Racial disparities in healthcare expenditures are well-documented, yet the underlying drivers remain complex and require further investigation. This study employs causal and counterfactual path-specific effects to quantify how various factors, including socioeconomic status, insurance access, health behaviors, and health status, mediate these disparities. Using data from the Medical Expenditures Panel Survey, we estimate how expenditures would differ under counterfactual scenarios in which the values of specific mediators were aligned across racial groups along selected causal pathways. A key challenge in this analysis is ensuring robustness against model misspecification while addressing the zero-inflation and right-skewness of healthcare expenditures. For reliable inference, we derive asymptotically linear estimators by integrating influence function-based techniques with flexible machine learning methods, including super learners and a two-part model tailored to the zero-inflated, right-skewed nature of healthcare expenditures.

The expansion of large-scale text-to-image diffusion models has raised growing concerns about their potential to generate undesirable or harmful content, ranging from fabricated depictions of public figures to sexually explicit images. To mitigate these risks, prior work has devised machine unlearning techniques that attempt to erase unwanted concepts through fine-tuning. However, in this paper, we introduce a new threat model, Toxic Erasure (ToxE), and demonstrate how recent unlearning algorithms, including those explicitly designed for robustness, can be circumvented through targeted backdoor attacks. The threat is realized by establishing a link between a trigger and the undesired content. Subsequent unlearning attempts fail to erase this link, allowing adversaries to produce harmful content. We instantiate ToxE via two established backdoor attacks: one targeting the text encoder and another manipulating the cross-attention layers. Further, we introduce Deep Intervention Score-based Attack (DISA), a novel, deeper backdoor attack that optimizes the entire U-Net using a score-based objective, improving the attack's persistence across different erasure methods. We evaluate five recent concept erasure methods against our threat model. For celebrity identity erasure, our deep attack circumvents erasure with up to 82% success, averaging 57% across all erasure methods. For explicit content erasure, ToxE attacks can elicit up to 9 times more exposed body parts, with DISA yielding an average increase by a factor of 2.9. These results highlight a critical security gap in current unlearning strategies.

Governance institutions must respond to societal risks, including those posed by generative AI. This study empirically examines how public trust in institutions and AI technologies, along with perceived risks, shape preferences for AI regulation. Using the nationally representative 2023 Artificial Intelligence, Morality, and Sentience (AIMS) survey, we assess trust in government, AI companies, and AI technologies, as well as public support for regulatory measures such as slowing AI development or outright bans on advanced AI. Our findings reveal broad public support for AI regulation, with risk perception playing a significant role in shaping policy preferences. Individuals with higher trust in government favor regulation, while those with greater trust in AI companies and AI technologies are less inclined to support restrictions. Trust in government and perceived risks significantly predict preferences for both soft (e.g., slowing development) and strong (e.g., banning AI systems) regulatory interventions. These results highlight the importance of public opinion in AI governance. As AI capabilities advance, effective regulation will require balancing public concerns about risks with trust in institutions. This study provides a foundational empirical baseline for policymakers navigating AI governance and underscores the need for further research into public trust, risk perception, and regulatory strategies in the evolving AI landscape.

Fairness auditing of AI systems can identify and quantify biases. However, traditional auditing using real-world data raises security and privacy concerns. It exposes auditors to security risks as they become custodians of sensitive information and targets for cyberattacks. Privacy risks arise even without direct breaches, as data analyses can inadvertently expose confidential information. To address these, we propose a framework that leverages differentially private synthetic data to audit the fairness of AI systems. By applying privacy-preserving mechanisms, it generates synthetic data that mirrors the statistical properties of the original dataset while ensuring privacy. This method balances the goal of rigorous fairness auditing and the need for strong privacy protections. Through experiments on real datasets like Adult, COMPAS, and Diabetes, we compare fairness metrics of synthetic and real data. By analyzing the alignment and discrepancies between these metrics, we assess the capacity of synthetic data to preserve the fairness properties of real data. Our results demonstrate the framework's ability to enable meaningful fairness evaluations while safeguarding sensitive information, proving its applicability across critical and sensitive domains.

--The use of machine learning in cyber-physical systems has attracted the interest of both industry and academia. However, no general solution has yet been found against the unpredictable behavior of neural networks and reinforcement learning agents. Nevertheless, the improvements of photo-realistic simulators have paved the way towards extensive testing of complex algorithms in different virtual scenarios, which would be expensive and dangerous to implement in the real world. PRIVE, a simulation framework for physical robot interaction with virtual environments, which operates as a vehicle-in-the-loop platform, rendering a virtual world while operating the vehicle in the real world. PRIVE, any physical mobile robot running on ROS 2 can easily be configured to move its digital twin in a virtual world built with the Unreal Engine 5 graphic engine, which can be populated with objects, people, or other vehicles with programmable behavior . PRIVE has been designed to accommodate custom or pre-built virtual worlds while being light-weight to contain execution times and allow fast rendering. Its main advantage lies in the possibility of testing complex algorithms on the full software and hardware stack while minimizing the risks and costs of a test campaign. The framework has been validated by testing a reinforcement learning agent trained for obstacle avoidance on an AgileX Scout Mini rover that navigates a virtual office environment where everyday objects and people are placed as obstacles. Recently, artificial intelligence (AI) models have achieved impressive performance in many applications, including cyber-physical systems (CPS).

Generative artificial intelligence (AI), particularly large language models (LLMs), is being rapidly deployed in recruitment and for candidate shortlisting. We audit several mid-sized open-source LLMs for gender bias using a dataset of 332,044 real-world online job postings. For each posting, we prompt the model to recommend whether an equally qualified male or female candidate should receive an interview callback. We find that most models tend to favor men, especially for higher-wage roles. Mapping job descriptions to the Standard Occupational Classification system, we find lower callback rates for women in male-dominated occupations and higher rates in female-associated ones, indicating occupational segregation. A comprehensive analysis of linguistic features in job ads reveals strong alignment of model recommendations with traditional gender stereotypes. To examine the role of recruiter identity, we steer model behavior by infusing Big Five personality traits and simulating the perspectives of historical figures. We find that less agreeable personas reduce stereotyping, consistent with an agreeableness bias in LLMs. Our findings highlight how AI-driven hiring may perpetuate biases in the labor market and have implications for fairness and diversity within firms.

Accurate imputation of race and ethnicity (R&E) is crucial for analyzing disparities and informing policy. Methods like Bayesian Improved Surname Geocoding (BISG) are widely used but exhibit limitations, including systematic misclassification biases linked to socioeconomic status. This paper introduces LSTM+Geo, a novel approach enhancing Long Short-Term Memory (LSTM) networks with census tract geolocation information. Using a large voter dataset, we demonstrate that LSTM+Geo (88.7% accuracy) significantly outperforms standalone LSTM (86.4%) and Bayesian methods like BISG (82.9%) and BIFSG (86.8%) in accuracy and F1-score on a held-out validation set. LSTM+Geo reduces the rate at which non-White individuals are misclassified as White (White FPR 19.3%) compared to name-only LSTMs (White FPR 24.6%). While sophisticated ensemble methods incorporating XGBoost achieve the highest overall accuracy (up to 89.4%) and lowest White FPR (17.8%), LSTM+Geo offers strong standalone performance with improved bias characteristics compared to baseline models. Integrating LSTM+Geo into an XGBoost ensemble further boosts accuracy, highlighting its utility as both a standalone model and a component for advanced systems. We give a caution at the end regarding the appropriate use of these methods.

Large Language Models (LLMs) currently struggle to sequentially add new memories and integrate new knowledge. These limitations contrast with the human ability to continuously learn from new experiences and acquire knowledge throughout life. Most existing approaches add memories either through large context windows or external memory buffers (e.g., Retrieval-Augmented Generation), and studies on knowledge injection rarely test scenarios resembling everyday life events. In this work, we introduce a continual learning framework, Memory Embedded in Gated LLMs (MEGa), which injects event memories directly into the weights of LLMs. Each memory is stored in a dedicated set of gated low-rank weights. During inference, a gating mechanism activates relevant memory weights by matching query embeddings to stored memory embeddings. This enables the model to both recall entire memories and answer related questions. On two datasets - fictional characters and Wikipedia events - MEGa outperforms baseline approaches in mitigating catastrophic forgetting. Our model draws inspiration from the complementary memory system of the human brain.

Wildlife trafficking remains a critical global issue, significantly impacting biodiversity, ecological stability, and public health. Despite efforts to combat this illicit trade, the rise of e-commerce platforms has made it easier to sell wildlife products, putting new pressure on wild populations of endangered and threatened species. The use of these platforms also opens a new opportunity: as criminals sell wildlife products online, they leave digital traces of their activity that can provide insights into trafficking activities as well as how they can be disrupted. The challenge lies in finding these traces. Online marketplaces publish ads for a plethora of products, and identifying ads for wildlife-related products is like finding a needle in a haystack. Learning classifiers can automate ad identification, but creating them requires costly, time-consuming data labeling that hinders support for diverse ads and research questions. This paper addresses a critical challenge in the data science pipeline for wildlife trafficking analytics: generating quality labeled data for classifiers that select relevant data. While large language models (LLMs) can directly label advertisements, doing so at scale is prohibitively expensive. We propose a cost-effective strategy that leverages LLMs to generate pseudo labels for a small sample of the data and uses these labels to create specialized classification models. Our novel method automatically gathers diverse and representative samples to be labeled while minimizing the labeling costs. Our experimental evaluation shows that our classifiers achieve up to 95% F1 score, outperforming LLMs at a lower cost. We present real use cases that demonstrate the effectiveness of our approach in enabling analyses of different aspects of wildlife trafficking.

Extreme heat is the deadliest weather-related hazard in the United States. Furthermore, it is increasing in intensity, frequency, and duration, making skillful forecasts vital to protecting life and property. Traditional numerical weather prediction (NWP) models struggle with extreme heat for medium-range and subseasonal-to-seasonal (S2S) timescales. Meanwhile, artificial intelligence-based weather prediction (AIWP) models are progressing rapidly. However, it is largely unknown how well AIWP models forecast extremes, especially for medium-range and S2S timescales. This study investigates 2-m temperature forecasts for 60 heat waves across the four boreal seasons and over four CONUS regions at lead times up to 20 days, using two AIWP models (Google GraphCast and Pangu-Weather) and one traditional NWP model (NOAA United Forecast System Global Ensemble Forecast System (UFS GEFS)). First, case study analyses show that both AIWP models and the UFS GEFS exhibit consistent cold biases on regional scales in the 5-10 days of lead time before heat wave onset. GraphCast is the more skillful AIWP model, outperforming UFS GEFS and Pangu-Weather in most locations. Next, the two AIWP models are isolated and analyzed across all heat waves and seasons, with events split among the model's testing (2018-2023) and training (1979-2017) periods. There are cold biases before and during the heat waves in both models and all seasons, except Pangu-Weather in winter, which exhibits a mean warm bias before heat wave onset. Overall, results offer encouragement that AIWP models may be useful for medium-range and S2S predictability of extreme heat.