Research Papers

In this research, we introduce BEATS, a novel framework for evaluating Bias, Ethics, Fairness, and Factuality in Large Language Models (LLMs). Building upon the BEATS framework, we present a bias benchmark for LLMs that measure performance across 29 distinct metrics. These metrics span a broad range of characteristics, including demographic, cognitive, and social biases, as well as measures of ethical reasoning, group fairness, and factuality related misinformation risk. These metrics enable a quantitative assessment of the extent to which LLM generated responses may perpetuate societal prejudices that reinforce or expand systemic inequities. To achieve a high score on this benchmark a LLM must show very equitable behavior in their responses, making it a rigorous standard for responsible AI evaluation. Empirical results based on data from our experiment show that, 37.65\% of outputs generated by industry leading models contained some form of bias, highlighting a substantial risk of using these models in critical decision making systems. BEATS framework and benchmark offer a scalable and statistically rigorous methodology to benchmark LLMs, diagnose factors driving biases, and develop mitigation strategies. With the BEATS framework, our goal is to help the development of more socially responsible and ethically aligned AI models.

The rise of social media has significantly increased the prevalence of cyberbullying (CB), posing serious risks to both mental and physical well-being. Effective detection systems are essential for mitigating its impact. While several machine learning (ML) models have been developed, few incorporate victims' psychological, demographic, and behavioral factors alongside bullying comments to assess severity. In this study, we propose an AI model intregrating user-specific attributes, including psychological factors (self-esteem, anxiety, depression), online behavior (internet usage, disciplinary history), and demographic attributes (race, gender, ethnicity), along with social media comments. Additionally, we introduce a re-labeling technique that categorizes social media comments into three severity levels: Not Bullying, Mild Bullying, and Severe Bullying, considering user-specific factors.Our LSTM model is trained using 146 features, incorporating emotional, topical, and word2vec representations of social media comments as well as user-level attributes and it outperforms existing baseline models, achieving the highest accuracy of 98\% and an F1-score of 0.97. To identify key factors influencing the severity of cyberbullying, we employ explainable AI techniques (SHAP and LIME) to interpret the model's decision-making process. Our findings reveal that, beyond hate comments, victims belonging to specific racial and gender groups are more frequently targeted and exhibit higher incidences of depression, disciplinary issues, and low self-esteem. Additionally, individuals with a prior history of bullying are at a greater risk of becoming victims of cyberbullying.

Explainable Artificial Intelligence (XAI) methods are increasingly used in safety-critical domains, yet there is no unified framework to jointly evaluate fidelity, interpretability, robustness, fairness, and completeness. We address this gap through two contributions. First, we propose a multi-criteria evaluation framework that formalizes these five criteria using principled metrics: fidelity via prediction-gap analysis; interpretability via a composite concentration-coherence-contrast score; robustness via cosine-similarity perturbation stability; fairness via Jensen-Shannon divergence across demographic groups; and completeness via feature-ablation coverage. These are integrated using an entropy-weighted dynamic scoring scheme that adapts to domain-specific priorities. Second, we introduce Perturbation-Gradient Consensus Attribution (PGCA), which fuses grid-based perturbation importance with Grad-CAM++ through consensus amplification and adaptive contrast enhancement, combining perturbation fidelity with gradient-based spatial precision. We evaluate across five domains (brain tumor MRI, plant disease, security screening, gender, and sunglass detection) using fine-tuned ResNet-50 models. PGCA achieves the best performance in fidelity $(2.22 \pm 1.62)$, interpretability $(3.89 \pm 0.33)$, and fairness $(4.95 \pm 0.03)$, with statistically significant improvements over baselines $(p < 10^{-7})$. Sensitivity analysis shows stable rankings (Kendall's $(τ\geq 0.88)$). Code and results are publicly available.

As Large Language Models (LLMs) become more prevalent, concerns about their safety, ethics, and potential biases have risen. Systematically evaluating LLMs' risk decision-making tendencies and attitudes, particularly in the ethical domain, has become crucial. This study innovatively applies the Domain-Specific Risk-Taking (DOSPERT) scale from cognitive science to LLMs and proposes a novel Ethical Decision-Making Risk Attitude Scale (EDRAS) to assess LLMs' ethical risk attitudes in depth. We further propose a novel approach integrating risk scales and role-playing to quantitatively evaluate systematic biases in LLMs. Through systematic evaluation and analysis of multiple mainstream LLMs, we assessed the "risk personalities" of LLMs across multiple domains, with a particular focus on the ethical domain, and revealed and quantified LLMs' systematic biases towards different groups. This research helps understand LLMs' risk decision-making and ensure their safe and reliable application. Our approach provides a tool for identifying and mitigating biases, contributing to fairer and more trustworthy AI systems. The code and data are available.

Explainable AI~(XAI) methods such as SHAP can help discover feature attributions in black-box models. If the method reveals a significant attribution from a ``protected feature'' (e.g., gender, race) on the model output, the model is considered unfair. However, adversarial attacks can subvert the detection of XAI methods. Previous approaches to constructing such an adversarial model require access to underlying data distribution, which may not be possible in many practical scenarios. We relax this constraint and propose a novel family of attacks, called shuffling attacks, that are data-agnostic. The proposed attack strategies can adapt any trained machine learning model to fool Shapley value-based explanations. We prove that Shapley values cannot detect shuffling attacks. However, algorithms that estimate Shapley values, such as linear SHAP and SHAP, can detect these attacks with varying degrees of effectiveness. We demonstrate the efficacy of the attack strategies by comparing the performance of linear SHAP and SHAP using real-world datasets.

Large pre-trained language models have become a crucial backbone for many downstream tasks in natural language processing (NLP), and while they are trained on a plethora of data containing a variety of biases, such as gender biases, it has been shown that they can also inherit such biases in their weights, potentially affecting their prediction behavior. However, it is unclear to what extent these biases also affect feature attributions generated by applying "explainable artificial intelligence" (XAI) techniques, possibly in unfavorable ways. To systematically study this question, we create a gender-controlled text dataset, GECO, in which the alteration of grammatical gender forms induces class-specific words and provides ground truth feature attributions for gender classification tasks. This enables an objective evaluation of the correctness of XAI methods. We apply this dataset to the pre-trained BERT model, which we fine-tune to different degrees, to quantitatively measure how pre-training induces undesirable bias in feature attributions and to what extent fine-tuning can mitigate such explanation bias. To this extent, we provide GECOBench, a rigorous quantitative evaluation framework for benchmarking popular XAI methods. We show a clear dependency between explanation performance and the number of fine-tuned layers, where XAI methods are observed to benefit particularly from fine-tuning or complete retraining of embedding layers.

Design Opportunities for Explainable AI Paraphrasing Tools: A User Study with Non-native English Speakers

Yewon Kim, Thanh-Long V. Le, Donghwi Kim +2

We investigate how non-native English speakers (NNESs) interact with diverse information aids to assess and select AI-generated paraphrases. We develop ParaScope, an AI paraphrasing assistant that integrates diverse information aids, such as back-translation, explanations, and usage examples, and logs user interaction data. Our in-lab study with 22 NNESs reveals that user preferences for information aids vary by language proficiency, with workflows progressing from global to more detailed information. While back-translation was the most frequently used aid, it was not a decisive factor in suggestion acceptance; users combined multiple information aids to make informed decisions. Our findings demonstrate the potential of explainable AI paraphrasing tools to enhance NNESs' confidence, autonomy, and writing efficiency, while also emphasizing the importance of thoughtful design to prevent information overload. Based on these findings, we offer design implications for explainable AI paraphrasing tools that support NNESs in making informed decisions when using AI writing systems.

SoUnD Framework: Analyzing (So)cial Representation in (Un)structured (D)ata

Mark Díaz, Sunipa Dev, Emily Reif +2

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.

Explaining Search Result Stances to Opinionated People

Z. Wu, T. Draws, F. Cau +3

People use web search engines to find information before forming opinions, which can lead to practical decisions with different levels of impact. The cognitive effort of search can leave opinionated users vulnerable to cognitive biases, e.g., the confirmation bias. In this paper, we investigate whether stance labels and their explanations can help users consume more diverse search results. We automatically classify and label search results on three topics (i.e., intellectual property rights, school uniforms, and atheism) as against, neutral, and in favor, and generate explanations for these labels. In a user study (N =203), we then investigate whether search result stance bias (balanced vs biased) and the level of explanation (plain text, label only, label and explanation) influence the diversity of search results clicked. We find that stance labels and explanations lead to a more diverse search result consumption. However, we do not find evidence for systematic opinion change among users in this context. We believe these results can help designers of search engines to make more informed design decisions.

Images are loaded with semantic information that pertains to real-world ontologies: dog breeds share mammalian similarities, food pictures are often depicted in domestic environments, and so on. However, when training machine learning models for image classification, the relative similarities amongst object classes are commonly paired with one-hot-encoded labels. According to this logic, if an image is labelled as 'spoon', then 'tea-spoon' and 'shark' are equally wrong in terms of training loss. To overcome this limitation, we explore the integration of additional goals that reflect ontological and semantic knowledge, improving model interpretability and trustworthiness. We suggest a generic approach that allows to derive an additional loss term starting from any kind of semantic information about the classification label. First, we show how to apply our approach to ontologies and word embeddings, and discuss how the resulting information can drive a supervised learning process. Second, we use our semantically enriched loss to train image classifiers, and analyse the trade-offs between accuracy, mistake severity, and learned internal representations. Finally, we discuss how this approach can be further exploited in terms of explainability and adversarial robustness. Code repository: https://github.com/S1M0N38/semantic-encodings

The application of Algorithmic Recourse in decision-making is a promising field that offers practical solutions to reverse unfavorable decisions. However, the inability of these methods to consider potential dependencies among variables poses a significant challenge due to the assumption of feature independence. Recent advancements have incorporated knowledge of causal dependencies, thereby enhancing the quality of the recommended recourse actions. Despite these improvements, the inability to incorporate the temporal dimension remains a significant limitation of these approaches. This is particularly problematic as identifying and addressing the root causes of undesired outcomes requires understanding time-dependent relationships between variables. In this work, we motivate the need to integrate the temporal dimension into causal algorithmic recourse methods to enhance recommendations' plausibility and reliability. The experimental evaluation highlights the significance of the role of time in this field.

Approaches to fair and ethical AI have recently fell under the scrutiny of the emerging, chiefly qualitative, field of critical data studies, placing emphasis on the lack of sensitivity to context and complex social phenomena of such interventions. We employ some of these lessons to introduce a tripartite decision-making toolkit, informed by dilemmas encountered in the pursuit of responsible AI/ML. These are: (a) the opportunity dilemma between the availability of data shaping problem statements vs problem statements shaping data; (b) the trade-off between scalability and contextualizability (too much data versus too specific data); and (c) the epistemic positioning between the pragmatic technical objectivism and the reflexive relativism in acknowledging the social. This paper advocates for a situated reasoning and creative engagement with the dilemmas surrounding responsible algorithmic/data-driven systems, and going beyond the formulaic bias elimination and ethics operationalization narratives found in the fair-AI literature.