Research Papers

Bridging the Data Provenance Gap Across Text, Speech and Video

Shayne Longpre, Nikhil Singh, Manuel Cherep +40

Progress in AI is driven largely by the scale and quality of training data. Despite this, there is a deficit of empirical analysis examining the attributes of well-established datasets beyond text. In this work we conduct the largest and first-of-its-kind longitudinal audit across modalities--popular text, speech, and video datasets--from their detailed sourcing trends and use restrictions to their geographical and linguistic representation. Our manual analysis covers nearly 4000 public datasets between 1990-2024, spanning 608 languages, 798 sources, 659 organizations, and 67 countries. We find that multimodal machine learning applications have overwhelmingly turned to web-crawled, synthetic, and social media platforms, such as YouTube, for their training sets, eclipsing all other sources since 2019. Secondly, tracing the chain of dataset derivations we find that while less than 33% of datasets are restrictively licensed, over 80% of the source content in widely-used text, speech, and video datasets, carry non-commercial restrictions. Finally, counter to the rising number of languages and geographies represented in public AI training datasets, our audit demonstrates measures of relative geographical and multilingual representation have failed to significantly improve their coverage since 2013. We believe the breadth of our audit enables us to empirically examine trends in data sourcing, restrictions, and Western-centricity at an ecosystem-level, and that visibility into these questions are essential to progress in responsible AI. As a contribution to ongoing improvements in dataset transparency and responsible use, we release our entire multimodal audit, allowing practitioners to trace data provenance across text, speech, and video.

The use of machine learning and AI on electronic health records (EHRs) holds substantial potential for clinical insight. However, this approach faces challenges due to data heterogeneity, sparsity, temporal misalignment, and limited labeled outcomes. In this context, we leverage a linked EHR dataset of approximately one million de-identified individuals from Bristol, North Somerset, and South Gloucestershire, UK, to characterize urinary tract infections (UTIs). We implemented a data pre-processing and curation pipeline that transforms the raw EHR data into a structured format suitable for developing predictive models focused on data fairness, accountability and transparency. Given the limited availability and biases of ground truth UTI outcomes, we introduce a UTI risk estimation framework informed by clinical expertise to estimate UTI risk across individual patient timelines. Pairwise XGBoost models are trained using this framework to differentiate UTI risk categories with explainable AI techniques applied to identify key predictors and support interpretability. Our findings reveal differences in clinical and demographic predictors across risk groups. While this study highlights the potential of AI-driven insights to support UTI clinical decision-making, further investigation of patient sub-strata and extensive validation are needed to ensure robustness and applicability in clinical practice.

As data-driven and AI-based decision making gains widespread adoption across disciplines, it is crucial that both data privacy and decision fairness are appropriately addressed. Although differential privacy (DP) provides a robust framework for guaranteeing privacy and methods are available to improve fairness, most prior work treats the two concerns separately. Even though there are existing approaches that consider privacy and fairness simultaneously, they typically focus on a single specific learning task, limiting their generalizability. In response, we introduce SAFES, a Sequential PrivAcy and Fairness Enhancing data Synthesis procedure that sequentially combines DP data synthesis with a fairness-aware data preprocessing step. SAFES allows users flexibility in navigating the privacy-fairness-utility trade-offs. We illustrate SAFES with different DP synthesizers and fairness-aware data preprocessing methods and run extensive experiments on multiple real datasets to examine the privacy-fairness-utility trade-offs of synthetic data generated by SAFES. Empirical evaluations demonstrate that for reasonable privacy loss, SAFES-generated synthetic data can achieve significantly improved fairness metrics with relatively low utility loss.

Large Language Models (LLMs) often perpetuate biases in pronoun usage, leading to misrepresentation or exclusion of queer individuals. This paper addresses the specific problem of biased pronoun usage in LLM outputs, particularly the inappropriate use of traditionally gendered pronouns ("he," "she") when inclusive language is needed to accurately represent all identities. We introduce a collaborative agent pipeline designed to mitigate these biases by analyzing and optimizing pronoun usage for inclusivity. Our multi-agent framework includes specialized agents for both bias detection and correction. Experimental evaluations using the Tango dataset-a benchmark focused on gender pronoun usage-demonstrate that our approach significantly improves inclusive pronoun classification, achieving a 32.6 percentage point increase over GPT-4o in correctly disagreeing with inappropriate traditionally gendered pronouns $(χ^2 = 38.57, p < 0.0001)$. These results accentuate the potential of agent-driven frameworks in enhancing fairness and inclusivity in AI-generated content, demonstrating their efficacy in reducing biases and promoting socially responsible AI.

Data augmentation has become a pivotal tool in enhancing the performance of computer vision tasks, with the KeepOriginalAugment method emerging as a standout technique for its intelligent incorporation of salient regions within less prominent areas, enabling augmentation in both regions. Despite its success in image classification, its potential in addressing biases remains unexplored. In this study, we introduce an extension of the KeepOriginalAugment method, termed FaceKeepOriginalAugment, which explores various debiasing aspects-geographical, gender, and stereotypical biases-in computer vision models. By maintaining a delicate balance between data diversity and information preservation, our approach empowers models to exploit both diverse salient and non-salient regions, thereby fostering increased diversity and debiasing effects. We investigate multiple strategies for determining the placement of the salient region and swapping perspectives to decide which part undergoes augmentation. Leveraging the Image Similarity Score (ISS), we quantify dataset diversity across a range of datasets, including Flickr Faces HQ (FFHQ), WIKI, IMDB, Labelled Faces in the Wild (LFW), UTK Faces, and Diverse Dataset. We evaluate the effectiveness of FaceKeepOriginalAugment in mitigating gender bias across CEO, Engineer, Nurse, and School Teacher datasets, utilizing the Image-Image Association Score (IIAS) in convolutional neural networks (CNNs) and vision transformers (ViTs). Our findings shows the efficacy of FaceKeepOriginalAugment in promoting fairness and inclusivity within computer vision models, demonstrated by reduced gender bias and enhanced overall fairness. Additionally, we introduce a novel metric, Saliency-Based Diversity and Fairness Metric, which quantifies both diversity and fairness while handling data imbalance across various datasets.

Representational harms in language technologies often occur in short spans within otherwise neutral text, where phrases may simultaneously convey generalizations, unfairness, or stereotypes. Framing bias detection as sentence-level classification obscures which words carry bias and what type is present, limiting both auditability and targeted mitigation. We introduce the GUS-Net Framework, comprising the GUS dataset and a multi-label token-level detector for span-level analysis of social bias. The GUS dataset contains 3,739 unique snippets across multiple domains, with over 69,000 token-level annotations. Each token is labeled using BIO tags (Begin, Inside, Outside) for three pathways of representational harm: Generalizations, Unfairness, and Stereotypes. To ensure reliable data annotation, we employ an automated multi-agent pipeline that proposes candidate spans which are subsequently verified and corrected by human experts. We formulate bias detection as multi-label token-level classification and benchmark both encoder-based models (e.g., BERT family variants) and decoder-based large language models (LLMs). Our evaluations cover token-level identification and span-level entity recognition on our test set, and out-of-distribution generalization. Empirical results show that encoder-based models consistently outperform decoder-based baselines on nuanced and overlapping spans while being more computationally efficient. The framework delivers interpretable, fine-grained diagnostics that enable systematic auditing and mitigation of representational harms in real-world NLP systems.

The Technology of Outrage: Bias in Artificial Intelligence

Will Bridewell, Paul F. Bello, Selmer Bringsjord

Artificial intelligence and machine learning are increasingly used to offload decision making from people. In the past, one of the rationales for this replacement was that machines, unlike people, can be fair and unbiased. Evidence suggests otherwise. We begin by entertaining the ideas that algorithms can replace people and that algorithms cannot be biased. Taken as axioms, these statements quickly lead to absurdity. Spurred on by this result, we investigate the slogans more closely and identify equivocation surrounding the word 'bias.' We diagnose three forms of outrage-intellectual, moral, and political-that are at play when people react emotionally to algorithmic bias. Then we suggest three practical approaches to addressing bias that the AI community could take, which include clarifying the language around bias, developing new auditing methods for intelligent systems, and building certain capabilities into these systems. We conclude by offering a moral regarding the conversations about algorithmic bias that may transfer to other areas of artificial intelligence.

AI fairness measurements, including tests for equal treatment, often take the form of disaggregated evaluations of AI systems. Such measurements are an important part of Responsible AI operations. These measurements compare system performance across demographic groups or sub-populations and typically require member-level demographic signals such as gender, race, ethnicity, and location. However, sensitive member-level demographic attributes like race and ethnicity can be challenging to obtain and use due to platform choices, legal constraints, and cultural norms. In this paper, we focus on the task of enabling AI fairness measurements on race/ethnicity for \emph{U.S. LinkedIn members} in a privacy-preserving manner. We present the Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) method for performing this task. PPRE combines the Bayesian Improved Surname Geocoding (BISG) model, a sparse LinkedIn survey sample of self-reported demographics, and privacy-enhancing technologies like secure two-party computation and differential privacy to enable meaningful fairness measurements while preserving member privacy. We provide details of the PPRE method and its privacy guarantees. We then illustrate sample measurement operations. We conclude with a review of open research and engineering challenges for expanding our privacy-preserving fairness measurement capabilities.

Our society collects data on people for a wide range of applications, from building a census for policy evaluation to running meaningful clinical trials. To collect data, we typically sample individuals with the goal of accurately representing a population of interest. However, current sampling processes often collect data opportunistically from data sources, which can lead to datasets that are biased and not representative, i.e., the collected dataset does not accurately reflect the distribution of demographics of the true population. This is a concern because subgroups within the population can be under- or over-represented in a dataset, which may harm generalizability and lead to an unequal distribution of benefits and harms from downstream tasks that use such datasets (e.g., algorithmic bias in medical decision-making algorithms). In this paper, we assess the relationship between dataset representativeness and group-fairness of classifiers trained on that dataset. We demonstrate that there is a natural tension between dataset representativeness and classifier fairness; empirically we observe that training datasets with better representativeness can frequently result in classifiers with higher rates of unfairness. We provide some intuition as to why this occurs via a set of theoretical results in the case of univariate classifiers. We also find that over-sampling underrepresented groups can result in classifiers which exhibit greater bias to those groups. Lastly, we observe that fairness-aware sampling strategies (i.e., those which are specifically designed to select data with high downstream fairness) will often over-sample members of majority groups. These results demonstrate that the relationship between dataset representativeness and downstream classifier fairness is complex; balancing these two quantities requires special care from both model- and dataset-designers.

This study addresses gender bias in image generation models using Reinforcement Learning from Artificial Intelligence Feedback (RLAIF) with a novel Denoising Diffusion Policy Optimization (DDPO) pipeline. By employing a pretrained stable diffusion model and a highly accurate gender classification Transformer, the research introduces two reward functions: Rshift for shifting gender imbalances, and Rbalance for achieving and maintaining gender balance. Experiments demonstrate the effectiveness of this approach in mitigating bias without compromising image quality or requiring additional data or prompt modifications. While focusing on gender bias, this work establishes a foundation for addressing various forms of bias in AI systems, emphasizing the need for responsible AI development. Future research directions include extending the methodology to other bias types, enhancing the RLAIF pipeline's robustness, and exploring multi-prompt fine-tuning to further advance fairness and inclusivity in AI.

In Explainable AI, rule extraction translates model knowledge into logical rules, such as IF-THEN statements, crucial for understanding patterns learned by black-box models. This could significantly aid in fields like disease diagnosis, disease progression estimation, or drug discovery. However, such application domains often contain imbalanced data, with the class of interest underrepresented. Existing methods inevitably compromise the performance of rules for the minor class to maximise the overall performance. As the first attempt in this field, we propose a model-agnostic approach for extracting rules from specific subgroups of data, featuring automatic rule generation for numerical features. This method enhances the regional explainability of machine learning models and offers wider applicability compared to existing methods. We additionally introduce a new method for selecting features to compose rules, reducing computational costs in high-dimensional spaces. Experiments across various datasets and models demonstrate the effectiveness of our methods.

Deep learning models have achieved remarkable success across diverse domains. However, the intricate nature of these models often impedes a clear understanding of their decision-making processes. This is where Explainable AI (XAI) becomes indispensable, offering intuitive explanations for model decisions. In this work, we propose a simple yet highly effective approach, ScoreCAM++, which introduces modifications to enhance the promising ScoreCAM method for visual explainability. Our proposed approach involves altering the normalization function within the activation layer utilized in ScoreCAM, resulting in significantly improved results compared to previous efforts. Additionally, we apply an activation function to the upsampled activation layers to enhance interpretability. This improvement is achieved by selectively gating lower-priority values within the activation layer. Through extensive experiments and qualitative comparisons, we demonstrate that ScoreCAM++ consistently achieves notably superior performance and fairness in interpreting the decision-making process compared to both ScoreCAM and previous methods.