Most ML datasets today contain biases. When we train models on these datasets, they often not only learn these biases but can worsen them -- a phenomenon known as bias amplification. Several co-occurrence-based metrics have been proposed to measure bias amplification in classification datasets. They measure bias amplification between a protected attribute (e.g., gender) and a task (e.g., cooking). These metrics also support fine-grained bias analysis by identifying the direction in which a model amplifies biases. However, co-occurrence-based metrics have limitations -- some fail to measure bias amplification in balanced datasets, while others fail to measure negative bias amplification.

Most ML datasets today contain biases. When we train models on these datasets, they often not only learn these biases but can worsen them --- a phenomenon known as bias amplification. Several co-occurrence-based metrics have been proposed to measure bias amplification in classification datasets. They measure bias amplification between a protected attribute (e.g., gender) and a task (e.g., cooking). These metrics also support fine-grained bias analysis by identifying the direction in which a model amplifies biases. However, co-occurrence-based metrics have limitations --- some fail to measure bias amplification in balanced datasets, while others fail to measure negative bias amplification.

Despite growing empirical evidence of bias amplification in machine learning, its theoretical foundations remain poorly understood. We develop a formal framework for majority-minority learning tasks, showing how standard training can favor majority groups and produce stereotypical predictors that neglect minority-specific features. Assuming population and variance imbalance, our analysis reveals three key findings: (i) the close proximity between full-data and stereotypical predictors, (ii) the dominance of a region where training the entire model tends to merely learn the majority traits, and (iii) a lower bound on the additional training required. Our results are illustrated through experiments in deep learning for tabular and image classification tasks.

As society increasingly relies on AI-based tools for decision-making in socially sensitive domains, investigating fairness and equity of such automated systems has become a critical field of inquiry. Most of the literature in fair machine learning focuses on defining and achieving fairness criteria in the context of prediction, while not explicitly focusing on how these predictions may be used later on in the pipeline. For instance, if commonly used criteria, such as independence or sufficiency, are satisfied for a prediction score $S$ used for binary classification, they need not be satisfied after an application of a simple thresholding operation on $S$ (as commonly used in practice). In this paper, we take an important step to address this issue in numerous statistical and causal notions of fairness. We introduce the notion of a margin complement, which measures how much a prediction score $S$ changes due to a thresholding operation.We then demonstrate that the marginal difference in the optimal 0/1 predictor $\widehat Y$ between groups, written $P(\hat y \mid x_1) - P(\hat y \mid x_0)$, can be causally decomposed into the influences of $X$ on the $L_2$-optimal prediction score $S$ and the influences of $X$ on the margin complement $M$, along different causal pathways (direct, indirect, spurious). We then show that under suitable causal assumptions, the influences of $X$ on the prediction score $S$ are equal to the influences of $X$ on the true outcome $Y$. This yields a new decomposition of the disparity in the predictor $\widehat Y$ that allows us to disentangle causal differences inherited from the true outcome $Y$ that exists in the real world vs. those coming from the optimization procedure itself. This observation highlights the need for more regulatory oversight due to the potential for bias amplification, and to address this issue we introduce new notions of weak and strong business necessity, together with an algorithm for assessing whether these notions are satisfied. We apply our method to three real-world datasets and derive new insights on bias amplification in prediction and decision-making.

Despite growing empirical evidence of bias amplification in machine learning, its theoretical foundations remain poorly understood. We develop a formal framework for majority-minority learning tasks, showing how standard training can favor majority groups and produce stereotypical predictors that neglect minority-specific features. Assuming population and variance imbalance, our analysis reveals three key findings: (i) the close proximity between ``full-data'' and stereotypical predictors, (ii) the dominance of a region where training the entire model tends to merely learn the majority traits, and (iii) a lower bound on the additional training required. Our results are illustrated through experiments in deep learning for tabular and image classification tasks.

Large multimodal models (LMMs) have revolutionized text-to-image generation, but they risk perpetuating the harmful social biases in their training data. Prior work has identified gender bias in these models, but methodological limitations prevented large-scale, comparable, cross-model analysis. To address this gap, we introduce the Aymara Image Fairness Evaluation, a benchmark for assessing social bias in AI-generated images. We test 13 commercially available LMMs using 75 procedurally-generated, gender-neutral prompts to generate people in stereotypically-male, stereotypically-female, and non-stereotypical professions. We then use a validated LLM-as-a-judge system to score the 965 resulting images for gender representation. Our results reveal (p < .001 for all): 1) LMMs systematically not only reproduce but actually amplify occupational gender stereotypes relative to real-world labor data, generating men in 93.0% of images for male-stereotyped professions but only 22.5% for female-stereotyped professions; 2) Models exhibit a strong default-male bias, generating men in 68.3% of the time for non-stereotyped professions; and 3) The extent of bias varies dramatically across models, with overall male representation ranging from 46.7% to 73.3%. Notably, the top-performing model de-amplified gender stereotypes and approached gender parity, achieving the highest fairness scores. This variation suggests high bias is not an inevitable outcome but a consequence of design choices. Our work provides the most comprehensive cross-model benchmark of gender bias to date and underscores the necessity of standardized, automated evaluation tools for promoting accountability and fairness in AI development.

As society increasingly relies on AI-based tools for decision-making in socially sensitive domains, investigating fairness and equity of such automated systems has become a critical field of inquiry. Most of the literature in fair machine learning focuses on defining and achieving fairness criteria in the context of prediction, while not explicitly focusing on how these predictions may be used later on in the pipeline. For instance, if commonly used criteria, such as independence or sufficiency, are satisfied for a prediction score S used for binary classification, they need not be satisfied after an application of a simple thresholding operation on S (as commonly used in practice). In this paper, we take an important step to address this issue in numerous statistical and causal notions of fairness. We introduce the notion of a margin complement, which measures how much a prediction score S changes due to a thresholding operation.We then demonstrate that the marginal difference in the optimal 0/1 predictor \widehat Y between groups, written P(\hat y \mid x_1) - P(\hat y \mid x_0), can be causally decomposed into the influences of X on the L_2 -optimal prediction score S and the influences of X on the margin complement M, along different causal pathways (direct, indirect, spurious).

With the widespread adoption of Large Language Models (LLMs), the prevalence of iterative interactions among these models is anticipated to increase. Notably, recent advancements in multi-round on-policy self-improving methods allow LLMs to generate new examples for training subsequent models. At the same time, multi-agent LLM systems, involving automated interactions among agents, are also increasing in prominence. Thus, in both short and long terms, LLMs may actively engage in an evolutionary process. We draw parallels between the behavior of LLMs and the evolution of human culture, as the latter has been extensively studied by cognitive scientists for decades. Our approach involves leveraging Iterated Learning (IL), a Bayesian framework that elucidates how subtle biases are magnified during human cultural evolution, to explain some behaviors of LLMs.

When we train models on biased ML datasets, they not only learn these biases but can inflate them at test time - a phenomenon called bias amplification. To measure bias amplification in ML datasets, many co-occurrence-based metrics have been proposed. Co-occurrence-based metrics are effective in measuring bias amplification in simple problems like image classification. However, these metrics are ineffective for complex problems like image captioning as they cannot capture the semantics of a caption. To measure bias amplification in captions, prior work introduced a predictability-based metric called Leakage in Captioning (LIC). While LIC captures the semantics and context of captions, it has limitations. LIC cannot identify the direction in which bias is amplified, poorly estimates dataset bias due to a weak vocabulary substitution strategy, and is highly sensitive to attacker models (a hyperparameter in predictability-based metrics). To overcome these issues, we propose Directional Predictability Amplification in Captioning (DPAC). DPAC measures directional bias amplification in captions, provides a better estimate of dataset bias using an improved substitution strategy, and is less sensitive to attacker models. Our experiments on the COCO captioning dataset show how DPAC is the most reliable metric to measure bias amplification in captions.

Multimodal machine learning models, such as those that combine text and image modalities, are increasingly used in critical domains including public safety, security, and healthcare. However, these systems inherit biases from their single modalities. This study proposes a systemic framework for analyzing dynamic multimodal bias interactions. Using the MMBias dataset, which encompasses categories prone to bias such as religion, nationality, and sexual orientation, this study adopts a simulation-based heuristic approach to compute bias scores for text-only, image-only, and multimodal embeddings. A framework is developed to classify bias interactions as amplification (multimodal bias exceeds both unimodal biases), mitigation (multimodal bias is lower than both), and neutrality (multimodal bias lies between unimodal biases), with proportional analyzes conducted to identify the dominant mode and dynamics in these interactions. The findings highlight that amplification (22\%) occurs when text and image biases are comparable, while mitigation (11\%) arises under the dominance of text bias, highlighting the stabilizing role of image bias. Neutral interactions (67\%) are related to a higher text bias without divergence. Conditional probabilities highlight the text's dominance in mitigation and mixed contributions in neutral and amplification cases, underscoring complex modality interplay. In doing so, the study encourages the use of this heuristic, systemic, and interpretable framework to analyze multimodal bias interactions, providing insight into how intermodal biases dynamically interact, with practical applications for multimodal modeling and transferability to context-based datasets, all essential for developing fair and equitable AI models.