Vinodkumar Prabhakaran

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin et al. · deepmind, stanford

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model PaLM. We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

Shaily Bhatt, Sunipa Dev, Partha Talukdar et al. · cmu

Recent research has revealed undesirable biases in NLP data and models. However, these efforts focus on social disparities in West, and are not directly portable to other geo-cultural contexts. In this paper, we focus on NLP fair-ness in the context of India. We start with a brief account of the prominent axes of social disparities in India. We build resources for fairness evaluation in the Indian context and use them to demonstrate prediction biases along some of the axes. We then delve deeper into social stereotypes for Region andReligion, demonstrating its prevalence in corpora and models. Finally, we outline a holistic research agenda to re-contextualize NLP fairness research for the Indian context, ac-counting for Indian societal context, bridging technological gaps in NLP capabilities and re-sources, and adapting to Indian cultural values. While we focus on India, this framework can be generalized to other geo-cultural contexts.

Shaily Bhatt, Sunipa Dev, Partha Talukdar et al. · cmu

Recent research has revealed undesirable biases in NLP data and models. However, these efforts largely focus on social disparities in the West, and are not directly portable to other geo-cultural contexts. In this position paper, we outline a holistic research agenda to re-contextualize NLP fairness research for the Indian context, accounting for Indian societal context, bridging technological gaps in capability and resources, and adapting to Indian cultural values. We also summarize findings from an empirical study on various social biases along different axes of disparities relevant to India, demonstrating their prevalence in corpora and models.

Ben Hutchinson, Jason Baldridge, Vinodkumar Prabhakaran

Questions regarding implicitness, ambiguity and underspecification are crucial for understanding the task validity and ethical concerns of multimodal image+text systems, yet have received little attention to date. This position paper maps out a conceptual framework to address this gap, focusing on systems which generate images depicting scenes from scene descriptions. In doing so, we account for how texts and images convey meaning differently. We outline a set of core challenges concerning textual and visual ambiguity, as well as risks that may be amplified by ambiguous and underspecified elements. We propose and discuss strategies for addressing these challenges, including generating visually ambiguous images, and generating a set of diverse images.

Mark Diaz, Ian D. Kivlichan, Rachel Rosen et al.

Human annotated data plays a crucial role in machine learning (ML) research and development. However, the ethical considerations around the processes and decisions that go into dataset annotation have not received nearly enough attention. In this paper, we survey an array of literature that provides insights into ethical considerations around crowdsourced dataset annotation. We synthesize these insights, and lay out the challenges in this space along two layers: (1) who the annotator is, and how the annotators' lived experiences can impact their annotations, and (2) the relationship between the annotators and the crowdsourcing platforms, and what that relationship affords them. Finally, we introduce a novel framework, CrowdWorkSheets, for dataset developers to facilitate transparent documentation of key decisions points at various stages of the data annotation pipeline: task formulation, selection of annotators, platform and infrastructure choices, dataset analysis and evaluation, and dataset release and maintenance.

Ben Hutchinson, Negar Rostamzadeh, Christina Greer et al.

Forming a reliable judgement of a machine learning (ML) model's appropriateness for an application ecosystem is critical for its responsible use, and requires considering a broad range of factors including harms, benefits, and responsibilities. In practice, however, evaluations of ML models frequently focus on only a narrow range of decontextualized predictive behaviours. We examine the evaluation gaps between the idealized breadth of evaluation concerns and the observed narrow focus of actual evaluations. Through an empirical study of papers from recent high-profile conferences in the Computer Vision and Natural Language Processing communities, we demonstrate a general focus on a handful of evaluation methods. By considering the metrics and test data distributions used in these methods, we draw attention to which properties of models are centered in the field, revealing the properties that are frequently neglected or sidelined during evaluation. By studying these properties, we demonstrate the machine learning discipline's implicit assumption of a range of commitments which have normative impacts; these include commitments to consequentialism, abstractability from context, the quantifiability of impacts, the limited role of model inputs in evaluation, and the equivalence of different failure modes. Shedding light on these assumptions enables us to question their appropriateness for ML system contexts, pointing the way towards more contextualized evaluation methodologies for robustly examining the trustworthiness of ML models

Vinodkumar Prabhakaran, Margaret Mitchell, Timnit Gebru et al.

Research on fairness, accountability, transparency and ethics of AI-based interventions in society has gained much-needed momentum in recent years. However it lacks an explicit alignment with a set of normative values and principles that guide this research and interventions. Rather, an implicit consensus is often assumed to hold for the values we impart into our models - something that is at odds with the pluralistic world we live in. In this paper, we put forth the doctrine of universal human rights as a set of globally salient and cross-culturally recognized set of values that can serve as a grounding framework for explicit value alignment in responsible AI - and discuss its efficacy as a framework for civil society partnership and participation. We argue that a human rights framework orients the research in this space away from the machines and the risks of their biases, and towards humans and the risks to their rights, essentially helping to center the conversation around who is harmed, what harms they face, and how those harms may be mitigated.

Vinodkumar Prabhakaran, Rida Qadri, Ben Hutchinson

Artificial intelligence (AI) systems attempt to imitate human behavior. How well they do this imitation is often used to assess their utility and to attribute human-like (or artificial) intelligence to them. However, most work on AI refers to and relies on human intelligence without accounting for the fact that human behavior is inherently shaped by the cultural contexts they are embedded in, the values and beliefs they hold, and the social practices they follow. Additionally, since AI technologies are mostly conceived and developed in just a handful of countries, they embed the cultural values and practices of these countries. Similarly, the data that is used to train the models also fails to equitably represent global cultural diversity. Problems therefore arise when these technologies interact with globally diverse societies and cultures, with different values and interpretive practices. In this position paper, we describe a set of cultural dependencies and incongruencies in the context of AI-based language and vision technologies, and reflect on the possibilities of and potential strategies towards addressing these incongruencies.

Sunipa Dev, Jaya Goyal, Dinesh Tewari et al.

With rapid development and deployment of generative language models in global settings, there is an urgent need to also scale our measurements of harm, not just in the number and types of harms covered, but also how well they account for local cultural contexts, including marginalized identities and the social biases experienced by them. Current evaluation paradigms are limited in their abilities to address this, as they are not representative of diverse, locally situated but global, socio-cultural perspectives. It is imperative that our evaluation resources are enhanced and calibrated by including people and experiences from different cultures and societies worldwide, in order to prevent gross underestimations or skews in measurements of harm. In this work, we demonstrate a socio-culturally aware expansion of evaluation resources in the Indian societal context, specifically for the harm of stereotyping. We devise a community engaged effort to build a resource which contains stereotypes for axes of disparity that are uniquely present in India. The resultant resource increases the number of stereotypes known for and in the Indian context by over 1000 stereotypes across many unique identities. We also demonstrate the utility and effectiveness of such expanded resources for evaluations of language models. CONTENT WARNING: This paper contains examples of stereotypes that may be offensive.

Nithish Kannen, Arif Ahmad, Marco Andreetto et al.

Text-to-Image (T2I) models are being increasingly adopted in diverse global communities where they create visual representations of their unique cultures. Current T2I benchmarks primarily focus on faithfulness, aesthetics, and realism of generated images, overlooking the critical dimension of cultural competence. In this work, we introduce a framework to evaluate cultural competence of T2I models along two crucial dimensions: cultural awareness and cultural diversity, and present a scalable approach using a combination of structured knowledge bases and large language models to build a large dataset of cultural artifacts to enable this evaluation. In particular, we apply this approach to build CUBE (CUltural BEnchmark for Text-to-Image models), a first-of-its-kind benchmark to evaluate cultural competence of T2I models. CUBE covers cultural artifacts associated with 8 countries across different geo-cultural regions and along 3 concepts: cuisine, landmarks, and art. CUBE consists of 1) CUBE-1K, a set of high-quality prompts that enable the evaluation of cultural awareness, and 2) CUBE-CSpace, a larger dataset of cultural artifacts that serves as grounding to evaluate cultural diversity. We also introduce cultural diversity as a novel T2I evaluation component, leveraging quality-weighted Vendi score. Our evaluations reveal significant gaps in the cultural awareness of existing models across countries and provide valuable insights into the cultural diversity of T2I outputs for under-specified prompts. Our methodology is extendable to other cultural regions and concepts, and can facilitate the development of T2I models that better cater to the global population.

Jiho Jin, Junho Myung, Juhyun Oh et al.

As large language models (LLMs) are increasingly deployed to users around the world, they are integrated into everyday tasks across diverse cultural contexts, from drafting personal communications to brainstorming creative ideas. These tasks are inherently cultural: they require contextual appropriateness, symbolic resonance, and tacit cultural expectations that native speakers draw on instinctively, meaning that a response can be factually plausible yet unmistakably wrong to a local reader. Existing cultural benchmarks have treated culture as a flat set of facts via fact verification or norm entailment methods, and have adopted LLM-as-a-Judge without examining whether they can capture such thick cultural errors. To address this gap, we present JuICE (Benchmark for LLM-Judge in Identifying Cultural Errors), a multilingual dataset of 7,470 span-level annotations of cultural and linguistic errors in long-form LLM responses. It covers 1,050 query-response pairs from four countries (the United States, South Korea, Indonesia, and Bangladesh), in both English and their countries' main languages. Using JuICE, we find that even the strongest LLM-judge achieves only an F1 of 0.52 in the erroneous span detection task. Furthermore, LLM-judges consistently miss thick cultural errors that local residents readily identify. Our findings suggest that robust cultural evaluation must move beyond surface-level detection toward frameworks that account for the depth and situatedness of cultural meaning.

Vinodkumar Prabhakaran, Christopher Homan, Lora Aroyo et al.

Human annotation plays a core role in machine learning -- annotations for supervised models, safety guardrails for generative models, and human feedback for reinforcement learning, to cite a few avenues. However, the fact that many of these human annotations are inherently subjective is often overlooked. Recent work has demonstrated that ignoring rater subjectivity (typically resulting in rater disagreement) is problematic within specific tasks and for specific subgroups. Generalizable methods to harness rater disagreement and thus understand the socio-cultural leanings of subjective tasks remain elusive. In this paper, we propose GRASP, a comprehensive disagreement analysis framework to measure group association in perspectives among different rater sub-groups, and demonstrate its utility in assessing the extent of systematic disagreements in two datasets: (1) safety annotations of human-chatbot conversations, and (2) offensiveness annotations of social media posts, both annotated by diverse rater pools across different socio-demographic axes. Our framework (based on disagreement metrics) reveals specific rater groups that have significantly different perspectives than others on certain tasks, and helps identify demographic axes that are crucial to consider in specific task contexts.

Mark Díaz, Sunipa Dev, Emily Reif et al.

The unstructured nature of data used in foundation model development is a challenge to systematic analyses for making data use and documentation decisions. From a Responsible AI perspective, these decisions often rely upon understanding how people are represented in data. We propose a framework designed to guide analysis of human representation in unstructured data and identify downstream risks. We apply the framework in two toy examples using the Common Crawl web text corpus (C4) and LAION-400M. We also propose a set of hypothetical action steps in service of dataset use, development, and documentation.

Sunipa Dev, Vinodkumar Prabhakaran, Rutledge Chin Feman et al.

As generative AI technologies are increasingly being launched across the globe, assessing their competence to operate in different cultural contexts is exigently becoming a priority. While recent years have seen numerous and much-needed efforts on cultural benchmarking, these efforts have largely focused on specific aspects of culture and evaluation. While these efforts contribute to our understanding of cultural competence, a unified and systematic evaluation approach is needed for us as a field to comprehensively assess diverse cultural dimensions at scale. Drawing on measurement theory, we present a principled framework to aggregate multifaceted indicators of cultural capabilities into a unified assessment of cultural intelligence. We start by developing a working definition of culture that includes identifying core domains of culture. We then introduce a broad-purpose, systematic, and extensible framework for assessing cultural intelligence of AI systems. Drawing on theoretical framing from psychometric measurement validity theory, we decouple the background concept (i.e., cultural intelligence) from its operationalization via measurement. We conceptualize cultural intelligence as a suite of core capabilities spanning diverse domains, which we then operationalize through a set of indicators designed for reliable measurement. Finally, we identify the considerations, challenges, and research pathways to meaningfully measure these indicators, specifically focusing on data collection, probing strategies, and evaluation metrics.

Aishwarya Verma, Laud Ammah, Olivia Nercy Ndlovu Lucas et al.

Stereotype repositories are critical to assess generative AI model safety, but currently lack adequate global coverage. It is imperative to prioritize targeted expansion, strategically addressing existing deficits, over merely increasing data volume. This work introduces a multilingual stereotype resource covering four sub-Saharan African countries that are severely underrepresented in NLP resources: Ghana, Kenya, Nigeria, and South Africa. By utilizing socioculturally-situated, community-engaged methods, including telephonic surveys moderated in native languages, we establish a reproducible methodology that is sensitive to the region's complex linguistic diversity and traditional orality. By deliberately balancing the sample across diverse ethnic and demographic backgrounds, we ensure broad coverage, resulting in a dataset of 3,534 stereotypes in English and 3,206 stereotypes across 15 native languages.

Wenbo Zhang, Hangzhi Guo, Ian D Kivlichan et al.

Toxicity is an increasingly common and severe issue in online spaces. Consequently, a rich line of machine learning research over the past decade has focused on computationally detecting and mitigating online toxicity. These efforts crucially rely on human-annotated datasets that identify toxic content of various kinds in social media texts. However, such annotations historically yield low inter-rater agreement, which was often dealt with by taking the majority vote or other such approaches to arrive at a single ground truth label. Recent research has pointed out the importance of accounting for the subjective nature of this task when building and utilizing these datasets, and this has triggered work on analyzing and better understanding rater disagreements, and how they could be effectively incorporated into the machine learning developmental pipeline. While these efforts are filling an important gap, there is a lack of a broader framework about the root causes of rater disagreement, and therefore, we situate this work within that broader landscape. In this survey paper, we analyze a broad set of literature on the reasons behind rater disagreements focusing on online toxicity, and propose a detailed taxonomy for the same. Further, we summarize and discuss the potential solutions targeting each reason for disagreement. We also discuss several open issues, which could promote the future development of online toxicity research.

Robin Schimmelpfennig, Mark Díaz, Vinodkumar Prabhakaran et al.

Over a billion users globally interact with AI systems engineered to mimic human traits. This development raises concerns that anthropomorphism, the attribution of human characteristics to AI, may foster over-reliance and misplaced trust. Yet, causal effects of humanlike AI design on users remain untested in ecologically valid, cross-cultural settings, leaving policy discussions to rely on theoretical assumptions derived largely from Western populations. Here we conducted two experiments (N=3,500) across ten countries representing a wide cultural spectrum, involving real-time, open-ended interactions with a state-of-the-art chatbot. We found users evaluate human-likeness based on pragmatic interactional cues (conversation flow, response speed, perspective-taking) rather than abstract theory-driven attributes emphasized in academic discourse (e.g., sentience, consciousness). Furthermore, while experimentally increasing chatbot's human-likeness reliably increased anthropomorphism across all sampled countries, it did not universally increase trust or engagement. Instead, effects were culturally contingent; design choices fostering engagement or trust in one country may reduce them in another. These findings challenge prevailing assumptions that humanlike AI poses uniform psychological risks and necessarily increases trust. Risk is not inherent to humanlike design but emerges from its interplay with cultural context. Consequently, governance frameworks must move beyond universalist approaches to account for this global heterogeneity.

Aida Davani, Mark Díaz, Dylan Baker et al.

Perception of offensiveness is inherently subjective, shaped by the lived experiences and socio-cultural values of the perceivers. Recent years have seen substantial efforts to build AI-based tools that can detect offensive language at scale, as a means to moderate social media platforms, and to ensure safety of conversational AI technologies such as ChatGPT and Bard. However, existing approaches treat this task as a technical endeavor, built on top of data annotated for offensiveness by a global crowd workforce without any attention to the crowd workers' provenance or the values their perceptions reflect. We argue that cultural and psychological factors play a vital role in the cognitive processing of offensiveness, which is critical to consider in this context. We re-frame the task of determining offensiveness as essentially a matter of moral judgment -- deciding the boundaries of ethically wrong vs. right language within an implied set of socio-cultural norms. Through a large-scale cross-cultural study based on 4309 participants from 21 countries across 8 cultural regions, we demonstrate substantial cross-cultural differences in perceptions of offensiveness. More importantly, we find that individual moral values play a crucial role in shaping these variations: moral concerns about Care and Purity are significant mediating factors driving cross-cultural differences. These insights are of crucial importance as we build AI models for the pluralistic world, where the values they espouse should aim to respect and account for moral values in diverse geo-cultural contexts.

Mukul Bhutani, Kevin Robinson, Vinodkumar Prabhakaran et al.

While generative multilingual models are rapidly being deployed, their safety and fairness evaluations are largely limited to resources collected in English. This is especially problematic for evaluations targeting inherently socio-cultural phenomena such as stereotyping, where it is important to build multi-lingual resources that reflect the stereotypes prevalent in respective language communities. However, gathering these resources, at scale, in varied languages and regions pose a significant challenge as it requires broad socio-cultural knowledge and can also be prohibitively expensive. To overcome this critical gap, we employ a recently introduced approach that couples LLM generations for scale with culturally situated validations for reliability, and build SeeGULL Multilingual, a global-scale multilingual dataset of social stereotypes, containing over 25K stereotypes, spanning 20 languages, with human annotations across 23 regions, and demonstrate its utility in identifying gaps in model evaluations. Content warning: Stereotypes shared in this paper can be offensive.

Akshita Jha, Vinodkumar Prabhakaran, Remi Denton et al.

Recent studies have shown that Text-to-Image (T2I) model generations can reflect social stereotypes present in the real world. However, existing approaches for evaluating stereotypes have a noticeable lack of coverage of global identity groups and their associated stereotypes. To address this gap, we introduce the ViSAGe (Visual Stereotypes Around the Globe) dataset to enable the evaluation of known nationality-based stereotypes in T2I models, across 135 nationalities. We enrich an existing textual stereotype resource by distinguishing between stereotypical associations that are more likely to have visual depictions, such as `sombrero', from those that are less visually concrete, such as 'attractive'. We demonstrate ViSAGe's utility through a multi-faceted evaluation of T2I generations. First, we show that stereotypical attributes in ViSAGe are thrice as likely to be present in generated images of corresponding identities as compared to other attributes, and that the offensiveness of these depictions is especially higher for identities from Africa, South America, and South East Asia. Second, we assess the stereotypical pull of visual depictions of identity groups, which reveals how the 'default' representations of all identity groups in ViSAGe have a pull towards stereotypical depictions, and that this pull is even more prominent for identity groups from the Global South. CONTENT WARNING: Some examples contain offensive stereotypes.

Aida Mostafazadeh Davani, Mark Díaz, Dylan Baker et al.

While human annotations play a crucial role in language technologies, annotator subjectivity has long been overlooked in data collection. Recent studies that have critically examined this issue are often situated in the Western context, and solely document differences across age, gender, or racial groups. As a result, NLP research on subjectivity have overlooked the fact that individuals within demographic groups may hold diverse values, which can influence their perceptions beyond their group norms. To effectively incorporate these considerations into NLP pipelines, we need datasets with extensive parallel annotations from various social and cultural groups. In this paper we introduce the \dataset dataset: a large-scale cross-cultural dataset of parallel annotations for offensive language in over 4.5K sentences annotated by a pool of over 4k annotators, balanced across gender and age, from across 21 countries, representing eight geo-cultural regions. The dataset contains annotators' moral values captured along six moral foundations: care, equality, proportionality, authority, loyalty, and purity. Our analyses reveal substantial regional variations in annotators' perceptions that are shaped by individual moral values, offering crucial insights for building pluralistic, culturally sensitive NLP models.

Charvi Rastogi, Tian Huey Teh, Pushkar Mishra et al.

AI systems crucially rely on human ratings, but these ratings are often aggregated, obscuring the inherent diversity of perspectives in real-world phenomenon. This is particularly concerning when evaluating the safety of generative AI, where perceptions and associated harms can vary significantly across socio-cultural contexts. While recent research has studied the impact of demographic differences on annotating text, there is limited understanding of how these subjective variations affect multimodal safety in generative AI. To address this, we conduct a large-scale study employing highly-parallel safety ratings of about 1000 text-to-image (T2I) generations from a demographically diverse rater pool of 630 raters balanced across 30 intersectional groups across age, gender, and ethnicity. Our study shows that (1) there are significant differences across demographic groups (including intersectional groups) on how severe they assess the harm to be, and that these differences vary across different types of safety violations, (2) the diverse rater pool captures annotation patterns that are substantially different from expert raters trained on specific set of safety policies, and (3) the differences we observe in T2I safety are distinct from previously documented group level differences in text-based safety tasks. To further understand these varying perspectives, we conduct a qualitative analysis of the open-ended explanations provided by raters. This analysis reveals core differences into the reasons why different groups perceive harms in T2I generations. Our findings underscore the critical need for incorporating diverse perspectives into safety evaluation of generative AI ensuring these systems are truly inclusive and reflect the values of all users.

Rida Qadri, Aida M. Davani, Kevin Robinson et al.

Large language models are increasingly being integrated into applications that shape the production and discovery of societal knowledge such as search, online education, and travel planning. As a result, language models will shape how people learn about, perceive and interact with global cultures making it important to consider whose knowledge systems and perspectives are represented in models. Recognizing this importance, increasingly work in Machine Learning and NLP has focused on evaluating gaps in global cultural representational distribution within outputs. However, more work is needed on developing benchmarks for cross-cultural impacts of language models that stem from a nuanced sociologically-aware conceptualization of cultural impact or harm. We join this line of work arguing for the need of metricizable evaluations of language technologies that interrogate and account for historical power inequities and differential impacts of representation on global cultures, particularly for cultures already under-represented in the digital corpora. We look at two concepts of erasure: omission: where cultures are not represented at all and simplification i.e. when cultural complexity is erased by presenting one-dimensional views of a rich culture. The former focuses on whether something is represented, and the latter on how it is represented. We focus our analysis on two task contexts with the potential to influence global cultural production. First, we probe representations that a language model produces about different places around the world when asked to describe these contexts. Second, we analyze the cultures represented in the travel recommendations produced by a set of language model applications. Our study shows ways in which the NLP community and application developers can begin to operationalize complex socio-cultural considerations into standard evaluations and benchmarks.

Piyawat Lertvittayakumjorn, David Kinney, Vinodkumar Prabhakaran et al.

Generative large language models (LLMs) have demonstrated gaps in diverse cultural awareness across the globe. We investigate the effect of retrieval augmented generation and search-grounding techniques on LLMs' ability to display familiarity with various national cultures. Specifically, we compare the performance of standard LLMs, LLMs augmented with retrievals from a bespoke knowledge base (i.e., KB grounding), and LLMs augmented with retrievals from a web search (i.e., search grounding) on multiple cultural awareness benchmarks. We find that search grounding significantly improves the LLM performance on multiple-choice benchmarks that test propositional knowledge (e.g., cultural norms, artifacts, and institutions), while KB grounding's effectiveness is limited by inadequate knowledge base coverage and a suboptimal retriever. However, search grounding also increases the risk of stereotypical judgments by language models and fails to improve evaluators' judgments of cultural familiarity in a human evaluation with adequate statistical power. These results highlight the distinction between propositional cultural knowledge and open-ended cultural fluency when it comes to evaluating LLMs' cultural awareness.

Charvi Rastogi, Tian Huey Teh, Pushkar Mishra et al.

Current text-to-image (T2I) models often fail to account for diverse human experiences, leading to misaligned systems. We advocate for pluralistic alignment, where an AI understands and is steerable towards diverse, and often conflicting, human values. Our work provides three core contributions to achieve this in T2I models. First, we introduce a novel dataset for Diverse Intersectional Visual Evaluation (DIVE) -- the first multimodal dataset for pluralistic alignment. It enable deep alignment to diverse safety perspectives through a large pool of demographically intersectional human raters who provided extensive feedback across 1000 prompts, with high replication, capturing nuanced safety perceptions. Second, we empirically confirm demographics as a crucial proxy for diverse viewpoints in this domain, revealing significant, context-dependent differences in harm perception that diverge from conventional evaluations. Finally, we discuss implications for building aligned T2I models, including efficient data collection strategies, LLM judgment capabilities, and model steerability towards diverse perspectives. This research offers foundational tools for more equitable and aligned T2I systems. Content Warning: The paper includes sensitive content that may be harmful.

Aida Mostafazadeh Davani, Sagar Gubbi, Sunipa Dev et al.

Generative language models are transforming our digital ecosystem, but they often inherit societal biases, for instance stereotypes associating certain attributes with specific identity groups. While whether and how these biases are mitigated may depend on the specific use cases, being able to effectively detect instances of stereotype perpetuation is a crucial first step. Current methods to assess presence of stereotypes in generated language rely on simple template or co-occurrence based measures, without accounting for the variety of sentential contexts they manifest in. We argue that understanding the sentential context is crucial for detecting instances of generalization. We distinguish two types of generalizations: (1) language that merely mentions the presence of a generalization ("people think the French are very rude"), and (2) language that reinforces such a generalization ("as French they must be rude"), from non-generalizing context ("My French friends think I am rude"). For meaningful stereotype evaluations, we need to reliably distinguish such instances of generalizations. We introduce the new task of detecting generalization in language, and build GeniL, a multilingual dataset of over 50K sentences from 9 languages (English, Arabic, Bengali, Spanish, French, Hindi, Indonesian, Malay, and Portuguese) annotated for instances of generalizations. We demonstrate that the likelihood of a co-occurrence being an instance of generalization is usually low, and varies across different languages, identity groups, and attributes. We build classifiers to detect generalization in language with an overall PR-AUC of 58.7, with varying degrees of performance across languages. Our research provides data and tools to enable a nuanced understanding of stereotype perpetuation, a crucial step towards more inclusive and responsible language technologies.

Myra Cheng, Vinodkumar Prabhakaran, Alice Oh et al.

Generative AI models ought to be useful and safe across cross-cultural contexts. One critical step toward this goal is understanding how AI models adhere to sociocultural norms. While this challenge has gained attention in NLP, existing work lacks both nuance and coverage in understanding and evaluating models' norm adherence. We address these gaps by introducing a taxonomy of norms that clarifies their contexts (e.g., distinguishing between human-human norms that models should recognize and human-AI interactional norms that apply to the human-AI interaction itself), specifications (e.g., relevant domains), and mechanisms (e.g., modes of enforcement). We demonstrate how our taxonomy can be operationalized to automatically evaluate models' norm adherence in naturalistic, open-ended settings. Our exploratory analyses suggest that state-of-the-art models frequently violate norms, though violation rates vary by model, interactional context, and country. We further show that violation rates also vary by prompt intent and situational framing. Our taxonomy and demonstrative evaluation pipeline enable nuanced, context-sensitive evaluation of cultural norm adherence in realistic settings.

Renee Shelby, Fernando Diaz, Vinodkumar Prabhakaran

The growing ubiquity of conversational AI highlights the need for frameworks that capture not only users' instrumental goals but also the situated, adaptive, and social practices through which they achieve them. Existing taxonomies of conversational behavior either overgeneralize, remain domain-specific, or reduce interactions to narrow dialogue functions. To address this gap, we introduce the Taxonomy of User Needs and Actions (TUNA), an empirically grounded framework developed through iterative qualitative analysis of 1193 human-AI conversations, supplemented by theoretical review and validation across diverse contexts. TUNA organizes user actions into a three-level hierarchy encompassing behaviors associated with information seeking, synthesis, procedural guidance, content creation, social interaction, and meta-conversation. By centering user agency and appropriation practices, TUNA enables multi-scale evaluation, supports policy harmonization across products, and provides a backbone for layering domain-specific taxonomies. This work contributes a systematic vocabulary for describing AI use, advancing both scholarly understanding and practical design of safer, more responsive, and more accountable conversational systems.

Ding Wang, Mark Díaz, Charvi Rastogi et al.

Understanding what constitutes safety in AI-generated content is complex. While developers often rely on predefined taxonomies, real-world safety judgments also involve personal, social, and cultural perceptions of harm. This paper examines how annotators evaluate the safety of AI-generated images, focusing on the qualitative reasoning behind their judgments. Analyzing 5,372 open-ended comments, we find that annotators consistently invoke moral, emotional, and contextual reasoning that extends beyond structured safety categories. Many reflect on potential harm to others more than to themselves, grounding their judgments in lived experience, collective risk, and sociocultural awareness. Beyond individual perceptions, we also find that the structure of the task itself -- including annotation guidelines -- shapes how annotators interpret and express harm. Guidelines influence not only which images are flagged, but also the moral judgment behind the justifications. Annotators frequently cite factors such as image quality, visual distortion, and mismatches between prompt and output as contributing to perceived harm dimensions, which are often overlooked in standard evaluation frameworks. Our findings reveal that existing safety pipelines miss critical forms of reasoning that annotators bring to the task. We argue for evaluation designs that scaffold moral reflection, differentiate types of harm, and make space for subjective, context-sensitive interpretations of AI-generated content.

Jennifer Chien, A. Stevie Bergman, Kevin R. McKee et al.

Algorithmic fairness has emerged as a critical concern in artificial intelligence (AI) research. However, the development of fair AI systems is not an objective process. Fairness is an inherently subjective concept, shaped by the values, experiences, and identities of those involved in research and development. To better understand the norms and values embedded in current fairness research, we conduct a meta-analysis of algorithmic fairness papers from two leading conferences on AI fairness and ethics, AIES and FAccT, covering a final sample of 139 papers over the period from 2018 to 2022. Our investigation reveals two concerning trends: first, a US-centric perspective dominates throughout fairness research; and second, fairness studies exhibit a widespread reliance on binary codifications of human identity (e.g., "Black/White", "male/female"). These findings highlight how current research often overlooks the complexities of identity and lived experiences, ultimately failing to represent diverse global contexts when defining algorithmic bias and fairness. We discuss the limitations of these research design choices and offer recommendations for fostering more inclusive and representative approaches to fairness in AI systems, urging a paradigm shift that embraces nuanced, global understandings of human identity and values.

Akshita Jha, Aida Davani, Chandan K. Reddy et al.

Stereotype benchmark datasets are crucial to detect and mitigate social stereotypes about groups of people in NLP models. However, existing datasets are limited in size and coverage, and are largely restricted to stereotypes prevalent in the Western society. This is especially problematic as language technologies gain hold across the globe. To address this gap, we present SeeGULL, a broad-coverage stereotype dataset, built by utilizing generative capabilities of large language models such as PaLM, and GPT-3, and leveraging a globally diverse rater pool to validate the prevalence of those stereotypes in society. SeeGULL is in English, and contains stereotypes about identity groups spanning 178 countries across 8 different geo-political regions across 6 continents, as well as state-level identities within the US and India. We also include fine-grained offensiveness scores for different stereotypes and demonstrate their global disparities. Furthermore, we include comparative annotations about the same groups by annotators living in the region vs. those that are based in North America, and demonstrate that within-region stereotypes about groups differ from those prevalent in North America. CONTENT WARNING: This paper contains stereotype examples that may be offensive.

Jacob Eisenstein, Vinodkumar Prabhakaran, Clara Rivera et al.

We introduce a new dataset of conversational speech representing English from India, Nigeria, and the United States. The Multi-Dialect Dataset of Dialogues (MD3) strikes a new balance between open-ended conversational speech and task-oriented dialogue by prompting participants to perform a series of short information-sharing tasks. This facilitates quantitative cross-dialectal comparison, while avoiding the imposition of a restrictive task structure that might inhibit the expression of dialect features. Preliminary analysis of the dataset reveals significant differences in syntax and in the use of discourse markers. The dataset, which will be made publicly available with the publication of this paper, includes more than 20 hours of audio and more than 200,000 orthographically-transcribed tokens.

Romal Thoppilan, Daniel De Freitas, Jamie Hall et al.

We present LaMDA: Language Models for Dialog Applications. LaMDA is a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. While model scaling alone can improve quality, it shows less improvements on safety and factual grounding. We demonstrate that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding. The first challenge, safety, involves ensuring that the model's responses are consistent with a set of human values, such as preventing harmful suggestions and unfair bias. We quantify safety using a metric based on an illustrative set of human values, and we find that filtering candidate responses using a LaMDA classifier fine-tuned with a small amount of crowdworker-annotated data offers a promising approach to improving model safety. The second challenge, factual grounding, involves enabling the model to consult external knowledge sources, such as an information retrieval system, a language translator, and a calculator. We quantify factuality using a groundedness metric, and we find that our approach enables the model to generate responses grounded in known sources, rather than responses that merely sound plausible. Finally, we explore the use of LaMDA in the domains of education and content recommendations, and analyze their helpfulness and role consistency.

Remi Denton, Mark Díaz, Ian Kivlichan et al.

Human annotations play a crucial role in machine learning (ML) research and development. However, the ethical considerations around the processes and decisions that go into building ML datasets has not received nearly enough attention. In this paper, we survey an array of literature that provides insights into ethical considerations around crowdsourced dataset annotation. We synthesize these insights, and lay out the challenges in this space along two layers: (1) who the annotator is, and how the annotators' lived experiences can impact their annotations, and (2) the relationship between the annotators and the crowdsourcing platforms and what that relationship affords them. Finally, we put forth a concrete set of recommendations and considerations for dataset developers at various stages of the ML data pipeline: task formulation, selection of annotators, platform and infrastructure choices, dataset analysis and evaluation, and dataset documentation and release.

Negar Rostamzadeh, Ben Hutchinson, Christina Greer et al.

Testing practices within the machine learning (ML) community have centered around assessing a learned model's predictive performance measured against a test dataset, often drawn from the same distribution as the training dataset. While recent work on robustness and fairness testing within the ML community has pointed to the importance of testing against distributional shifts, these efforts also focus on estimating the likelihood of the model making an error against a reference dataset/distribution. We argue that this view of testing actively discourages researchers and developers from looking into other sources of robustness failures, for instance corner cases which may have severe undesirable impacts. We draw parallels with decades of work within software engineering testing focused on assessing a software system against various stress conditions, including corner cases, as opposed to solely focusing on average-case behaviour. Finally, we put forth a set of recommendations to broaden the view of machine learning testing to a rigorous practice.

Aida Mostafazadeh Davani, Mark Díaz, Vinodkumar Prabhakaran

Majority voting and averaging are common approaches employed to resolve annotator disagreements and derive single ground truth labels from multiple annotations. However, annotators may systematically disagree with one another, often reflecting their individual biases and values, especially in the case of subjective tasks such as detecting affect, aggression, and hate speech. Annotator disagreements may capture important nuances in such tasks that are often ignored while aggregating annotations to a single ground truth. In order to address this, we investigate the efficacy of multi-annotator models. In particular, our multi-task based approach treats predicting each annotators' judgements as separate subtasks, while sharing a common learned representation of the task. We show that this approach yields same or better performance than aggregating labels in the data prior to training across seven different binary classification tasks. Our approach also provides a way to estimate uncertainty in predictions, which we demonstrate better correlate with annotation disagreements than traditional methods. Being able to model uncertainty is especially useful in deployment scenarios where knowing when not to make a prediction is important.

Vinodkumar Prabhakaran, Aida Mostafazadeh Davani, Mark Díaz

A common practice in building NLP datasets, especially using crowd-sourced annotations, involves obtaining multiple annotator judgements on the same data instances, which are then flattened to produce a single "ground truth" label or score, through majority voting, averaging, or adjudication. While these approaches may be appropriate in certain annotation tasks, such aggregations overlook the socially constructed nature of human perceptions that annotations for relatively more subjective tasks are meant to capture. In particular, systematic disagreements between annotators owing to their socio-cultural backgrounds and/or lived experiences are often obfuscated through such aggregations. In this paper, we empirically demonstrate that label aggregation may introduce representational biases of individual and group perspectives. Based on this finding, we propose a set of recommendations for increased utility and transparency of datasets for downstream use cases.

Sayan Ghosh, Dylan Baker, David Jurgens et al.

Online social media platforms increasingly rely on Natural Language Processing (NLP) techniques to detect abusive content at scale in order to mitigate the harms it causes to their users. However, these techniques suffer from various sampling and association biases present in training data, often resulting in sub-par performance on content relevant to marginalized groups, potentially furthering disproportionate harms towards them. Studies on such biases so far have focused on only a handful of axes of disparities and subgroups that have annotations/lexicons available. Consequently, biases concerning non-Western contexts are largely ignored in the literature. In this paper, we introduce a weakly supervised method to robustly detect lexical biases in broader geocultural contexts. Through a case study on a publicly available toxicity detection model, we demonstrate that our method identifies salient groups of cross-geographic errors, and, in a follow up, demonstrate that these groupings reflect human judgments of offensive and inoffensive language in those geographic contexts. We also conduct analysis of a model trained on a dataset with ground truth labels to better understand these biases, and present preliminary mitigation experiments.

Vinodkumar Prabhakaran, Marek Rei, Ekaterina Shutova

Metaphors are widely used in political rhetoric as an effective framing device. While the efficacy of specific metaphors such as the war metaphor in political discourse has been documented before, those studies often rely on small number of hand-coded instances of metaphor use. Larger-scale topic-agnostic studies are required to establish the general persuasiveness of metaphors as a device, and to shed light on the broader patterns that guide their persuasiveness. In this paper, we present a large-scale data-driven study of metaphors used in political discourse. We conduct this study on a publicly available dataset of over 85K posts made by 412 US politicians in their Facebook public pages, up until Feb 2017. Our contributions are threefold: we show evidence that metaphor use correlates with ideological leanings in complex ways that depend on concurrent political events such as winning or losing elections; we show that posts with metaphors elicit more engagement from their audience overall even after controlling for various socio-political factors such as gender and political party affiliation; and finally, we demonstrate that metaphoricity is indeed the reason for increased engagement of posts, through a fine-grained linguistic analysis of metaphorical vs. literal usages of 513 words across 70K posts.

Nithya Sambasivan, Erin Arnesen, Ben Hutchinson et al.

Conventional algorithmic fairness is West-centric, as seen in its sub-groups, values, and methods. In this paper, we de-center algorithmic fairness and analyse AI power in India. Based on 36 qualitative interviews and a discourse analysis of algorithmic deployments in India, we find that several assumptions of algorithmic fairness are challenged. We find that in India, data is not always reliable due to socio-economic factors, ML makers appear to follow double standards, and AI evokes unquestioning aspiration. We contend that localising model fairness alone can be window dressing in India, where the distance between models and oppressed communities is large. Instead, we re-imagine algorithmic fairness in India and provide a roadmap to re-contextualise data and models, empower oppressed communities, and enable Fair-ML ecosystems.

Nithya Sambasivan, Erin Arnesen, Ben Hutchinson et al.

Conventional algorithmic fairness is Western in its sub-groups, values, and optimizations. In this paper, we ask how portable the assumptions of this largely Western take on algorithmic fairness are to a different geo-cultural context such as India. Based on 36 expert interviews with Indian scholars, and an analysis of emerging algorithmic deployments in India, we identify three clusters of challenges that engulf the large distance between machine learning models and oppressed communities in India. We argue that a mere translation of technical fairness work to Indian subgroups may serve only as a window dressing, and instead, call for a collective re-imagining of Fair-ML, by re-contextualising data and models, empowering oppressed communities, and more importantly, enabling ecosystems.

Dorottya Demszky, Devyani Sharma, Jonathan H. Clark et al.

Building NLP systems that serve everyone requires accounting for dialect differences. But dialects are not monolithic entities: rather, distinctions between and within dialects are captured by the presence, absence, and frequency of dozens of dialect features in speech and text, such as the deletion of the copula in "He {} running". In this paper, we introduce the task of dialect feature detection, and present two multitask learning approaches, both based on pretrained transformers. For most dialects, large-scale annotated corpora for these features are unavailable, making it difficult to train recognizers. We train our models on a small number of minimal pairs, building on how linguists typically define dialect features. Evaluation on a test set of 22 dialect features of Indian English demonstrates that these models learn to recognize many features with high accuracy, and that a few minimal pairs can be as effective for training as thousands of labeled examples. We also demonstrate the downstream applicability of dialect feature detection both as a measure of dialect density and as a dialect classifier.

Donald Martin, Vinodkumar Prabhakaran, Jill Kuhlberg et al.

Machine learning (ML) fairness research tends to focus primarily on mathematically-based interventions on often opaque algorithms or models and/or their immediate inputs and outputs. Such oversimplified mathematical models abstract away the underlying societal context where ML models are conceived, developed, and ultimately deployed. As fairness itself is a socially constructed concept that originates from that societal context along with the model inputs and the models themselves, a lack of an in-depth understanding of societal context can easily undermine the pursuit of ML fairness. In this paper, we outline three new tools to improve the comprehension, identification and representation of societal context. First, we propose a complex adaptive systems (CAS) based model and definition of societal context that will help researchers and product developers to expand the abstraction boundary of ML fairness work to include societal context. Second, we introduce collaborative causal theory formation (CCTF) as a key capability for establishing a sociotechnical frame that incorporates diverse mental models and associated causal theories in modeling the problem and solution space for ML-based products. Finally, we identify community based system dynamics (CBSD) as a powerful, transparent and rigorous approach for practicing CCTF during all phases of the ML product development process. We conclude with a discussion of how these systems theoretic approaches to understand the societal context within which sociotechnical systems are embedded can improve the development of fair and inclusive ML-based products.

Donald Martin, Vinodkumar Prabhakaran, Jill Kuhlberg et al.

Recent research on algorithmic fairness has highlighted that the problem formulation phase of ML system development can be a key source of bias that has significant downstream impacts on ML system fairness outcomes. However, very little attention has been paid to methods for improving the fairness efficacy of this critical phase of ML system development. Current practice neither accounts for the dynamic complexity of high-stakes domains nor incorporates the perspectives of vulnerable stakeholders. In this paper we introduce community based system dynamics (CBSD) as an approach to enable the participation of typically excluded stakeholders in the problem formulation phase of the ML system development process and facilitate the deep problem understanding required to mitigate bias during this crucial stage.

Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton et al.

Building equitable and inclusive NLP technologies demands consideration of whether and how social attitudes are represented in ML models. In particular, representations encoded in models often inadvertently perpetuate undesirable social biases from the data on which they are trained. In this paper, we present evidence of such undesirable biases towards mentions of disability in two different English language models: toxicity prediction and sentiment analysis. Next, we demonstrate that the neural embeddings that are the critical first step in most NLP pipelines similarly contain undesirable biases towards mentions of disability. We end by highlighting topical biases in the discourse about disability which may contribute to the observed model biases; for instance, gun violence, homelessness, and drug addiction are over-represented in texts discussing mental illness.

Vinodkumar Prabhakaran, Ben Hutchinson, Margaret Mitchell

Data-driven statistical Natural Language Processing (NLP) techniques leverage large amounts of language data to build models that can understand language. However, most language data reflect the public discourse at the time the data was produced, and hence NLP models are susceptible to learning incidental associations around named referents at a particular point in time, in addition to general linguistic meaning. An NLP system designed to model notions such as sentiment and toxicity should ideally produce scores that are independent of the identity of such entities mentioned in text and their social associations. For example, in a general purpose sentiment analysis system, a phrase such as I hate Katy Perry should be interpreted as having the same sentiment as I hate Taylor Swift. Based on this idea, we propose a generic evaluation framework, Perturbation Sensitivity Analysis, which detects unintended model biases related to named entities, and requires no new annotations or corpora. We demonstrate the utility of this analysis by employing it on two different NLP models --- a sentiment model and a toxicity model --- applied on online comments in English language from four different genres.

Michelle Lam, Catherina Xu, Angela Kong et al.

Can language analysis reveal the underlying social power relations that exist between participants of an interaction? Prior work within NLP has shown promise in this area, but the performance of automatically predicting power relations using NLP analysis of social interactions remains wanting. In this paper, we present a novel neural architecture that captures manifestations of power within individual emails which are then aggregated in an order-preserving way in order to infer the direction of power between pairs of participants in an email thread. We obtain an accuracy of 80.4%, a 10.1% improvement over state-of-the-art methods, in this task. We further apply our model to the task of predicting power relations between individuals based on the entire set of messages exchanged between them; here also, our model significantly outperforms the70.0% accuracy using prior state-of-the-art techniques, obtaining an accuracy of 83.0%.

Vinodkumar Prabhakaran, Premkumar Ganeshkumar, Owen Rambow

Understanding how social power structures affect the way we interact with one another is of great interest to social scientists who want to answer fundamental questions about human behavior, as well as to computer scientists who want to build automatic methods to infer the social contexts of interactions. In this paper, we employ advancements in extra-propositional semantics extraction within NLP to study how author commitment reflects the social context of an interaction. Specifically, we investigate whether the level of commitment expressed by individuals in an organizational interaction reflects the hierarchical power structures they are part of. We find that subordinates use significantly more instances of non-commitment than superiors. More importantly, we also find that subordinates attribute propositions to other agents more often than superiors do --- an aspect that has not been studied before. Finally, we show that enriching lexical features with commitment labels captures important distinctions in social meanings.

Vinodkumar Prabhakaran, Owen Rambow

Understanding how the social context of an interaction affects our dialog behavior is of great interest to social scientists who study human behavior, as well as to computer scientists who build automatic methods to infer those social contexts. In this paper, we study the interaction of power, gender, and dialog behavior in organizational interactions. In order to perform this study, we first construct the Gender Identified Enron Corpus of emails, in which we semi-automatically assign the gender of around 23,000 individuals who authored around 97,000 email messages in the Enron corpus. This corpus, which is made freely available, is orders of magnitude larger than previously existing gender identified corpora in the email domain. Next, we use this corpus to perform a large-scale data-oriented study of the interplay of gender and manifestations of power. We argue that, in addition to one's own gender, the "gender environment" of an interaction, i.e., the gender makeup of one's interlocutors, also affects the way power is manifested in dialog. We focus especially on manifestations of power in the dialog structure --- both, in a shallow sense that disregards the textual content of messages (e.g., how often do the participants contribute, how often do they get replies etc.), as well as the structure that is expressed within the textual content (e.g., who issues requests and how are they made, whose requests get responses etc.). We find that both gender and gender environment affect the ways power is manifested in dialog, resulting in patterns that reveal the underlying factors. Finally, we show the utility of gender information in the problem of automatically predicting the direction of power between pairs of participants in email interactions.

Vinodkumar Prabhakaran, Michael Bloodgood, Mona Diab et al.

We explore training an automatic modality tagger. Modality is the attitude that a speaker might have toward an event or state. One of the main hurdles for training a linguistic tagger is gathering training data. This is particularly problematic for training a tagger for modality because modality triggers are sparse for the overwhelming majority of sentences. We investigate an approach to automatically training a modality tagger where we first gathered sentences based on a high-recall simple rule-based modality tagger and then provided these sentences to Mechanical Turk annotators for further annotation. We used the resulting set of training data to train a precise modality tagger using a multi-class SVM that delivers good performance.