Although large language models (LLMs) are increasingly trained using human feedback for safety and alignment with human values, alignment decisions often overlook human social diversity. This study examines how incorporating pluralistic values affects LLM behavior by systematically evaluating demographic variation and design parameters in the alignment pipeline. We collect alignment data from US and German participants (N = 1,095 participants, 27,375 ratings) who rated LLM responses across five dimensions: Toxicity, Emotional Awareness (EA), Sensitivity, Stereotypical Bias, and Helpfulness. We fine-tuned multiple Large Language Models and Large Reasoning Models using preferences from different social groups while varying rating scales, disagreement handling methods, and optimization techniques. The results revealed systematic demographic effects: male participants rated responses 18% less toxic than female participants; conservative and Black participants rated responses 27.9% and 44% higher on EA than liberal and White participants, respectively. Models fine-tuned on group-specific preferences exhibited distinct behaviors. Technical design choices showed strong effects: the preservation of rater disagreement achieved roughly 53% greater toxicity reduction than majority voting, and 5-point scales yielded about 22% more reduction than binary formats; and Direct Preference Optimization (DPO) consistently outperformed Group Relative Policy Optimization (GRPO) in multi-value optimization. These findings represent a preliminary step in answering a critical question: How should alignment balance expert-driven and user-driven signals to ensure both safety and fair representation?

Despite the growing integration of retrieval-enabled AI agents into society, their safety and ethical behavior remain inadequately understood. In particular, the growing integration of LLMs and AI agents with external information sources and real-world environments raises critical questions about how they engage with and are influenced by these external data sources and interactive contexts. This study investigates how expanding retrieval access—from no external sources to Wikipedia-based retrieval and open web search—affects model reliability, bias propagation, and harmful content generation. Through extensive benchmarking of censored and uncensored LLMs and AI Agents, our findings reveal a consistent degradation in refusal rates, bias sensitivity, and harmfulness safeguards as models gain broader access to external sources, culminating in a phenomenon we term safety degradation. Notably, retrieval-enabled agents built on aligned LLMs often behave more unsafely than uncensored models without retrieval. This effect persists even under strong retrieval accuracy and prompt-based mitigation, suggesting that the mere presence of retrieved content reshapes model behavior in structurally unsafe ways. These findings underscore the need for robust mitigation strategies to ensure fairness and reliability in retrieval-enabled and increasingly autonomous AI systems.

Yusuke Hirota, Jerone Andrews, Dora Zhao, Orestis Papakyriakopoulos, Apostolos Modas, Yuta Nakashima, Alice Xiang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Machine learning (ML) datasets, often perceived as neutral, inherently encapsulate abstract and disputed social constructs. Dataset curators frequently employ value-laden terms such as diversity, bias, and quality to characterize datasets. Despite their prevalence, these terms lack clear definitions and validation. Our research explores the implications of this issue by analyzing "diversity" across 135 image and text datasets. Drawing from social sciences, we apply principles from measurement theory to identify considerations and offer recommendations for conceptualizing, operationalizing, and evaluating diversity in datasets. Our findings have broader implications for ML research, advocating for a more nuanced and precise approach to handling value-laden properties in dataset construction.

Beyond the well-known giants like Uber Eats and DoorDash, there are hundreds of independent food delivery platforms in the United States. However, little is known about the sociotechnical landscape of these "indie'' platforms. In this paper, we analyzed these platforms to understand why they were created, how they operate, and what technologies they use. We collected data on 495 indie platforms and detailed survey responses from 29 platforms. We found that personalized, timely service is a central value of indie platforms, as is a sense of responsibility to the local community they serve. Indie platforms are motivated to provide fair rates for restaurants and couriers. These alternative business practices differentiate them from mainstream platforms. Though indie platforms have plans to expand, a lack of customizability in off-the-shelf software prevents independent platforms from personalizing services for their local communities. We show that these platforms are a widespread and longstanding fixture of the food delivery market. We illustrate the diversity of motivations and values to explain why a one-size-fits-all support is insufficient, and we discuss the siloing of technology that inhibits platforms' growth. Through these insights, we aim to promote future HCI research into the potential development of public-interest technologies for local food delivery.

A significant share of political discourse occurs online on social media platforms. Policymakers and researchers try to understand the role of social media design in shaping the quality of political discourse around the globe. In the past decades, scholarship on political discourse theory has produced distinct characteristics of different types of prominent political rhetoric such as deliberative, civic, or demagogic discourse. This study investigates the relationship between social media reaction mechanisms (i.e., upvotes, downvotes) and political rhetoric in user discussions by engaging in an in-depth conceptual analysis of political discourse theory. First, we analyze 155 million user comments in 55 political subforums on Reddit between 2010 and 2018 to explore whether users’ style of political discussion aligns with the essential components of deliberative, civic, and demagogic discourse. Second, we perform a quantitative study that combines confirmatory factor analysis with difference in differences models to explore whether different reaction mechanism schemes (e.g., upvotes only, upvotes and downvotes, no reaction mechanisms) correspond with political user discussion that is more or less characteristic of deliberative, civic, or demagogic discourse. We produce three main takeaways. First, despite being “ideal constructs of political rhetoric,” we find that political discourse theories describe political discussions on Reddit to a large extent. Second, we find that discussions in subforums with only upvotes, or both up- and downvotes are associated with user discourse that is more deliberate and civic. Third, and perhaps most strikingly, social media discussions are most demagogic in subreddits with no reaction mechanisms at all. These findings offer valuable contributions for ongoing policy discussions on the relationship between social media interface design and respectful political discussion among users.1

We investigate how representations of Syrian refugees (2011-2021) differ across US partisan news outlets. We analyze 47,388 articles from the online US media about Syrian refugees to detail differences in reporting between left- and right-leaning media. We use various NLP techniques to understand these differences. Our polarization and question answering results indicated that left-leaning media tended to represent refugees as child victims, welcome in the US, and right-leaning media cast refugees as Islamic terrorists. We noted similar results with our sentiment and offensive speech scores over time, which detail possibly unfavorable representations of refugees in right-leaning media. A strength of our work is how the different techniques we have applied validate each other. Based on our results, we provide several recommendations. Stakeholders may utilize our findings to intervene around refugee representations, and design communications campaigns that improve the way society sees refugees and possibly aid refugee outcomes.

Human-centric computer vision (HCCV) data curation practices often neglect privacy and bias concerns, leading to dataset retractions and unfair models. HCCV datasets constructed through nonconsensual web scraping lack crucial metadata for comprehensive fairness and robustness evaluations. Current remedies are post hoc, lack persuasive justification for adoption, or fail to provide proper contextualization for appropriate application. Our research focuses on proactive, domain-specific recommendations, covering purpose, privacy and consent, and diversity, for curating HCCV evaluation datasets, addressing privacy and bias concerns. We adopt an ante hoc reflective perspective, drawing from current practices, guidelines, dataset withdrawals, and audits, to inform our considerations and recommendations.

Speech AI Technologies are largely trained on publicly available datasets or by the massive web-crawling of speech. In both cases, data acquisition focuses on minimizing collection effort, without necessarily taking the data subjects' protection or user needs into consideration. This results to models that are not robust when used on users who deviate from the dominant demographics in the training set, discriminating individuals having different dialects, accents, speaking styles, and disfluencies. In this talk, we use automatic speech recognition as a case study and examine the properties that ethical speech datasets should possess towards responsible AI applications. We showcase diversity issues, inclusion practices, and necessary considerations that can improve trained models, while facilitating model explainability and protecting users and data subjects. We argue for the legal & privacy protection of data subjects, targeted data sampling corresponding to user demographics & needs, appropriate meta data that ensure explainability & accountability in cases of model failure, and the sociotechnical & situated model design. We hope this talk can inspire researchers & practitioners to design and use more human-centric datasets in speech technologies and other domains, in ways that empower and respect users, while improving machine learning models' robustness and utility.

Online platforms play an increasingly important role in shaping democracy by influencing the distribution of political information to the electorate. In recent years, political campaigns have spent heavily on the platforms' algorithmic tools to target voters with online advertising. While the public interest in understanding how platforms perform the task of shaping the political discourse has never been higher, the efforts of the major platforms to make the necessary disclosures to understand their practices falls woefully short. In this study, we collect and analyze a dataset containing over 800,000 ads and 2.5 million videos about the 2020 U.S. presidential election from Facebook, Google, and TikTok. We conduct the first large scale data analysis of public data to critically evaluate how these platforms amplified or moderated the distribution of political advertisements. We conclude with recommendations for how to improve the disclosures so that the public can hold the platforms and political advertisers accountable.

Social media platforms are performing “soft moderation” by attaching warning labels to misinformation to reduce dissemination of, and engagement with, such content. This study investigates the warning labels that Twitter placed on Donald Trump’s false tweets about the 2020 US Presidential election. It specifically studies their relation to misinformation spread, and the magnitude and nature of user engagement. We categorize the warning labels by type –“veracity labels” calling out falsity and “contextual labels” providing more information. In addition, we categorize labels by their rebuttal strength and textual overlap (linguistic, topical) with the underlying tweet. We look at user interactions (liking, retweeting, quote tweeting, and replying), the content of user replies, and the type of user involved (partisanship and Twitter activity level) according to various standard metrics. Using appropriate statistical tools, we find that, overall, label placement did not change the propensity of users to share and engage with labeled content, but the falsity of content did. However, we show that the presence of textual overlap in labels did reduce user interactions, while stronger rebuttals reduced the toxicity in comments. We also find that users were more likely to discuss their positions on the underlying tweets in replies when the labels contained rebuttals. When false content was labeled, results show that liberals engaged more than conservatives. Labels also increased the engagement of more passive Twitter users. This case study has direct implications for the design of effective soft moderation and related policies.