Large language models (LLMs) rely heavily on web-derived training datasets, yet understanding how filtering and curation decisions affect these datasets remains challenging. This paper presents two complementary datasets designed to enable systematic analysis of LLM training data composition. The first dataset captures domain-level statistics across 96 Common Crawl snapshots, providing baseline data about web content distribution before filtering. The second dataset contains standardized URL information from three major LLM training corpora (C4, Falcon RefinedWeb, and CulturaX), allowing researchers to analyze how different filtering approaches affect content inclusion. By making these datasets publicly available in a consistent format, we aim to (1) facilitate research into training data composition, (2) enable systematic auditing of filtering effects, and (3) support more transparent approaches to dataset development. Our datasets can help researchers investigate questions related to content diversity, source representation, and the impact of different filtering decisions on training data composition. Overall, this work provides a foundation for understanding how curation choices shape the content that ultimately trains widely-deployed language models.

Social media platforms can help people find connection and entertainment, but they can also show potentially abusive content such as insults and targeted cursing. While platforms do remove some abusive content for rule violation, some is considered "margin content" that does not violate any rules and thus stays on the platform. This paper presents a focused analysis of exposure to such content on Twitter, asking (RQ1) how exposure to marginally abusive content varies across Twitter users, and (RQ2) how algorithmically-ranked timelines impact exposure to marginally abusive content. Based on one month of impression data from November 2021, descriptive analyses (RQ1) show significant variation in exposure, with more active users experiencing higher rates and higher volumes of marginal impressions. Experimental analyses (RQ2) show that users with algorithmically-ranked timelines experience slightly lower rates of marginal impressions. However, they tend to register more total impression activity and thus experience a higher cumulative volume of marginal impressions. The paper concludes by discussing implications of the observed concentration, the multifaceted impact of algorithmically-ranked timelines, and potential directions for future work.

With billions of users, algorithmic platforms like Facebook, Twitter, and TikTok are among the most influential communication technologies in the history of civilization. These platforms offer society a mixed bag of benefits and harms, but much remains ambiguous about their true dynamics, effects, and interworkings. As a step toward improved accountability, this document outlines a dissertation that inspects (i.e. audits) algorithmic platforms as mass communication flows. Broadly, I aim to characterize the temporal dynamics, content quality, and sustainability of four different platforms (Apple News, Google Ads, Twitter, and Facebook).

While algorithm audits are growing rapidly in commonality and public importance, relatively little scholarly work has gone toward synthesizing prior work and strategizing future research in the area. This systematic literature review aims to do just that, following PRISMA guidelines in a review of over 500 English articles that yielded 62 algorithm audit studies. The studies are synthesized and organized primarily by behavior (discrimination, distortion, exploitation, and misjudgement), with codes also provided for domain (e.g. search, vision, advertising, etc.), organization (e.g. Google, Facebook, Amazon, etc.), and audit method (e.g. sock puppet, direct scrape, crowdsourcing, etc.). The review shows how previous audit studies have exposed public-facing algorithms exhibiting problematic behavior, such as search algorithms culpable of distortion and advertising algorithms culpable of discrimination. Based on the studies reviewed, it also suggests some behaviors (e.g. discrimination on the basis of intersectional identities), domains (e.g. advertising algorithms), methods (e.g. code auditing), and organizations (e.g. Twitter, TikTok, LinkedIn) that call for future audit attention. The paper concludes by offering the common ingredients of successful audits, and discussing algorithm auditing in the context of broader research working toward algorithmic justice.

Algorithmic timeline curation is now an integral part of Twitter's platform, affecting information exposure for more than 150 million daily active users. Despite its large-scale and high-stakes impact, especially during a public health emergency such as the COVID-19 pandemic, the exact effects of Twitter's curation algorithm generally remain unknown. In this work, we present a sock-puppet audit that aims to characterize the effects of algorithmic curation on source diversity and topic diversity in Twitter timelines. We created eight sock puppet accounts to emulate representative real-world users, selected through a large-scale network analysis. Then, for one month during early 2020, we collected the puppets' timelines twice per day. Broadly, our results show that algorithmic curation increases source diversity in terms of both Twitter accounts and external domains, even though it drastically decreases the number of external links in the timeline. In terms of topic diversity, algorithmic curation had a mixed effect, slightly amplifying a cluster of politically-focused tweets while squelching clusters of tweets focused on COVID-19 fatalities and health information. Finally, we present some evidence that the timeline algorithm may exacerbate partisan differences in exposure to different sources and topics. The paper concludes by discussing broader implications in the context of algorithmic gatekeeping.

The rise of geotargeted online advertising has disrupted the business model of local journalism, but it remains ambiguous whether online advertising platforms can effectively reach local audiences. To address this ambiguity, we present a focused study auditing the positional accuracy of geotargeted display advertisements on Google. We measure the frequency and severity of geotargeting errors by targeting display ads to random ZIP codes across the United States, collecting self-reported location information from users who click on the advertisement. We find evidence that geotargeting errors are common, but minor in terms of advertising goals. While 41% of respondents lived outside the target ZIP code, only 11% lived outside the target county, and only 2% lived outside the target state. We also present details regarding a high volume of suspicious clicks in our data, which made the cost per sample extremely expensive. The paper concludes by discussing implications for advertisers, the business of local journalism, and future research.