Research Papers

The rapid integration of Artificial Intelligence (AI) into K-12 STEM education presents transformative opportunities alongside significant ethical challenges. While AI-powered tools such as Intelligent Tutoring Systems (ITS), automated assessments, and predictive analytics enhance personalized learning and operational efficiency, they also risk perpetuating algorithmic bias, eroding student privacy, and exacerbating educational inequities. This paper examines the dual-edged impact of AI in STEM classrooms, analyzing its benefits (e.g., adaptive learning, real-time feedback) and drawbacks (e.g., surveillance risks, pedagogical limitations) through an ethical lens. We identify critical gaps in current AI education research, particularly the lack of subject-specific frameworks for responsible integration and propose a three-phased implementation roadmap paired with a tiered professional development model for educators. Our framework emphasizes equity-centered design, combining technical AI literacy with ethical reasoning to foster critical engagement among students. Key recommendations include mandatory bias audits, low-resource adaptation strategies, and policy alignment to ensure AI serves as a tool for inclusive, human-centered STEM education. By bridging theory and practice, this work advances a research-backed approach to AI integration that prioritizes pedagogical integrity, equity, and student agency in an increasingly algorithmic world. Keywords: Artificial Intelligence, STEM education, algorithmic bias, ethical AI, K-12 pedagogy, equity in education

Detecting fraudulent auto-insurance claims remains a challenging classification problem, largely due to the extreme imbalance between legitimate and fraudulent cases. Standard learning algorithms tend to overfit to the majority class, resulting in poor detection of economically significant minority events. This paper proposes a structured three-stage training framework that integrates a convex surrogate of focal loss for stable initialization, a controlled non-convex intermediate loss to improve feature discrimination, and the standard focal loss to refine minority-class sensitivity. We derive conditions under which the surrogate retains convexity in the prediction space and show how this facilitates more reliable optimization when combined with deep sequential models. Using a proprietary auto-insurance dataset, the proposed method improves minority-class F1-scores and AUC relative to conventional focal-loss training and resampling baselines. The approach also provides interpretable feature-attribution patterns through SHAP analysis, offering transparency for actuarial and fraud-analytics applications.

OMNIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities

Sahil Verma, Keegan Hines, Jeff Bilmes +4

The emerging capabilities of large language models (LLMs) have sparked concerns about their immediate potential for harmful misuse. The core approach to mitigate these concerns is the detection of harmful queries to the model. Current detection approaches are fallible, and are particularly susceptible to attacks that exploit mismatched generalization of model capabilities (e.g., prompts in low-resource languages or prompts provided in non-text modalities such as image and audio). To tackle this challenge, we propose Omniguard, an approach for detecting harmful prompts across languages and modalities. Our approach (i) identifies internal representations of an LLM/MLLM that are aligned across languages or modalities and then (ii) uses them to build a language-agnostic or modality-agnostic classifier for detecting harmful prompts. Omniguard improves harmful prompt classification accuracy by 11.57\% over the strongest baseline in a multilingual setting, by 20.44\% for image-based prompts, and sets a new SOTA for audio-based prompts. By repurposing embeddings computed during generation, Omniguard is also very efficient ($\approx\!120 \times$ faster than the next fastest baseline). Code and data are available at: https://github.com/vsahil/OmniGuard.

Existing depression screening predominantly relies on standardized questionnaires (e.g., PHQ-9, BDI), which suffer from high misdiagnosis rates (18-34% in clinical studies) due to their static, symptom-counting nature and susceptibility to patient recall bias. This paper presents an AI-powered depression prevention system that leverages large language models (LLMs) to analyze real-time conversational cues--including subtle emotional expressions (e.g., micro-sentiment shifts, self-referential language patterns)--for more accurate and dynamic mental state assessment. Our system achieves three key innovations: (1) Continuous monitoring through natural dialogue, detecting depression-indicative linguistic features (anhedonia markers, hopelessness semantics) with 89% precision (vs. 72% for PHQ-9); (2) Adaptive risk stratification that updates severity levels based on conversational context, reducing false positives by 41% compared to scale-based thresholds; and (3) Personalized intervention strategies tailored to users' emotional granularity, demonstrating 2.3x higher adherence rates than generic advice. Clinical validation with 450 participants shows the system identifies 92% of at-risk cases missed by traditional scales, while its explainable AI interface bridges the gap between automated analysis and clinician judgment. This work establishes conversational AI as a paradigm shift from episodic scale-dependent diagnosis to continuous, emotionally intelligent mental health monitoring.

This study presents the design and development of the 21st Century Teacher Educator for Ghana GPT, a customized Generative AI (GenAI) tool created using OpenAI's Retrieval-Augmented Generation (RAG) and Interactive Semi-Automated Prompting Strategy (ISA). Anchored in a Glocalized design approach, this tool supports pre-service teachers (PSTs) in Ghana by embedding localized linguistic, cultural, and curricular content within globally aligned principles of ethical and responsible AI use. The model utilizes structured, preloaded datasets-including Ghana's National Teacher Education Curriculum Framework (NTECF), UNESCO's (2023) AI guidelines, and culturally responsive pedagogies-to offer curriculum-aligned, linguistically adaptive, and pedagogically grounded learning support. The ISA enables users to input their institution, year, and semester, generating tailored academic content such as lecture notes, assessment practice, practicum resources, and action research guidance. The design incorporates the Culture and Context-Aware Framework, GenAI-CRSciA, and frameworks addressing GenAI neocolonialism to ensure equity, curriculum fidelity, and local relevance. Pilot implementation revealed notable strengths in language adaptation and localization, delivering bilingual support in English and Ghanaian languages like Twi, Dagbani, Mampruli, and Dagaare, with contextualized examples for deeper understanding. The GPT also generated practice assessments aligned with course objectives, reinforcing learner engagement. Challenges included occasional hallucinations due to limited corpora in some indigenous languages and access barriers tied to premium subscriptions. This design case contributes to discourse on Glocalized GenAI and calls for collaboration with OpenAI NextGen to expand access and empirically assess usage across diverse African educational contexts.

Misgendering is the act of referring to someone by a gender that does not match their chosen identity. It marginalizes and undermines a person's sense of self, causing significant harm. English-based approaches have clear-cut approaches to avoiding misgendering, such as the use of the pronoun ``they''. However, other languages pose unique challenges due to both grammatical and cultural constructs. In this work we develop methodologies to assess and mitigate misgendering across 42 languages and dialects using a participatory-design approach to design effective and appropriate guardrails across all languages. We test these guardrails in a standard LLM-based application (meeting transcript summarization), where both the data generation and the annotation steps followed a human-in-the-loop approach. We find that the proposed guardrails are very effective in reducing misgendering rates across all languages in the summaries generated, and without incurring loss of quality. Our human-in-the-loop approach demonstrates a method to feasibly scale inclusive and responsible AI-based solutions across multiple languages and cultures. We release the guardrails and synthetic dataset encompassing 42 languages, along with human and LLM-judge evaluations, to encourage further research on this subject.

The battle for AI leadership is on, with OpenAI in the United States and DeepSeek in China as key contenders. In response to these global trends, the Spanish government has proposed ALIA, a public and transparent AI infrastructure incorporating small language models designed to support Spanish and co-official languages such as Basque. This paper presents the results of Red Teaming sessions, where ten participants applied their expertise and creativity to manually test three of the latest models from these initiatives$\unicode{x2013}$OpenAI o3-mini, DeepSeek R1, and ALIA Salamandra$\unicode{x2013}$focusing on biases and safety concerns. The results, based on 670 conversations, revealed vulnerabilities in all the models under test, with biased or unsafe responses ranging from 29.5% in o3-mini to 50.6% in Salamandra. These findings underscore the persistent challenges in developing reliable and trustworthy AI systems, particularly those intended to support Spanish and Basque languages.

As machine learning models become increasingly prevalent in time series applications, Explainable Artificial Intelligence (XAI) methods are essential for understanding their predictions. Within XAI, feature attribution methods aim to identify which input features contribute the most to a model's prediction, with their evaluation typically relying on perturbation-based metrics. Through systematic empirical analysis across multiple datasets, model architectures, and perturbation strategies, we reveal previously overlooked class-dependent effects in these metrics: they show varying effectiveness across classes, achieving strong results for some while remaining less sensitive to others. In particular, we find that the most effective perturbation strategies often demonstrate the most pronounced class differences. Our analysis suggests that these effects arise from the learned biases of classifiers, indicating that perturbation-based evaluation may reflect specific model behaviors rather than intrinsic attribution quality. We propose an evaluation framework with a class-aware penalty term to help assess and account for these effects in evaluating feature attributions, offering particular value for class-imbalanced datasets. Although our analysis focuses on time series classification, these class-dependent effects likely extend to other structured data domains where perturbation-based evaluation is common.

Ensuring awareness of fairness and privacy in Large Language Models (LLMs) is critical. Interestingly, we discover a counter-intuitive trade-off phenomenon that enhancing an LLM's privacy awareness through Supervised Fine-Tuning (SFT) methods significantly decreases its fairness awareness with thousands of samples. To address this issue, inspired by the information theory, we introduce a training-free method to \textbf{S}uppress the \textbf{P}rivacy and fa\textbf{I}rness coupled \textbf{N}eurons (\textbf{SPIN}), which theoretically and empirically decrease the mutual information between fairness and privacy awareness. Extensive experimental results demonstrate that SPIN eliminates the trade-off phenomenon and significantly improves LLMs' fairness and privacy awareness simultaneously without compromising general capabilities, \eg improving Qwen-2-7B-Instruct's fairness awareness by 12.2\% and privacy awareness by 14.0\%. More crucially, SPIN remains robust and effective with limited annotated data or even when only malicious fine-tuning data is available, whereas SFT methods may fail to perform properly in such scenarios. Furthermore, we show that SPIN could generalize to other potential trade-off dimensions. We hope this study provides valuable insights into concurrently addressing fairness and privacy concerns in LLMs and can be integrated into comprehensive frameworks to develop more ethical and responsible AI systems. Our code is available at https://github.com/ChnQ/SPIN.

Federated Learning (FL) is a technique that allows multiple parties to train a shared model collaboratively without disclosing their private data. It has become increasingly popular due to its distinct privacy advantages. However, FL models can suffer from biases against certain demographic groups (e.g., racial and gender groups) due to the heterogeneity of data and party selection. Researchers have proposed various strategies for characterizing the group fairness of FL algorithms to address this issue. However, the effectiveness of these strategies in the face of deliberate adversarial attacks has not been fully explored. Although existing studies have revealed various threats (e.g., model poisoning attacks) against FL systems caused by malicious participants, their primary aim is to decrease model accuracy, while the potential of leveraging poisonous model updates to exacerbate model unfairness remains unexplored. In this paper, we propose a new type of model poisoning attack, EAB-FL, with a focus on exacerbating group unfairness while maintaining a good level of model utility. Extensive experiments on three datasets demonstrate the effectiveness and efficiency of our attack, even with state-of-the-art fairness optimization algorithms and secure aggregation rules employed.

The Exame Nacional do Ensino Médio (ENEM) is a pivotal test for Brazilian students, required for admission to a significant number of universities in Brazil. The test consists of four objective high-school level tests on Math, Humanities, Natural Sciences and Languages, and one writing essay. Students' answers to the test and to the accompanying socioeconomic status questionnaire are made public every year (albeit anonymized) due to transparency policies from the Brazilian Government. In the context of large language models (LLMs), these data lend themselves nicely to comparing different groups of humans with AI, as we can have access to human and machine answer distributions. We leverage these characteristics of the ENEM dataset and compare GPT-3.5 and 4, and MariTalk, a model trained using Portuguese data, to humans, aiming to ascertain how their answers relate to real societal groups and what that may reveal about the model biases. We divide the human groups by using socioeconomic status (SES), and compare their answer distribution with LLMs for each question and for the essay. We find no significant biases when comparing LLM performance to humans on the multiple-choice Brazilian Portuguese tests, as the distance between model and human answers is mostly determined by the human accuracy. A similar conclusion is found by looking at the generated text as, when analyzing the essays, we observe that human and LLM essays differ in a few key factors, one being the choice of words where model essays were easily separable from human ones. The texts also differ syntactically, with LLM generated essays exhibiting, on average, smaller sentences and less thought units, among other differences. These results suggest that, for Brazilian Portuguese in the ENEM context, LLM outputs represent no group of humans, being significantly different from the answers from Brazilian students across all tests.

Algorithmic bias in healthcare mirrors existing data biases. However, the factors driving unfairness are not always known. Medical devices capture significant amounts of data but are prone to errors; for instance, pulse oximeters overestimate the arterial oxygen saturation of darker-skinned individuals, leading to worse outcomes. The impact of this bias in machine learning (ML) models remains unclear. This study addresses the technical challenges of quantifying the impact of medical device bias in downstream ML. Our experiments compare a "perfect world", without pulse oximetry bias, using SaO2 (blood-gas), to the "actual world", with biased measurements, using SpO2 (pulse oximetry). Under this counterfactual design, two models are trained with identical data, features, and settings, except for the method of measuring oxygen saturation: models using SaO2 are a "control" and models using SpO2 a "treatment". The blood-gas oximetry linked dataset was a suitable test-bed, containing 163,396 nearly-simultaneous SpO2 - SaO2 paired measurements, aligned with a wide array of clinical features and outcomes. We studied three classification tasks: in-hospital mortality, respiratory SOFA score in the next 24 hours, and SOFA score increase by two points. Models using SaO2 instead of SpO2 generally showed better performance. Patients with overestimation of O2 by pulse oximetry of > 3% had significant decreases in mortality prediction recall, from 0.63 to 0.59, P < 0.001. This mirrors clinical processes where biased pulse oximetry readings provide clinicians with false reassurance of patients' oxygen levels. A similar degradation happened in ML models, with pulse oximetry biases leading to more false negatives in predicting adverse outcomes.