Negative patient descriptions and stigmatizing language can contribute to generating healthcare disparities in two ways: (1) read by patients, they can harm their trust and engagement with the medical center; (2) read by physicians, they may negatively influence their perspective of a future patient. In psychiatry, the patient-clinician therapeutic alliance is a major determinant of clinical outcomes. Therefore, language usage in psychiatric clinical notes may not only create healthcare disparities, but also perpetuate them. Recent advances in natural language processing systems have facilitated the efforts to detect discriminatory language in healthcare. However, such attempts have only focused on the perspectives of the medical center and its physicians. Considering both physicians’ and non-physicians’ subjective points of view is a more equitable approach to identifying harmful language in clinical notes. By leveraging large language models (LLMs), this work aims to characterize potentially harmful language usage in psychiatric notes by identifying the sentiment expressed in sentences describing patients based on the reader’s point of view. First, we curated a psychiatric lexicon containing words commonly used to describe patients in psychiatry. Sentences (N=39) were extracted from clinical text containing psychiatric lexicon at a medical center, with which a set of physicians (N=10) and non-physicians (N=10) annotated them as negative, neutral, or positive. Three LLMs (GPT-3.5, Llama-3.1, and Mistral) used zero-shot/few-shot in-context learning (ICL) approaches to classify the sentiment of the sentences according to the physician or non-physician point of view. Results showed that GPT-3.5 aligned best to physician point of view and Mistral aligned best to non-physician point of view, both with an ICL approach. These results underline the importance of recognizing subjectivity in clinical annotation tasks, not only for improving the note writing process, but also for the quantification, identification, and reduction of bias in computational systems for downstream analyses.

Speech Continuation (SC) is the task of generating a coherent extension of a spoken prompt while preserving both semantic context and speaker identity. Because SC is constrained to a single audio stream, it offers a more direct setting for probing biases in speech foundation models than dialogue does. In this work we present the first systematic evaluation of bias in SC, investigating how gender and phonation type (breathy, creaky, end-creak) affect continuation behaviour. We evaluate three recent models: SpiritLM (base and expressive), VAE-GSLM, and SpeechGPT across speaker similarity, voice quality preservation, and text-based bias metrics. Results show that while both speaker similarity and coherence remain a challenge, textual evaluations reveal significant model and gender interactions: once coherence is sufficiently high (for VAE-GSLM), gender effects emerge on text-metrics such as agency and sentence polarity. In addition, continuations revert toward modal phonation more strongly for female prompts than for male ones, revealing a systematic voice-quality bias. These findings highlight SC as a controlled probe of socially relevant representational biases in speech foundation models, and suggest that it will become an increasingly informative diagnostic as continuation quality improves.

Honorifics encode social hierarchies and relational nuances, making their correct use a culturally sensitive yet challenging aspect of translation. In doing so, they reflect and shape how individuals position themselves and others within a social world. In this work, we investigate how different translation models handle Korean honorifics, both in implicit scenarios, where only the sentence is given, and explicit scenarios. Our findings are as follows: (i) large language models (LLMs) fine-tuned for translation (MTLMs) consistently prefer polite forms more than their instruction-tuned counterparts in both scenarios; (ii) sequence-to-sequence models produce less polite outputs in implicit contexts but shift toward more polite forms when the addressee is explicitly provided; and (iii) both types of LM-based models tend to become more casual when the addressee is known. When compared with human preferences, MTLMs diverge more strongly, exhibiting a systematic overuse of polite forms relative to human judgments.

Fairness in classification tasks has traditionally focused on bias removal from neural representations, but recent approaches have shifted towards algorithmic methods that embed fairness into the training process. These methods steer models towards fair performance, preventing potential elimination of valuable information that arises from representation manipulation. Reinforcement Learning (RL), with its ability to learn through interaction and adjust reward functions to encourage desired behaviors, presents a promising approach in this domain. In this paper, we conduct an exploratory evaluation of RL for addressing bias in imbalanced classification by scaling the reward function. We employ the contextual multi-armed bandit framework, adapt three popular RL algorithms, and conduct an extensive empirical evaluation of their relative strengths and limitations. Through this analysis, we contribute meaningful evidence to the ongoing debate between algorithmic and representational fairness approaches.

Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we conduct a comprehensive benchmark study on English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task of detecting targeted identities using a demographic-focused taxonomy. We then systematically evaluate models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across identity axes and multi-demographic targeted biases, underscoring the need for more effective and scalable detection frameworks.

AI systems embed majority-group defaults into training data, evaluation metrics, and category definitions, producing documented harms for queer communities including erasure, misclassification, and discrimination. Standard technical audits often rely on aggregate measures and cannot detect harms that be come visible only through the lived experience of affected communities. We conducted a participatory auditing workshop at EurIPS 2025 where 16 queer community members audited four case studies using the 4Cs harm taxonomy (Content, Conduct, Contact, Contract) applied across the AI lifecycle. Participants used structured worksheets and plenary synthesis to classify harms and trace them to their origins in the development pipeline. Across all four cases, participants traced harms to problem definition and data collection, and they identified contractual structures that extract value from vulnerable populations while providing minimal recourse. These findings illustrate that community-informed auditing surfaces concrete, identity-specific harms that aggregate evaluation methods risk overlooking.