypes projected cleanly onto the race dimension, showing race bias and correlated strongly with human judgements; thus, validating the results. The analysis revealed the institutions behind the aggregate bias. Business, finance, and investment outlets (Moneyweb, Financial Mail, Business Day, and Fin24) were the most strongly White-stereotyped, their language capturing wealth, development, and economic agency that the race dimension identifies as White-associated. At the other end, outlets oriented toward working-class Black audiences, government service communication, and metropolitan coverage in unequal urban settings (Daily Sun, Cape Argus, Cape Times, and SA News) were the most strongly Black-stereotyped, their language drawing on service delivery, crime, welfare, poverty, and inequality. Importantly, all platforms showed meaningful association with both stereotype sets, which means race bias is not confined to a subset of outlets but is a feature of South African news across the board. What varies is its direction, and that direction maps almost perfectly onto the institutional character of each outlet. This pattern corresponds with the account of media racism developed in the literature. Earlier work showed that racism persists after apartheid mainly through subtle, inferential means, narrative framing, source selection, and editorial norms that lack racial transformation in a concentrated, capitalist, market-driven media (Braude, 1999 ; Berger, 2001 ; Durrheim et al., 2005 ; MMP, 1999, 2006; Govenden, 2022 ; Radebe, 2022 , 2023 ; Duncan, 2014 ; Wasserman, 2010 ; Ohamadike et al., 2025a , b ). Our findings extend that literature by demonstrating, computationally and at scale, that such bias is institutionally patterned. Business outlets are White-stereotyped mainly because they cover the world of capital, which in South Africa remains a predominantly White world; the community, metropolitan, and government outlets are Black-stereotyped largely because they cover the world of poverty, inequality, public and welfare services, which remains a predominantly Black experience. This is the “inferential racism” Govenden ( 2022 ) described – subtle bias embedded in routine choices about what to cover and whose reality to centre, rather than in overt racial animus. This is vital because news representations can amplify racism by shaping public perceptions, narratives, and actions through their wide reach, especially during a public health crisis (Green, 2012 ). These findings have implications for media accountability. The method introduced here functions as a scalable institutional audit – it identifies which outlets most warrant scrutiny and in what direction, without requiring manual analysis of thousands of articles. For practitioners, the results highlight that racial bias in news language often operates through the accumulated weight of coverage focus and audience orientation. Business and finance outlets, in particular, may not perceive themselves as racially biased, yet their language consistently associates with the White end of the socioeconomic spectrum. Addressing this does not require abandoning their editorial focus; it requires greater attention to whose business and economic experiences are represented within it. More broadly, the methodology provides media regulators, researchers, and civil society organisations a replicable tool for monitoring racial bias across the news ecosystem over time. This study has several limitations. The corpus is restricted to COVID-19 vaccination coverage, so the distribution of racial associations across platforms may differ in general news contexts. The 100-article cap, while necessary to equalise platform representation, means platform vectors are derived from samples rather than each outlet’s full output. Word2vec is also limited in how it accounts for context: it represents all instances of a word with a single vector, whereas contextualised models such as BERT (Devlin et al., 2018) model each instance separately and may produce different results (Arseniev-Koehler & Foster, 2022 ; Schuld et al., 2024 ). We used Word2Vec for consistency with Ohamadike et al. ( 2025a ) and for simplicity. The analysis also focuses on Black and White racial groups, as these dominate South African news discourse and are the groups for which validated stereotype vocabularies exist. Finally, as with all embedding-based analyses, the scores measure language patterns that cannot fully account for editorial intent. What this study ultimately establishes is that the aggregate racial bias in South African news is a differentiated phenomenon with identifiable institutional sources. Word embeddings, applied at the platform level, are a tractable and scalable tool for locating them, and for holding specific outlets accountable in ways that aggregate analyses of the media as a whole cannot. Abbreviations BERT Bidirectional Encoder Representations from Transformers DA Democratic Alliance HRC Human Rights Commission HTML Hypertext Markup Language ICC Intraclass Correlation Coefficients MMA Media Monitoring Africa MMP Media Monitoring Project NSFAS National Student Financial Aid Scheme SA South Africa SABC South African Broadcasting Corporation SASSA South African Social Security Agency TRC Truth and Reconciliation Commission WEAT Word Embedding Association Test Declarations Ethics approval and consent to participate Not applicable Consent for publication Not applicable Availability of data and materials The code and embeddings used for the platform-level analyses in this paper are available at https://doi.org/10.6084/m9.figshare.31940451. Schuld et al. (2024)’s speaker landscapes, along with the code to construct and analyse them, are available at https://github.com/mariaschuld/speaker-landscapes. The dataset is available from the authors upon reasonable request and with permission of Media Monitoring Africa. Competing interests The authors declare that they have no competing interests. Funding This research was supported by funding from the International Society of Political Psychology (ISPP), the National Research Foundation of South Africa (Grant UID: 137755), the South African Centre for Digital Language Resources (SADiLaR) (Grant OR-AAALV), and the University of Johannesburg. We also gratefully acknowledge the Swiss National Science Foundation for their financial support of this research (grant number: 10001AL_205032). Author contributions NO analysed the data, conceptualised and wrote the paper. KD and MP designed the study and provided guidance. All authors contributed to, reviewed, and approved the paper. Acknowledgements We thank Media Monitoring Africa for kindly granting us access to their vaccination news archive. We are also grateful to Dr. Ross Harvey for asking the questions that informed and motivated this research. References Arseniev-Koehler, A., & Foster, J. G. (2022). Machine learning as a model for cultural learning: Teaching an algorithm what it means to be fat. Sociological Methods & Research , 51(4), 1484–1539. Berger, G. (2001). Problematising race for journalists: Critical reflections on the South African Human Rights Commission Inquiry into media racism. Critical Arts , 15(1-2), 69-96. Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in Neural Information Processing Systems , 29. Braude, C. (1999) ‘Cultural bloodstains: Towards understanding the legacy of apartheid and the perpetuation of racial stereotypes in the contemporary South African media’, in Human Rights Commission, Interim Report into Racism in the Media. Johannesburg: South African Human Rights Commission. Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science , 356(6334), 183–186. https://doi.org/10.1126/science.aal4230 Duncan, J. (2014). “It’s not just the unions that are cut off from people, but the media too”: Reconstituting South Africa’s mediated public sphere. Acta Academica , 46(1), 73–97. Durrheim, K., Quayle, M., Whitehead, K., & Kriel, A. (2005). Denying racism: Discursive strategies used by the South African media. Critical Arts , 19(1-2), 167-186. Govenden, P. (2022) Does Black Economic Empowerment Ownership Matter? A Decolonial Analysis of “Black Visibility” in South Africa’s Print Media Content, 1994–2014, African Journalism Studies , 43:2, 1-20, DOI: 10.1080/23743670.2022.2096090 Green, M. (2012). Media representations of gratuitous violence in South Africa (Doctoral dissertation). Human Rights Commission (HRC). (1999). Interim report of the inquiry into racism in the media. South African Human Rights Commission, Johannesburg. Kozlowski, A. C., Taddy, M., & Evans, J. A. (2019). The geometry of culture: Analyzing the meanings of class through word embeddings. American Sociological Review , 84(5), 905–949. https://doi.org/10.1177/0003122419877135 Media Monitoring Project (1999) ‘The news in black and white: An investigation into racial stereotyping in the media’, in Human Rights Commission, Interim Report into Racism in the Media. Johannesburg: South African Human Rights Commission. Media Monitoring Project (MMP). (2006). Revealing race: An analysis of the coverage of race and xenophobia in the South African print media. Media Monitoring Africa. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781. Ohamadike, N., Durrheim, K., & Primus, M. (2025a). The news in black and white: word embeddings quantify racism in South African news. EPJ Data Science , 14 (1), 83. Ohamadike, N., Durrheim, K., & Primus, M. (2025b). Whose voice matters? Word embeddings reveal identity bias in news quotes. EPJ Data Science , 14 (1), 30. Pityana, N. B. (2000). South Africa's inquiry into racism in the media: the role of national institutions in the promotion and protection of human rights. African Affairs , 99(397), 525-532. Posel, D. (1990). Symbolizing violence: State and media discourse in TV coverage of township protest, 1985–7. In Political violence and the struggle in South Africa (pp. 154-171). London: Palgrave Macmillan UK. Radebe, M. J. (2022). Framing the poor during COVID-19: Townships and informal settlements in South African online news. Communicatio , 48(2), 41–62. Radebe, M. J. (2023). Constructing Hegemony: The South African Commercial Media and the (Mis) Representation of Nationalisation. Taylor & Francis. Řehůřek, R., & Sojka, P. (2011). Gensim — Python framework for vector space modelling. NLP Centre, Masaryk University. Schuld, M., Durrheim, K., & Mafunda, M. (2024). Speaker landscapes: Machine learning opens a window on the everyday language of opinion. Communication Methods and Measures , 18(4), 315-331. Talbot, H., & Durrheim, K. (2012). Racial stereotyping in South African discourse. South African Journal of Psychology , 42(3), 329–340. Truth and Reconciliation Commission of South Africa (TRC) (1998) Report of the Truth and Reconciliation Commission: Volume Four, Chapter Six. [Online]. Institutional Hearing. https://www.justice.gov.za/trc/report/finalreport/Volume%204.pdf. Accessed 25 March 2026. Van Loon, A., & Freese, J. (2023). Word embeddings reveal how fundamental sentiments structure natural language. American Behavioral Scientist , 67(2), 175–200. Wasserman, H. (2007). Is our media (still) racist? Global Media Journal: African Edition, 1(1). https://doi.org/10.5789/1-1-53 Wasserman, H. (2010). Tabloid journalism in South Africa: True story! Indiana University Press. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9327726","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":619838206,"identity":"7a8f31e9-7c66-4088-bd0d-04545c34e95e","order_by":0,"name":"Nnaemeka Ohamadike","email":"data:image/png;base64,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","orcid":"","institution":"University of Johannesburg","correspondingAuthor":true,"prefix":"","firstName":"Nnaemeka","middleName":"","lastName":"Ohamadike","suffix":""},{"id":619838207,"identity":"21936c13-2070-4db5-b9c9-3805cf010f83","order_by":1,"name":"Kevin Durrheim","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Kevin","middleName":"","lastName":"Durrheim","suffix":""},{"id":619838208,"identity":"071d7476-6b85-44af-ae42-468a8ee7cc20","order_by":2,"name":"Mpho Primus","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Mpho","middleName":"","lastName":"Primus","suffix":""}],"badges":[],"createdAt":"2026-04-05 17:23:23","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9327726/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9327726/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106474406,"identity":"e4c215b5-7e92-4bab-a302-baeec56ca5fb","added_by":"auto","created_at":"2026-04-09 02:40:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1933404,"visible":true,"origin":"","legend":"\u003cp\u003eCosine similarity of socioeconomic stereotypes with the Black–White race dimension, averaged across 10 resampled embeddings. Error bars represent cross-embedding standard error.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-9327726/v1/d224bb7c0f8ce0b8bc1b4421.png"},{"id":106474443,"identity":"6c9d4564-1aab-4d4f-9d83-b2c66a67155e","added_by":"auto","created_at":"2026-04-09 02:40:59","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1342199,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation between word embedding cosine similarity scores and human ratings for socioeconomic stereotypes.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-9327726/v1/745aa6358bcd7971e8831c85.png"},{"id":106474408,"identity":"22474812-659d-427c-8dcb-d9f388ecd678","added_by":"auto","created_at":"2026-04-09 02:40:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1017669,"visible":true,"origin":"","legend":"\u003cp\u003eScatter plot of absolute platform associations with White (y-axis) and Black (x-axis) stereotype centroids, averaged across 10 resampled embeddings. Dashed lines indicate cross-platform means. Error bars represent cross-embedding standard error.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-9327726/v1/aedd388370139eedb74cf5f9.png"},{"id":106474388,"identity":"a7346736-2203-4b52-8aaa-d7d0d480848e","added_by":"auto","created_at":"2026-04-09 02:40:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1365693,"visible":true,"origin":"","legend":"\u003cp\u003eRanked bar chart of the White minus Black stereotype association difference score per platform. Positive values (orange) indicate greater White stereotype association; negative values (blue) indicate greater Black stereotype association. Error bars represent cross-embedding standard error.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-9327726/v1/9935ac035e11c9213d64ea50.png"},{"id":107480373,"identity":"588a4db8-5f6e-4eca-ae93-bca6042c7139","added_by":"auto","created_at":"2026-04-22 02:09:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5253324,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9327726/v1/cb24f489-8259-4412-b938-b254228d066e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Tracing Bias to Its Sources: A Word Embedding Audit of Racism in South African News Outlets","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSouth Africa has struggled with racism in its news media, a struggle that predates democracy and continues to shape how different racial groups are represented in public discourse today. Apartheid-era news platforms were instruments of state-sanctioned racism \u0026ndash; controlled largely by white-owned corporations, they legitimised white supremacy, spread propaganda, actively suppressed or marginalised Black voices, and were a tool for the apartheid government to control information (Braude, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Berger, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2001\u003c/span\u003e; Durrheim et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). The South African Truth and Reconciliation Commission later documented the media\u0026rsquo;s complicity in gross human rights violations during this period (TRC, 1998).\u003c/p\u003e \u003cp\u003eThe transition to democracy in 1994 brought policy reforms and efforts to diversify newsrooms, but the underlying patterns of racism were harder to dislodge. Investigations by the South African Human Rights Commission (HRC, 1999) and the Media Monitoring Project (MMP, 1999, 2006) found racism in post-apartheid news coverage, with Black communities frequently framed through stereotypes of criminality, poverty, and social disorder. More recently, Govenden (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) showed that despite significant changes in media ownership since 1994, \u0026ldquo;inferential racism\u0026rdquo; continued to shape representations of Black South Africans in print media through to 2014.\u003c/p\u003e \u003cp\u003eFor most of this history, evidence of media racism came from manual qualitative analyses. These approaches have produced valuable insights but are limited by relying on individual interpretation, are difficult to replicate, and are poorly suited to the scale of analysis needed to draw representative conclusions across dozens of outlets and thousands of articles (Kozlowski et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). They are also better at identifying explicit bias than the recent subtler forms of racism embedded in news language (see Govenden, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWord embeddings provide a way around these limitations. They represent words as vectors in a high-dimensional space, where the distance between vectors encodes semantic association based on how words co-occur in a large corpus. Crucially, embedding models are not programmed with any assumptions about race \u0026ndash; they learn associations from the text they are trained on. This makes them well-suited to detecting implicit bias: if a news corpus consistently places \u0026ldquo;Black\u0026rdquo; near words like \u0026ldquo;crime\u0026rdquo; and \u0026ldquo;poverty,\u0026rdquo; the model will learn and reflect that from its training data. Researchers have used this property to study ethnic and gender bias in historical corpora (Garg et al., 2018), human-like biases encoded in language (Caliskan et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), gender stereotypes and debiasing (Bolukbasi et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), broad cultural dimensions such as class and race (Kozlowski et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), obesity-related schemata in news (Arseniev-Koehler \u0026amp; Foster, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), and the sentiment structures underlying natural language (Van Loon \u0026amp; Freese, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOhamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) used this computational approach to examine racism in 27,140 South African news articles from 76 outlets. They found racism in the coverage of socioeconomic phenomena across corpora, with \u0026ldquo;Black\u0026rdquo; frequently associated with crime, poverty, and poor socioeconomic conditions, and \u0026ldquo;White\u0026rdquo; with safety, wealth, civic virtue, and better socioeconomic conditions. These results were validated in many ways, including through human judgment, but they stop short of identifying which specific news platforms are responsible for driving it.\u003c/p\u003e \u003cp\u003eThis is the gap the present research addresses. Knowing that bias exists in aggregate tells us that something is wrong; knowing which platforms produce it tells us where to intervene. Using the same news corpus and embedding approach from Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e), we adapt the speaker landscape method (Schuld et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) to embed each news outlet as a vector based on its language and measure that vector\u0026rsquo;s association with racial stereotypes. This allows us to identify those whose language most closely mirrors the stereotypes previously found in the aggregate corpus, providing, for the first time, a computational outlet-level account of institutional responsibility for racial bias.\u003c/p\u003e \u003cp\u003eIn what follows, we review the literature on racism in South African news and on the use of word embeddings to study racial bias, before describing the methodology, results, and implications of the analysis.\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Racism in South African News\u003c/h2\u003e \u003cp\u003eThe story of racism in South African news cannot be told without starting with apartheid, because the structures, practices, and assumptions of that era did not disappear when the regime ended. During apartheid, the mainstream press was predominantly owned and controlled by white corporations aligned with the ruling government. News coverage systematically dehumanised Black South Africans \u0026ndash; depicting resistance movements as threats to order, portraying Black communities through imagery of violence and primitivism, and framing state violence against protesters as a legitimate response to chaos (Posel, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e1990\u003c/span\u003e; Braude, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). Posel (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e1990\u003c/span\u003e) documented how the South African Broadcasting Corporation (SABC) news in the 1980s described township uprisings in language that cast Black protesters as mindless mobs, effectively justifying the apartheid state\u0026rsquo;s violent suppression of dissent. The Truth and Reconciliation Commission later confirmed the extent of the media\u0026rsquo;s complicity in apartheid-era human rights violations (TRC, 1998).\u003c/p\u003e \u003cp\u003eThese patterns persisted after apartheid. Just five years into democracy, the Human Rights Commission launched a formal investigation into racism in the South African media following complaints that outlets continued to produce racially biased coverage (HRC, 1999; Pityana, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2000\u003c/span\u003e). The Media Monitoring Project\u0026rsquo;s contribution to that inquiry documented recurring racial themes in news content, such as \u0026lsquo;Blacks are criminals\u0026rsquo;, and the investigation established that racism had not been resolved by political transition (MMP, 1999; Durrheim et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; Pityana, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2000\u003c/span\u003e). A follow-up MMP report years later found much the same: coverage of race and xenophobia was characterised by dramatic, stereotype-reinforcing framing that lacked context, and marginalised groups continued to be depicted in ways that reproduced their social exclusion (MMP, 2006; Wasserman, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2007\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eGovenden (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) examined the longer arc of this problem, tracking print media content from 1994 to 2014 and finding that, despite the growth of Black media ownership during this period, \u0026ldquo;inferential racism\u0026rdquo; remained the dominant mode \u0026ndash; bias embedded in framing choices, source selections, and narrative structures of everyday journalism. The persistence of these patterns is partly a product of structural factors that outlast formal policy change. Commercialised, market-driven press tends to reproduce the assumptions and interests of dominant economic groups (Duncan, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Radebe, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWasserman (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2010\u003c/span\u003e) points to the influence of Western journalistic models imported through South Africa\u0026rsquo;s colonial and apartheid-era ties to Britain, which carry their own assumptions about whose voices are authoritative and which events are newsworthy. Newsroom diversification in terms of race and gender has improved, but has not translated into transformed content, often prioritising the worldview of elites in a capitalist framework (Duncan, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Govenden, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Radebe, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wasserman, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Testing this computationally, Ohamadike et al. (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e) found that South African news journalists often quoted White speakers on global and technical vaccination topics, while confining Black speakers, especially South African ones, to localised coverage.\u003c/p\u003e \u003cp\u003eWhat is missing from this body of research is an account of institutional responsibility. Studies have established that racial bias exists in South African news. But they have largely treated the media as a collective entity, leaving unanswered the question of which specific platforms contribute most to these patterns. This matters for two reasons. First, South Africa\u0026rsquo;s news landscape is not homogeneous: it includes public broadcasters, commercial print and digital outlets, specialised business publications, and community media, with different ownership structures, editorial cultures, and audience bases. Second, targeted accountability requires knowing who is responsible. Aggregate findings establish that a problem exists; they cannot, on their own, inform the institutional interventions \u0026ndash; in editorial policy, journalist training, and media regulation \u0026ndash; needed to address it.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Word Embeddings as a Tool for Detecting Racism\u003c/h2\u003e \u003cp\u003eWord embeddings have become one of the more productive tools for studying implicit bias in large text corpora. The core idea is that words that appear in similar contexts will have similar vector representations, and the geometric relationships between these vectors encode semantic associations based on how words co-occur in the text. A word\u0026rsquo;s position in the embedding space reflects the company it keeps in the corpus, making the space a kind of map of the cultural associations encoded in the language.\u003c/p\u003e \u003cp\u003eCaliskan et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) provided an influential early demonstration. Using GloVe embeddings trained on web text, they showed that European-American names were more closely associated with positive attributes than African-American names in the embedding space, mirroring the results of Implicit Association Test studies in social psychology. Garg et al. (2018) extended this work by tracking how ethnic stereotypes in American English shifted across the twentieth century, training GloVe embeddings on historical text corpora. They found that negative traits associated with Chinese immigrants in the early 1900s gave way to more positive associations by the century\u0026rsquo;s end, a shift that correlated with Asian-American population growth. Their work demonstrated that embeddings can serve as a longitudinal instrument, not merely detecting bias but tracking how it changes over time.\u003c/p\u003e \u003cp\u003eKozlowski et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) improved the specific methodological approach most relevant to the present study. They showed that, for race, a semantic dimension defined by the difference between a cluster of \u0026ldquo;Black\u0026rdquo; terms and a cluster of \u0026ldquo;White\u0026rdquo; terms allows any word in the embedding to be scored for how closely it associates with each pole. They found, using Google News embeddings, that jazz and rap were more strongly associated with the Black pole, and opera and bluegrass with the White pole.\u003c/p\u003e \u003cp\u003eIn the South African context, Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) trained word2vec embeddings on a corpus of South African vaccination news and found that the race dimension reliably separated socioeconomic stereotypes, with poverty, crime, and welfare-associated terms clustering at the Black pole and wealth, safety, and terms representing good socioeconomic conditions at the White pole, and the results corresponded with human judgments.\u003c/p\u003e \u003cp\u003eThe speaker landscape method, introduced by Schuld et al. (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), allows for extending this analysis from words to people and institutions. By inserting speaker or source identifiers as tokens into the training corpus alongside the text associated with each actor, the method produces a vector for each speaker or source that captures their surrounding language. This vector can then be subjected to the same cosine similarity analyses used for words \u0026ndash; locating actors in the embedding space, measuring their proximity to ideological or thematic issues, and comparing groups. Ohamadike et al. (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e) used this method to study identity bias at the speaker level in South African news quotes, finding that \u0026lsquo;journalists privilege White voices as more authoritative and connected to global and technical vaccination discourse but confine black voices to primarily localised contexts\u0026rsquo; (p.1). The present study adapts this framework to the platform level, treating each news outlet as the unit of analysis and measuring its vector\u0026rsquo;s association with racial stereotypes.\u003c/p\u003e \u003cp\u003eOverall, embeddings remain an effective tool for studying implicit bias at scale, and when applied at the platform level, one of the few methods capable of attributing aggregate patterns of media bias to their institutional sources.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Data\u003c/h2\u003e \u003cp\u003eThis study builds on Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e), using the same corpus of 27,140 South African news articles from 76 outlets covering COVID-19 vaccination. Articles span from January 2020 to May 2023 and were collected by Media Monitoring Africa (MMA). The same preprocessing pipeline from the original study was applied: all text was lowercased, punctuation and HTML tags removed, and no stopwords were excluded.\u003c/p\u003e \u003cp\u003eTo attribute bias to its institutional sources, each article was prefixed with a single-token label representing its news platform (e.g., \u003cem\u003esource_news24\u003c/em\u003e), enabling word2vec to learn a vector for each outlet based on its language. This approach was adapted from Schuld et al. (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u0026rsquo;s speaker landscape. It allows each platform to acquire a unique vector representing its overall linguistic profile, used in coverage and can then be associated with stereotype centroids. Only platforms with at least 100 articles were included to ensure representationally stable vectors. For platforms exceeding this threshold, 100 articles were sampled per run (described below), equalising representation across outlets and ensuring that bias estimates reflect language use rather than publication volume. This yielded 39 eligible platforms and 3,900 articles in each resampled data set for training each of the 10 embeddings.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Word Embeddings\u003c/h2\u003e \u003cp\u003eWord embeddings were trained using the Skip-Gram architecture of the word2vec algorithm (Mikolov et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) implemented in Python\u0026rsquo;s Gensim library (Řehůřek \u0026amp; Sojka, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). Word2vec is a neural embedding method that learns vector representations for words by predicting the context in which they appear. When trained on large corpora, these representations encode the semantic relationships present in the text, including latent biases absorbed from human language (Caliskan et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Kozlowski et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). We used word2vec word embeddings for simplicity and in line with the previous research.\u003c/p\u003e \u003cp\u003eTo account for stochastic variation in training and produce stable estimates, we trained an ensemble of 10 embeddings with different random seeds, resampling 100 articles per platform in each run. This resampling procedure serves two purposes: it controls for the disproportionate influence of high-volume outlets on the embedding space and allows uncertainty in our estimates to be quantified as cross-run variability. The hyperparameter settings used in training are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Compared with Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e), the vector dimensionality was reduced from 250 to 200 and the minimum word count from 5 to 3, to accommodate the smaller per-run corpus.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHyperparameter settings for word2vec training\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParameter Name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMinimum word count (m)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWindow size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArchitecture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSkip-Gram (sg\u0026thinsp;=\u0026thinsp;1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative samples\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVector dimension size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e200\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEpochs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of bootstrap embeddings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of worker threads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Constructing the Race Dimension\u003c/h2\u003e \u003cp\u003eWe constructed a term-based race dimension using the same Black/White terms employed and validated by Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The authors validated these terms through different means, including by projecting individual-specific Black and White speaker names, Euro-American and African country names, and non-individual-specific Black and White names onto the race dimension. They found that these names were racialised \u0026ndash; the model and race dimension, in most cases, learned to associate the names with the relevant region (Black or White).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eWord Pairs used to Construct Term-based Race Dimension\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWord Pairs\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eblack-white\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eblacks-whites\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eafro-anglo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eafrican-european\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eafrican-american\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eindigenous-settler\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ezulu-afrikaner\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFor each group in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, we computed a centroid \u0026ndash; the average of all corresponding word vectors across all 10 embeddings. The race dimension is the vector connecting the two centroids (Black minus White), normalised to unit length. Positive cosine similarity with this direction indicates association with the Black pole; negative similarity indicates association with the White pole.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Analysis 1: Replicating Socioeconomic Bias\u003c/h2\u003e \u003cp\u003eTo establish that the socioeconomic biases identified in the previous study persist in the platform-level corpus, we projected the same set of socioeconomic stereotype keywords onto the race dimension. These stereotypes were drawn from three sources: the Media Monitoring Project (MMP, 1999), Talbot and Durrheim (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2012\u003c/span\u003e), and ChatGPT (2024a), and represent socially constructed and historically grounded associations tied to race in South Africa (see Ohamadike et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e for the full list and sources).\u003c/p\u003e \u003cp\u003eFor each stereotype that appeared in at least 5 of the 10 embeddings, we computed its cosine similarity with the race dimension in each embedding and averaged these scores across embeddings. The resulting mean cosine similarity score indicates each term\u0026rsquo;s directional association with the Black or White pole, with standard error across embeddings quantifying estimate uncertainty.\u003c/p\u003e \u003cp\u003eTo formally test whether White stereotypes cluster more strongly with the White pole and Black stereotypes with the Black pole, we applied the Word Embedding Association Test (WEAT) (Caliskan et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). The WEAT effect size is calculated as Cohen\u0026rsquo;s \u003cem\u003ed\u003c/em\u003e \u0026ndash; the standardised mean difference in cosine similarities between the two stereotype groups \u0026ndash; divided by their pooled standard deviation (Ohamadike et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). Statistical significance was evaluated with a permutation test over 10,000 iterations, randomly reassigning group labels to estimate the null distribution of the test statistic.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Analysis 2: Human Validation of Bias Scores\u003c/h2\u003e \u003cp\u003eTo validate whether the cosine similarity scores derived from the platform-level embeddings reflect the racial associations held by South African participants, we correlated the stereotype scores from Analysis 1 with human ratings collected in Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). In that study, 26 South African participants (21 Black, 4 White, 1 Indian) rated each socioeconomic stereotype on a 5-point Likert scale: 1 (Strongly Associated with White) to 5 (Strongly Associated with Black). Inter-rater reliability was evaluated using Krippendorff\u0026rsquo;s alpha with bootstrapped 95% confidence intervals and intraclass correlation coefficients (ICC), both scaled from 0 to 1. The results indicated moderate agreement among individual raters (Krippendorff α\u0026thinsp;=\u0026thinsp;0.54, 95% CI [0.46, 0.60]; ICC(A,1)\u0026thinsp;=\u0026thinsp;0.55, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003cp\u003eFor the present study, we averaged human ratings per stereotype and merged them with the corresponding mean cosine similarity scores. We then computed Pearson\u0026rsquo;s \u003cem\u003er\u003c/em\u003e between the two sets of scores. A significant positive correlation would indicate that the platform-level embeddings capture racial associations in ways that accord with human judgement, validating the bias scores.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Analysis 3: Platform-Level Bias Measurement\u003c/h2\u003e \u003cp\u003eThe central analysis of this paper measures the degree to which each news platform\u0026rsquo;s learned vector is associated with Black and White stereotype language. We adapted the speaker landscape approach (Schuld et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) in which news platforms are positioned relative to their language use.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.6.1 Stereotype Centroids\u003c/h2\u003e \u003cp\u003eFor each of the 10 embeddings, we computed a White stereotype centroid as the mean vector of White stereotype terms present in that embedding that projected cleanly onto their theorised side of the race dimension, and analogously for Black stereotypes. Using embedding-specific centroids rather than a single grand centroid preserves within-embedding coherence: each platform vector is compared to centroids derived from the same vector space, avoiding cross-embedding vector incompatibility.\u003c/p\u003e \u003cp\u003eNot all stereotypes project cleanly onto their theorised side of the race dimension and may introduce noise into centroid construction. We therefore retained only those stereotypes whose mean cosine similarity from Analysis 1 was directionally consistent with their theorised category (Black stereotypes with positive mean projection; White stereotypes with negative mean projection). Terms projecting to the opposite pole or at zero were excluded, although robustness checks including all terms did not materially change the results.\u003c/p\u003e \u003cp\u003eAcross embeddings, each platform\u0026rsquo;s cosine similarity with these centroids was averaged to produce stable estimates, with standard error quantifying cross-embedding variability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.6.2 Platform Association Scores\u003c/h2\u003e \u003cp\u003eFor each embedding, we computed two cosine similarity scores: the similarity of the platform vector to the embedding\u0026rsquo;s White stereotype centroid and to the Black stereotype centroid. These scores capture the degree to which each platform\u0026rsquo;s language aligns with each set of stereotypes.\u003c/p\u003e \u003cp\u003eThe two platform association scores capture the degree to which each platform\u0026rsquo;s language is associated with each stereotype set (Black/White). We report both scores together to enable a full characterisation of each platform\u0026rsquo;s racial framing. Outlets in the high-White, high-Black quadrant of the resulting space exhibit strong stereotyping across both racial groups; those in the low-low quadrant show minimal stereotypical language. Asymmetric positions reveal directional framing. We also compute a difference score (White similarity minus Black similarity) to rank platforms on the net direction of their racial associations.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4. Results","content":"\u003cp\u003eWe present the results in four parts. We begin by establishing whether the socioeconomic race biases documented by Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) are present in the platform-level corpus, then validate those bias scores against human judgments, and finally report the main analysis attributing racial bias to individual news platforms.\u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Socioeconomic Stereotypes and the Race Dimension\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the projection of the socioeconomic stereotypes onto the race dimension derived from the platform-level embeddings. The x-axis represents each term\u0026rsquo;s cosine similarity with the bias direction, where positive values indicate association with the Black pole and negative values indicate association with the White pole. Error bars reflect the standard error of each term\u0026rsquo;s score across the 10 resampled embeddings.\u003c/p\u003e \u003cp\u003eThe results replicate the core finding from Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) (WEAT\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;1.31, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001): race bias in socioeconomic language is clearly present in our results, with most Black stereotypes projecting on the Black side and most White stereotypes projecting on the White side (WEAT\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;1.04, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Further confirming this replication, cosine similarity scores from the platform-level embeddings correlate strongly with those from the original study (r\u0026thinsp;=\u0026thinsp;0.75, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), showing that despite differences in corpus construction and hyperparameters, both embeddings learned similar racial associations from the same news data. The strongest associations with the Black pole in this study include terms like \u0026ldquo;marginalized,\u0026rdquo; \u0026ldquo;traditional,\u0026rdquo; \u0026ldquo;township,\u0026rdquo; \u0026ldquo;rural,\u0026rdquo; and \u0026ldquo;shack,\u0026rdquo; as well as government welfare identifiers specific to the South African context such as \u0026ldquo;SASSA\u0026rdquo; (the South African Social Security Agency), its former Social Relief of Distress Grant amount of \u0026ldquo;R350,\u0026rdquo; and \u0026ldquo;NSFAS\u0026rdquo; (South African government\u0026rsquo;s National Student Financial Aid Scheme targeted at low-income households). At the other end, the strongest associations with the White pole include White-stereotyped economic and civic terms like \u0026ldquo;advantage,\u0026rdquo; \u0026ldquo;investor,\u0026rdquo; \u0026ldquo;taxpayers,\u0026rdquo; \u0026ldquo;ambitious,\u0026rdquo; and \u0026ldquo;developed.\u0026rdquo;\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Validation Against Human Judgements\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the relationship between the embedding-derived cosine similarity scores from Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and the human ratings collected by Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e), in which 26 South African participants rated each stereotype\u0026rsquo;s association with Black or White on a five-point Likert scale. The correlation between the two sets of scores is positive and highly significant (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.61, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), indicating that the racial biases or associations encoded in the platform-level embeddings reflect the intuitions of South African human raters, validating our results.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Platform-Level Racial Bias\u003c/h2\u003e \u003cp\u003eFigures \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e present the main findings of the study \u0026ndash; the association of each news platform\u0026rsquo;s embedding vector with the White and Black stereotype centroids. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e plots each platform\u0026rsquo;s Black stereotype association (x-axis) against its White stereotype association (y-axis), locating platforms in a two-dimensional bias space. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e ranks platforms by the difference between these two scores (White minus Black), computed per platform and averaged across the 10 resampled embeddings, providing a single directional measure of each outlet\u0026rsquo;s net racial orientation.\u003c/p\u003e \u003cp\u003eThe most immediately noteworthy feature of Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e is that all 39 outlets show meaningful association with both stereotype sets, indicating that race bias is not confined to a subset of outlets but is a feature of South African news language across the board.\u003c/p\u003e \u003cp\u003eThe clearest pattern in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e is the separation of platforms by institutional character. Business, finance, and investment-focused outlets (such as Moneyweb, Financial Mail, Business Day, and Fin24) cluster in the upper-left quadrant, characterised by high White stereotype association and relatively low Black stereotype association, with Moneyweb positioned furthest left on the Black side. The language of these outlets captures the world of investment, corporate ownership, wealth, and economic development, precisely the vocabulary that the race dimension has absorbed as White-associated.\u003c/p\u003e \u003cp\u003eRelatedly, PoliticsWeb is one of the platforms strongly associated with White-stereotyped language in both figures, positioned very close to the business and finance outlets in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This is analytically interesting because PoliticsWeb is not a business but a political commentary platform that predominantly publishes opinion pieces from right-leaning and opposition voices in South Africa, like the Democratic Alliance (DA). Its strong White-stereotype association likely captures the political and economic worldview of its contributors rather than business coverage. This suggests that the White-stereotype language cluster goes beyond economic reporting into a particular aspect of political discourse. eNCA Online is another notable case in that region, ranking fifth on the White-stereotype difference score in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e despite being a mainstream broadcast news outlet rather than a specialist business or political commentary platform. It is also considerably closer to the finance and business outlets compared to other general news broadcasters like SABC.\u003c/p\u003e \u003cp\u003ePlatforms in the high-Black, low-White quadrant include Daily Sun, Weekend Argus, Cape Argus, Cape Times, SA News, and City Press. Daily Sun is South Africa\u0026rsquo;s most widely read tabloid, targeting working-class Black South African readers in urban centres, and its strong Black stereotype association likely reflects the lived realities of the communities it covers. Compared with Sunday World, both are Black-owned publications historically targeting Black South African audiences, yet Sunday World is less Black-stereotyped, while Daily Sun is the most strongly Black-stereotyped platform in the dataset. Weekend Argus, Cape Argus, and Cape Times are all Cape Town-based Independent Media outlets, and their positioning likely captures the prominence of township, housing, community, and poor socioeconomic coverage in a city with one of South Africa\u0026rsquo;s most racially segregated socioeconomic and residential landscapes. SA News \u0026ndash; the South African government\u0026rsquo;s official news agency \u0026ndash; is in the high Black region and shows the lowest White stereotype association of any platform, likely because government communication in South Africa is heavily oriented toward service delivery, social welfare, crime, and developmental policy, the very stereotypes the race dimension identifies as Black-associated. These patterns should also be read in the context of COVID-19, which intensified existing inequalities and made poverty, vulnerability, and unequal access to services even more salient in public discourse.\u003c/p\u003e \u003cp\u003eOne of the most interesting results is The Conversation, which ranks among the highest on both stereotype dimensions simultaneously. As an academic journalism outlet commissioning pieces from university researchers, its high scores on both stereotypes likely capture the breadth of its coverage \u0026ndash; it engages with the language of wealth and development on the White-associated side, and the language of poverty, inequality, and marginalisation on the Black-associated side. Similarly, Bhekisisa, a health journalism outlet, is in the high-Black, high-White region. In substantive terms, the language in both cases appears less one-sidedly racialised: it produces both the wealth, high-status, development-oriented register associated with whiteness and the crisis or deprivation-oriented register associated with blackness.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e also shows outlets that score relatively low on both stereotypes. SABC Online News is the clearest example. It shows a modest association with both centroids and does not lean strongly toward either side. This makes it one of the least stereotyped platforms in the sample. As the public broadcaster mandated to serve all South Africans, this likely captures editorial choices against the kind of sectoral, racially biased language that characterises more specialised outlets. The Times, The Citizen and, to a lesser extent, Daily Dispatch and The Herald are also nearer the middle of the plot than the more clearly polarised outlets. News24 is also relatively low on the White stereotype side, although it leans somewhat more toward the Black stereotype centroid than SABC. The implication is that some high-volume national outlets appear less semantically saturated with either stereotype category, even when they are not entirely neutral.\u003c/p\u003e \u003cp\u003eThe technology-oriented platforms (e.g., MyBroadband and BusinessTech) are low on the Black stereotype axis, but they are not neutral overall. Both lean toward the White stereotype centroid, especially in the difference plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This does not suggest an absence of stereotyping, but rather a discursive environment more aligned with business, infrastructure, consumer technology, and aspirational middle-class frames than with the cluster of Black-coded social vulnerability terms. Thus, they can be more accurately described as white-leaning with comparatively weak Black-stereotype association.\u003c/p\u003e \u003cp\u003eTaken together, the figures reveal that all platforms reveal a meaningful association with both stereotype sets. White-leaning platforms are overwhelmingly those serving business, finance, and professional-class audiences; Black-leaning platforms are often those targeted at the Black working-class South African, focused on government services, or metropolitan newspapers in unequal urban settings where township life is especially salient; and the least stereotyped platforms are often those with a broad national and public mandate or those that tend to cover outside of socioeconomic phenomena (e.g., tech-oriented platforms).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study set out to identify which South African news platforms contribute to the racial biases documented in previous research (Ohamadike et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e). Using word2vec embeddings trained on resamples of 3,900 articles across 39 eligible platforms drawn from a corpus of 27,140 COVID-19 vaccination articles, and adapting the speaker landscape method (Schuld et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) to embed each outlet as a vector, we measured how closely each platform\u0026rsquo;s language associates with the Black and White racial stereotype vocabularies identified in the previous study. The results of the present study replicated the original findings \u0026ndash; socioeconomic stereotypes projected cleanly onto the race dimension, showing race bias and correlated strongly with human judgements; thus, validating the results.\u003c/p\u003e \u003cp\u003eThe analysis revealed the institutions behind the aggregate bias. Business, finance, and investment outlets (Moneyweb, Financial Mail, Business Day, and Fin24) were the most strongly White-stereotyped, their language capturing wealth, development, and economic agency that the race dimension identifies as White-associated. At the other end, outlets oriented toward working-class Black audiences, government service communication, and metropolitan coverage in unequal urban settings (Daily Sun, Cape Argus, Cape Times, and SA News) were the most strongly Black-stereotyped, their language drawing on service delivery, crime, welfare, poverty, and inequality. Importantly, all platforms showed meaningful association with both stereotype sets, which means race bias is not confined to a subset of outlets but is a feature of South African news across the board. What varies is its direction, and that direction maps almost perfectly onto the institutional character of each outlet.\u003c/p\u003e \u003cp\u003eThis pattern corresponds with the account of media racism developed in the literature. Earlier work showed that racism persists after apartheid mainly through subtle, inferential means, narrative framing, source selection, and editorial norms that lack racial transformation in a concentrated, capitalist, market-driven media (Braude, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Berger, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2001\u003c/span\u003e; Durrheim et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; MMP, 1999, 2006; Govenden, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Radebe, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Duncan, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Wasserman, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2010\u003c/span\u003e; Ohamadike et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003eb\u003c/span\u003e). Our findings extend that literature by demonstrating, computationally and at scale, that such bias is institutionally patterned. Business outlets are White-stereotyped mainly because they cover the world of capital, which in South Africa remains a predominantly White world; the community, metropolitan, and government outlets are Black-stereotyped largely because they cover the world of poverty, inequality, public and welfare services, which remains a predominantly Black experience. This is the \u0026ldquo;inferential racism\u0026rdquo; Govenden (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) described \u0026ndash; subtle bias embedded in routine choices about what to cover and whose reality to centre, rather than in overt racial animus. This is vital because news representations can amplify racism by shaping public perceptions, narratives, and actions through their wide reach, especially during a public health crisis (Green, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2012\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese findings have implications for media accountability. The method introduced here functions as a scalable institutional audit \u0026ndash; it identifies which outlets most warrant scrutiny and in what direction, without requiring manual analysis of thousands of articles. For practitioners, the results highlight that racial bias in news language often operates through the accumulated weight of coverage focus and audience orientation. Business and finance outlets, in particular, may not perceive themselves as racially biased, yet their language consistently associates with the White end of the socioeconomic spectrum. Addressing this does not require abandoning their editorial focus; it requires greater attention to whose business and economic experiences are represented within it. More broadly, the methodology provides media regulators, researchers, and civil society organisations a replicable tool for monitoring racial bias across the news ecosystem over time.\u003c/p\u003e \u003cp\u003eThis study has several limitations. The corpus is restricted to COVID-19 vaccination coverage, so the distribution of racial associations across platforms may differ in general news contexts. The 100-article cap, while necessary to equalise platform representation, means platform vectors are derived from samples rather than each outlet\u0026rsquo;s full output. Word2vec is also limited in how it accounts for context: it represents all instances of a word with a single vector, whereas contextualised models such as BERT (Devlin et al., 2018) model each instance separately and may produce different results (Arseniev-Koehler \u0026amp; Foster, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Schuld et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). We used Word2Vec for consistency with Ohamadike et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e) and for simplicity. The analysis also focuses on Black and White racial groups, as these dominate South African news discourse and are the groups for which validated stereotype vocabularies exist. Finally, as with all embedding-based analyses, the scores measure language patterns that cannot fully account for editorial intent.\u003c/p\u003e \u003cp\u003eWhat this study ultimately establishes is that the aggregate racial bias in South African news is a differentiated phenomenon with identifiable institutional sources. Word embeddings, applied at the platform level, are a tractable and scalable tool for locating them, and for holding specific outlets accountable in ways that aggregate analyses of the media as a whole cannot.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eBERT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eBidirectional Encoder Representations from Transformers\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDemocratic Alliance\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHRC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHuman Rights Commission\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHTML\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eHypertext Markup Language\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntraclass Correlation Coefficients\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMMA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMedia Monitoring Africa\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMMP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMedia Monitoring Project\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eNSFAS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eNational Student Financial Aid Scheme\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSouth Africa\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSABC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSouth African Broadcasting Corporation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSASSA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSouth African Social Security Agency\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTRC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTruth and Reconciliation Commission\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eWEAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eWord Embedding Association Test\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code and embeddings used for the platform-level analyses in this paper are available at https://doi.org/10.6084/m9.figshare.31940451. Schuld et al. (2024)’s speaker landscapes, along with the code to construct and analyse them, are available at https://github.com/mariaschuld/speaker-landscapes. The dataset is available from the authors upon reasonable request and with permission of Media Monitoring Africa.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by funding from the International Society of Political Psychology (ISPP), the National Research Foundation of South Africa (Grant UID: 137755), the South African Centre for Digital Language Resources (SADiLaR) (Grant OR-AAALV), and the University of Johannesburg. We also gratefully acknowledge the Swiss National Science Foundation for their financial support of this research (grant number: 10001AL_205032).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNO analysed the data, conceptualised and wrote the paper. KD and MP designed the study and provided guidance. All authors contributed to, reviewed, and approved the paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank Media Monitoring Africa for kindly granting us access to their vaccination news archive. We are also grateful to Dr. Ross Harvey for asking the questions that informed and motivated this research.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eArseniev-Koehler, A., \u0026amp; Foster, J. G. (2022). Machine learning as a model for cultural learning: Teaching an algorithm what it means to be fat. \u003cem\u003eSociological Methods \u0026amp; Research\u003c/em\u003e, 51(4), 1484\u0026ndash;1539.\u003c/li\u003e\n \u003cli\u003eBerger, G. (2001). Problematising race for journalists: Critical reflections on the South African Human Rights Commission Inquiry into media racism. \u003cem\u003eCritical Arts\u003c/em\u003e, 15(1-2), 69-96.\u003c/li\u003e\n \u003cli\u003eBolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., \u0026amp; Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? debiasing word embeddings. \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e, 29.\u003c/li\u003e\n \u003cli\u003eBraude, C. (1999) \u0026lsquo;Cultural bloodstains: Towards understanding the legacy of apartheid and the perpetuation of racial stereotypes in the contemporary South African media\u0026rsquo;, in Human Rights Commission, Interim Report into Racism in the Media. Johannesburg: South African Human Rights Commission.\u003c/li\u003e\n \u003cli\u003eCaliskan, A., Bryson, J. J., \u0026amp; Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. \u003cem\u003eScience\u003c/em\u003e, 356(6334), 183\u0026ndash;186. https://doi.org/10.1126/science.aal4230\u003c/li\u003e\n \u003cli\u003eDuncan, J. (2014). \u0026ldquo;It\u0026rsquo;s not just the unions that are cut off from people, but the media too\u0026rdquo;: Reconstituting South Africa\u0026rsquo;s mediated public sphere. \u003cem\u003eActa Academica\u003c/em\u003e, 46(1), 73\u0026ndash;97.\u003c/li\u003e\n \u003cli\u003eDurrheim, K., Quayle, M., Whitehead, K., \u0026amp; Kriel, A. (2005). Denying racism: Discursive strategies used by the South African media. \u003cem\u003eCritical Arts\u003c/em\u003e, 19(1-2), 167-186.\u003c/li\u003e\n \u003cli\u003eGovenden, P. (2022) Does Black Economic Empowerment Ownership Matter? A Decolonial Analysis of \u0026ldquo;Black Visibility\u0026rdquo; in South Africa\u0026rsquo;s Print Media Content, 1994\u0026ndash;2014, \u003cem\u003eAfrican Journalism Studies\u003c/em\u003e, 43:2, 1-20, DOI: 10.1080/23743670.2022.2096090\u003c/li\u003e\n \u003cli\u003eGreen, M. (2012). Media representations of gratuitous violence in South Africa (Doctoral dissertation).\u003c/li\u003e\n \u003cli\u003eHuman Rights Commission (HRC). (1999). Interim report of the inquiry into racism in the media. South African Human Rights Commission, Johannesburg.\u003c/li\u003e\n \u003cli\u003eKozlowski, A. C., Taddy, M., \u0026amp; Evans, J. A. (2019). The geometry of culture: Analyzing the meanings of class through word embeddings. \u003cem\u003eAmerican Sociological Review\u003c/em\u003e, 84(5), 905\u0026ndash;949. https://doi.org/10.1177/0003122419877135\u003c/li\u003e\n \u003cli\u003eMedia Monitoring Project (1999) \u0026lsquo;The news in black and white: An investigation into racial stereotyping in the media\u0026rsquo;, in Human Rights Commission, Interim Report into Racism in the Media. Johannesburg: South African Human Rights Commission.\u003c/li\u003e\n \u003cli\u003eMedia Monitoring Project (MMP). (2006). Revealing race: An analysis of the coverage of race and xenophobia in the South African print media. Media Monitoring Africa.\u003c/li\u003e\n \u003cli\u003eMikolov, T., Chen, K., Corrado, G., \u0026amp; Dean, J. (2013). Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.\u003c/li\u003e\n \u003cli\u003eOhamadike, N., Durrheim, K., \u0026amp; Primus, M. (2025a). The news in black and white: word embeddings quantify racism in South African news. \u003cem\u003eEPJ Data Science\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(1), 83.\u003c/li\u003e\n \u003cli\u003eOhamadike, N., Durrheim, K., \u0026amp; Primus, M. (2025b). Whose voice matters? Word embeddings reveal identity bias in news quotes. \u003cem\u003eEPJ Data Science\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(1), 30.\u003c/li\u003e\n \u003cli\u003ePityana, N. B. (2000). South Africa\u0026apos;s inquiry into racism in the media: the role of national institutions in the promotion and protection of human rights. \u003cem\u003eAfrican Affairs\u003c/em\u003e, 99(397), 525-532.\u003c/li\u003e\n \u003cli\u003ePosel, D. (1990). Symbolizing violence: State and media discourse in TV coverage of township protest, 1985\u0026ndash;7. In Political violence and the struggle in South Africa (pp. 154-171). London: Palgrave Macmillan UK.\u003c/li\u003e\n \u003cli\u003eRadebe, M. J. (2022). Framing the poor during COVID-19: Townships and informal settlements in South African online news. \u003cem\u003eCommunicatio\u003c/em\u003e, 48(2), 41\u0026ndash;62.\u003c/li\u003e\n \u003cli\u003eRadebe, M. J. (2023). Constructing Hegemony: The South African Commercial Media and the (Mis) Representation of Nationalisation. Taylor \u0026amp; Francis.\u003c/li\u003e\n \u003cli\u003eŘehůřek, R., \u0026amp; Sojka, P. (2011). Gensim \u0026mdash; Python framework for vector space modelling. NLP Centre, Masaryk University.\u003c/li\u003e\n \u003cli\u003eSchuld, M., Durrheim, K., \u0026amp; Mafunda, M. (2024). Speaker landscapes: Machine learning opens a window on the everyday language of opinion. \u003cem\u003eCommunication Methods and Measures\u003c/em\u003e, 18(4), 315-331.\u003c/li\u003e\n \u003cli\u003eTalbot, H., \u0026amp; Durrheim, K. (2012). Racial stereotyping in South African discourse. \u003cem\u003eSouth African Journal of Psychology\u003c/em\u003e, 42(3), 329\u0026ndash;340.\u003c/li\u003e\n \u003cli\u003eTruth and Reconciliation Commission of South Africa (TRC) (1998) Report of the Truth and Reconciliation Commission: Volume Four, Chapter Six. [Online]. Institutional Hearing. https://www.justice.gov.za/trc/report/finalreport/Volume%204.pdf. Accessed 25 March 2026.\u003c/li\u003e\n \u003cli\u003eVan Loon, A., \u0026amp; Freese, J. (2023). Word embeddings reveal how fundamental sentiments structure natural language. \u003cem\u003eAmerican Behavioral Scientist\u003c/em\u003e, 67(2), 175\u0026ndash;200.\u003c/li\u003e\n \u003cli\u003eWasserman, H. (2007). Is our media (still) racist? Global Media Journal: African Edition, 1(1). https://doi.org/10.5789/1-1-53\u003c/li\u003e\n \u003cli\u003eWasserman, H. (2010). Tabloid journalism in South Africa: True story! Indiana University Press.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"word embeddings, racial bias, news outlets, South Africa, outlet accountability, computational social science","lastPublishedDoi":"10.21203/rs.3.rs-9327726/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9327726/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eRace bias in South African news is well documented, but existing research has treated the media as a collective entity, leaving open the question of which specific outlets drive these patterns. This study addresses that gap by tracing racial bias to its institutional sources using word embeddings. We trained an ensemble of 10 Word2Vec models on resamples drawn from a corpus of 27,140 COVID-19 vaccination news articles, with each resample comprising 3,900 articles across 39 outlets. Each outlet was embedded as a vector based on its language, and that vector\u0026rsquo;s association was measured with validated racial stereotype vocabularies. We first found that socioeconomic race bias from a prior study replicates in the present study and correlates strongly with South African human judgments. The outlet-level analysis in this study reveals that business and finance outlets are most strongly associated with White stereotype language, while metropolitan, community newspapers, and government media are most associated with Black stereotype language. All outlets showed meaningful associations with both stereotype sets, suggesting that racial bias is a feature of South African news language rather than the product of a few outliers. The method provides a scalable, replicable framework for auditing racial bias at the institutional level and informing targeted interventions in newsroom practice and media policy.\u003c/p\u003e","manuscriptTitle":"Tracing Bias to Its Sources: A Word Embedding Audit of Racism in South African News Outlets","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-09 02:40:21","doi":"10.21203/rs.3.rs-9327726/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f0e89d85-ae38-4e84-89c9-0375a9b34ff8","owner":[],"postedDate":"April 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-09T02:40:46+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-09 02:40:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9327726","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9327726","identity":"rs-9327726","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}