Demographic Bias in Language Models

Hundreds of millions of people use AI for writing assistance. We asked how AI distorts perceptions of AI users and their opinions. Writers (N=2,939) wrote political opinion paragraphs with and without AI assistance. Separate readers (N=11,091) blindly evaluated these paragraphs and made inferences about the writers. With AI, writers seemed more opinionated, competent, positive, and demographically privileged. Writers objected to many of these distortions yet preferred AI-assisted text even when made aware of them. Reward models trained on our data mitigated objectionable distortions but reduced user acceptance. Additional readers (N=8,798) trusted AI-assisted writers more and were more persuaded by AI writing when AI was more distortive. Overall, distortions from AI writing assistance are pervasive, persistent under human oversight, and likely to have consequential effects on human behaviours and attitudes.

Prior work evaluates code generation bias primarily through simple conditional statements, which represent only a narrow slice of real-world programming and reveal solely overt, explicitly encoded bias. We demonstrate that this approach dramatically underestimates bias in practice by examining a more realistic task: generating machine learning (ML) pipelines. Testing both code-specialized and general-instruction large language models, we find that generated pipelines exhibit significant bias during feature selection. Sensitive attributes appear in 87.7% of cases on average, despite models demonstrably excluding irrelevant features (e.g., including "race" while dropping "favorite color" for credit scoring). This bias is substantially more prevalent than that captured by conditional statements, where sensitive attributes appear in only 59.2% of cases. These findings are robust across prompt mitigation strategies, varying numbers of attributes, and different pipeline difficulty levels. Our results challenge simple conditionals as valid proxies for bias evaluation and suggest current benchmarks underestimate bias risk in practical deployments.

As state-of-the-art Large Language Models (LLMs) have become ubiquitous, ensuring equitable performance across diverse demographics is critical. However, it remains unclear whether these disparities arise from the explicitly stated identity itself or from the way identity is signaled. In real-world interactions, users' identity is often conveyed implicitly through a complex combination of various socio-linguistic factors. This study disentangles these signals by employing a factorial design with over 24,000 responses from two open-weight LLMs (Gemma-3-12B and Qwen-3-VL-8B), comparing prompts with explicitly announced user profiles against implicit dialect signals (e.g., AAVE, Singlish) across various sensitive domains. Our results uncover a unique paradox in LLM safety where users achieve better'' performance by sounding like a demographic than by stating they belong to it. Explicit identity prompts activate aggressive safety filters, increasing refusal rates and reducing semantic similarity compared to our reference text for Black users. In contrast, implicit dialect cues trigger a powerful dialect jailbreak,'' reducing refusal probability to near zero while simultaneously achieving a greater level of semantic similarity to the reference texts compared to Standard American English prompts. However, this dialect jailbreak'' introduces a critical safety trade-off regarding content sanitization. We find that current safety alignment techniques are brittle and over-indexed on explicit keywords, creating a bifurcated user experience where standard'' users receive cautious, sanitized information while dialect speakers navigate a less sanitized, more raw, and potentially a more hostile information landscape and highlights a fundamental tension in alignment--between equitable and linguistic diversity--and underscores the need for safety mechanisms that generalize beyond explicit cues.

Large Language Models (LLMs) are increasingly deployed in socially sensitive settings, raising concerns about fairness and biases, particularly across intersectional demographic attributes. In this paper, we systematically evaluate intersectional fairness in six LLMs using ambiguous and disambiguated contexts from two benchmark datasets. We assess LLM behavior using bias scores, subgroup fairness metrics, accuracy, and consistency through multi-run analysis across contexts and negative and non-negative question polarities. Our results show that while modern LLMs generally perform well in ambiguous contexts, this limits the informativeness of fairness metrics due to sparse non-unknown predictions. In disambiguated contexts, LLM accuracy is influenced by stereotype alignment, with models being more accurate when the correct answer reinforces a stereotype than when it contradicts it. This pattern is especially pronounced in race-gender intersections, where directional bias toward stereotypes is stronger. Subgroup fairness metrics further indicate that, despite low observed disparity in some cases, outcome distributions remain uneven across intersectional groups. Across repeated runs, responses also vary in consistency, including stereotype-aligned responses. Overall, our findings show that apparent model competence is partly associated with stereotype-consistent cues, and no evaluated LLM achieves consistently reliable or fair behavior across intersectional settings. These findings highlight the need for evaluation beyond accuracy, emphasizing the importance of combining bias, subgroup fairness, and consistency metrics across intersectional groups, contexts, and repeated runs.

Large language models (LLMs), trained on vast datasets, encode extensive real-world knowledge within their parameters, yet their black-box nature obscures the mechanisms and extent of this encoding. Surrogate modeling, which uses simplified models to approximate complex systems, can offer a path toward better interpretability of black-box models. We propose a surrogate modeling framework that quantitatively explains LLM-encoded knowledge. For a specific hypothesis derived from domain knowledge, this framework approximates the latent LLM knowledge space using observable elements (input-output pairs) through extensive prompting across a comprehensive range of simulated scenarios. Through proof-of-concept experiments in medical predictions, we demonstrate our framework's effectiveness in revealing the extent to which LLMs "perceive" each input variable in relation to the output. Particularly, given concerns that LLMs may perpetuate inaccuracies and societal biases embedded in their training data, our experiments using this framework quantitatively revealed both associations that contradict established medical knowledge and the persistence of scientifically refuted racial assumptions within LLM-encoded knowledge. By disclosing these issues, our framework can act as a red-flag indicator to support the safe and reliable application of these models.

Increasingly, studies are exploring using Large Language Models (LLMs) for accelerated or scaled qualitative analysis of text data. While we can compare LLM accuracy against human labels directly for deductive coding, or labeling text, it is more challenging to judge the ethics and effectiveness of using LLMs in abstractive methods such as inductive thematic analysis. We collaborate with psychologists to study the abstractive claims LLMs make about human life stories, asking, how does using an LLM as an interpreter of meaning affect the conclusions and perspectives of a study? We propose a summarization-based pipeline for surfacing biases in perspective-taking an LLM might employ in interpreting these life stories. We demonstrate that our pipeline can identify both race and gender bias with the potential for representational harm. Finally, we encourage the use of this analysis in future studies involving LLM-based interpretation of study participants' written text or transcribed speech to characterize a positionality portrait for the study.

Research has documented LLMs' name-based bias in hiring and salary recommendations. In this paper, we instead consider a setting where LLMs generate candidate summaries for downstream assessment. In a large-scale controlled study, we analyze nearly one million resume summaries produced by 4 models under systematic race-gender name perturbations, using synthetic resumes and real-world job postings. By decomposing each summary into resume-grounded factual content and evaluative framing, we find that factual content remains largely stable, while evaluative language exhibits subtle name-conditioned variation concentrated in the extremes of the distribution, especially in open-source models. Our hiring simulation demonstrates how evaluative summary transforms directional harm into symmetric instability that might evade conventional fairness audit, highlighting a potential pathway for LLM-to-LLM automation bias.

Large language models are increasingly deployed to simulate patients for clinical training, research, and mental health tools, yet population-level validity remains largely untested. We introduce PsychBench, the first epidemiological audit of LLM patient simulation: 28,800 profiles from four frontier models (GPT-4o-mini, DeepSeek-V3, Gemini-3-Flash, GLM-4.7) evaluated against NHANES and NESARC-III baselines across 120 intersectional cohorts. The central finding is a coherence-fidelity dissociation: models produce clinically plausible individuals while misrepresenting the populations they are drawn from. Variance compression ranges from 14 percent (GLM-4.7) to 62 percent (DeepSeek-V3), eliminating the distributional tails of clinical reality. Despite test-retest correlations above r = 0.90, 36.66 percent of cases cross diagnostic thresholds between runs. Symptom correlation matrices diverge across demographic groups beyond split-half noise, with transgender populations diverging three to five times more than racial differences. Calibration bias is systematic and asymmetric. Models overestimate depression severity for most groups by 3.6 to 6.1 points (Cohen d = 1.13 to 1.91), consistent with training on clinical corpora with elevated base rates. For transgender women the direction inverts: models capture only 8 to 46 percent of documented minority stress elevation, yielding a -5.42 residual (d = -1.55). Models also attribute irritability to Black men and fatigue to women beyond matched controls, encoding racialized and gendered assumptions. Patterns replicate across US and Chinese architectures, indicating failures tied to current training paradigms rather than isolated implementations. For most users, LLM mental health tools risk pathologizing ordinary distress; for transgender users, algorithmic erasure of genuine need. The patients look right. They do not represent real populations.

Safe clinical deployment of Large Language Models (LLMs) requires not only high accuracy but also robust uncertainty calibration to ensure models defer to clinicians when appropriate. Our paper investigates how social descriptors of a patient (specifically sexual orientation and religious affiliation) distort these uncertainty signals and model accuracy. Evaluating nine general-purpose and biomedical LLMs on 2,364 medical questions and their counterfactual variants, we demonstrate that identity markers cause a "calibration crisis". "Homosexual" markers consistently trigger performance drops, and intersectional identities produce idiosyncratic, non-additive harms to calibration. Moreover, a clinician-validated case study in an open-ended generation setting confirms that these failures are not an artifact of the multiple-choice format. Our results demonstrate that the presence of social identity cues does not merely shift predictions; it affects the reliability of confidence signals, posing a significant risk to equitable care and safe deployment in confidence-based clinical workflows.

Fairness in language models is typically studied as a property of a single, centrally optimized model. As large language models become increasingly agentic, we propose that fairness emerges through interaction and exchange. We study this via a controlled hospital triage framework in which two agents negotiate over three structured debate rounds. One agent is aligned to a specific ethical framework via retrieval-augmented generation (RAG), while the other is either unaligned or adversarially prompted to favor demographic groups over clinical need. We find that alignment systematically shapes negotiation strategies and allocation patterns, and that neither agent's allocation is ethically adequate in isolation, yet their joint final allocation can satisfy fairness criteria that neither would have reached alone. Aligned agents partially moderate bias through contestation rather than override, acting as corrective patches that restore access for marginalized groups without fully converting a biased counterpart. We further observe that even explicitly aligned agents exhibit intrinsic biases toward certain frameworks, consistent with known left-leaning tendencies in LLMs. We connect these limits to Arrow's Impossibility Theorem: no aggregation mechanism can simultaneously satisfy all desiderata of collective rationality, and multi-agent deliberation navigates rather than resolves this constraint. Our results reposition fairness as an emergent, procedural property of decentralized agent interaction, and the system rather than the individual agent as the appropriate unit of evaluation.

Large language models exhibit sycophantic tendencies, but whether this behavior varies systematically with perceived user demographics is underexplored. Inspired by intersectionality (overlapping identities produce compounded effects), we probe whether frontier models conditionally exhibit sycophancy. Across 768 multi-turn conversations spanning 128 personas (varying race, age, gender, confidence) and three domains (mathematics, philosophy, conspiracy theories), we find that sycophancy varies sharply with target model and domain, and emerges from combinations of perceived user traits rather than any single dimension. GPT-5-nano scores far higher than Claude Haiku 4.5 (average sycophancy scores of $\bar{x}=2.96$ vs.\ $1.74$, $p < 10^{-32}$); within GPT-5-nano, philosophy elicits 41% more sycophancy than mathematics and Hispanic personas receive the highest scores across races. The worst-scoring persona, a confident, 23-year-old Hispanic woman, averages 5.33/10 (max 6/10), while Claude Haiku 4.5 remains uniformly low with no significant demographic variation. We argue that safety evaluations should incorporate identity-aware adversarial testing.

Large language models are increasingly used as writing tools and pedagogical resources in English for Academic Purposes, but it remains unclear whether they encode culturally differentiated representations when generating academic text. This study tests whether Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating research article introductions conditioned by British and Chinese academic personas. A corpus of 270 texts was generated from 45 prompt templates crossed with six persona conditions in a 2 x 3 design. Logistic regression probes were trained on hidden-state activations across all 35 layers, with shuffled-label baselines, a surface-text skyline classifier, cross-family tests, and sentence-level baselines used as controls. Probe-selected token positions were annotated for structural, lexical, and stance features using the Stanza NLP pipeline. The nationality probe reached 0.968 cross-validated accuracy at Layer 18, with perfect held-out classification. Nationality encoding followed a non-monotonic trajectory across layers, with structural effects strongest in the middle to upper network and lexical-domain effects peaking earlier. At high-signal token positions, British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative or process-oriented vocabulary, while Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural or internationalisation vocabulary. However, sentence-level analysis found no significant nationality differences in the full generated surface text. The findings extend probing methodology to a sociolinguistic attribute and have practical implications for EAP and language pedagogy.

Skin-toned emojis are crucial for fostering personal identity and social inclusion in online communication. As AI models, particularly Large Language Models (LLMs), increasingly mediate interactions on web platforms, the risk that these systems perpetuate societal biases through their representation of such symbols is a significant concern. This paper presents the first large-scale comparative study of bias in skin-toned emoji representations across two distinct model classes. We systematically evaluate dedicated emoji embedding models (emoji2vec, emoji-sw2v) against four modern LLMs (Llama, Gemma, Qwen, and Mistral). Our analysis first reveals a critical performance gap: while LLMs demonstrate robust support for skin tone modifiers, widely-used specialized emoji models exhibit severe deficiencies. More importantly, a multi-faceted investigation into semantic consistency, representational similarity, sentiment polarity, and core biases uncovers systemic disparities. We find evidence of skewed sentiment and inconsistent meanings associated with emojis across different skin tones, highlighting latent biases within these foundational models. Our findings underscore the urgent need for developers and platforms to audit and mitigate these representational harms, ensuring that AI's role on the web promotes genuine equity rather than reinforcing societal biases.

While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.

As Large Language Models (LLMs) are increasingly used to support search and information retrieval, it is critical that they accurately attribute content to its original authors. In this work, we introduce AttriBench, the first fame- and demographically-balanced quote attribution benchmark dataset. By explicitly balancing author fame and demographics, AttriBench enables controlled investigation of demographic bias in quote attribution. Using this dataset, we evaluate 11 widely used LLMs across different prompt settings and find that quote attribution remains a challenging task even for frontier models. We observe large and systematic disparities in attribution accuracy between race, gender, and intersectional groups. We further introduce and investigate suppression, a distinct failure mode in which models omit attribution entirely, even when the model has access to authorship information. We find that suppression is widespread and unevenly distributed across demographic groups, revealing systematic biases not captured by standard accuracy metrics. Our results position quote attribution as a benchmark for representational fairness in LLMs.

As Large Language Models (LLMs) are increasingly integrated into educational settings, understanding their potential biases is critical. This study examines sociodemographic biases in LLM-based educational counselling. We evaluate responses from six LLMs answering questions about 900 vignettes describing students in diverse circumstances. Each vignette is systematically tested across 14 sociodemographic identifiers - spanning race and gender, socioeconomic status, and immigrant background - along with a control condition, yielding 243,000 model responses. Our findings indicate that (1) all models exhibit measurable biases, (2) bias patterns partially align with documented human biases but diverge in notable ways, (3) the magnitude of these biases is strongly influenced by the precision of the student descriptions, where vague or minimal information amplifies disparities nearly threefold, while concrete, individualised metrics substantially reduce them, and (4) bias profiles vary substantially across models. These results demonstrate the importance of context-rich and personalised educational representations, suggesting that AI-driven educational decisions benefit from detailed student-specific information to promote fairness and equity.

As Large Language Models (LLMs) increasingly power decision-making systems across critical domains, understanding and mitigating their biases becomes essential for responsible AI deployment. Although bias assessment frameworks have proliferated for attributes such as race and gender, socioeconomic status bias remains significantly underexplored despite its widespread implications in the real world. We introduce SocioEval, a template-based framework for systematically evaluating socioeconomic bias in foundation models through decision-making tasks. Our hierarchical framework encompasses 8 themes and 18 topics, generating 240 prompts across 6 class-pair combinations. We evaluated 13 frontier LLMs on 3,120 responses using a rigorous three-stage annotation protocol, revealing substantial variation in bias rates (0.42%-33.75%). Our findings demonstrate that bias manifests differently across themes lifestyle judgments show 10$\times$ higher bias than education-related decisions and that deployment safeguards effectively prevent explicit discrimination but show brittleness to domain-specific stereotypes. SocioEval provides a scalable, extensible foundation for auditing class-based bias in language models.

Large Language Models increasingly suppress biased outputs when demographic identity is stated explicitly, yet may still exhibit implicit biases when identity is conveyed indirectly. Existing benchmarks use name based proxies to detect implicit biases, which carry weak associations with many social demographics and cannot extend to dimensions like age or socioeconomic status. We introduce ImplicitBBQ, a QA benchmark that evaluates implicit bias through characteristic based cues, demographically associated attributes that signal implicitly, across age, gender, region, religion, caste, and socioeconomic status. Evaluating 11 models, we find that implicit bias in ambiguous contexts is over six times higher than explicit bias in open weight models. Notably, this bias is distributed unevenly across demographics: caste emerges as the most severe while gender is the least affected. Safety prompting and chain-of-thought reasoning fail to substantially close this gap; even few-shot prompting, which reduces implicit bias by 79%, leaves caste bias at four times the level of any other dimension. These findings indicate that current alignment and prompting strategies address the surface of bias evaluation while leaving demographically associated stereotypic associations largely unresolved. We publicly release our code and dataset for model providers and researchers to benchmark potential mitigation techniques.

Demographic prompts are routed differently from neutral prompts in Mixture-of-Experts (MoE) language models, motivating tests of routing-level stereotype control. We introduce Fairness-Aware Routing Equilibrium (FARE), a diagnostic framework combining demographic routing profiles, empirical layer selection, and fixed inference-time reweighting, and evaluate five MoE architectures in English. At the selected operating points, CrowS-Pairs preference changes by at most 1.3 percentage points; DeepSeekMoE selects no intervention. Paired 95% confidence intervals exclude decreases larger than 2.2 points on each intervened model, and the only nominally significant change (Qwen1.5, p = 0.015) does not survive multiple-comparison correction. OLMoE and Qwen3 nevertheless change nearly every top-k expert set. Noise controls, random and truncated synthetic profiles, and hard masking also move preference by at most 1.5 points. Four generation protocols on OLMoE, Mixtral, Qwen1.5, and Qwen3 show no consistent change in the measured toxicity, lexical, or reference-overlap metrics. Selection and evaluation items overlap, so these comparisons are not independent evaluations. The tested reweighting procedure offers limited stereotype control.

Large Language Models (LLMs) have introduced new capabilities to recommender systems, enabling dynamic, context-aware, and conversational recommendations. However, LLM-based recommender systems inherit and may amplify social biases embedded in their pre-training data, especially when demographic cues are present. Existing fairness solutions either require extra parameters fine-tuning, or suffer from optimization instability. We propose a lightweight and scalable bias mitigation method that combines a kernelized Iterative Null-space Projection (INLP) with a gated Mixture-of-Experts (MoE) adapter. Our approach estimates a closed-form projection that removes single or multiple sensitive attributes from LLM representations with no additional trainable parameters. To preserve task utility, we introduce a two-level MoE adapter that selectively restores useful signals without reintroducing bias. Experiments on two public datasets show that our method reduces attribute leakage across multiple protected variables while maintaining competitive recommendation accuracy.

Large language models (LLMs) are increasingly utilised for social simulation and persona generation, necessitating an understanding of how they represent geopolitical identities. In this paper, we analyse personas generated for Palestinian and Israeli identities by five popular LLMs across 640 experimental conditions, varying context (war vs non-war) and assigned roles. We observe significant distributional patterns in the generated attributes: Palestinian profiles in war contexts are frequently associated with lower socioeconomic status and survival-oriented roles, whereas Israeli profiles predominantly retain middle-class status and specialised professional attributes. When prompted with explicit instructions to avoid harmful assumptions, models exhibit diverse distributional changes, e.g., marked increases in non-binary gender inferences or a convergence toward generic occupational roles (e.g., "student"), while the underlying socioeconomic distinctions often remain. Furthermore, analysis of reasoning traces reveals an interesting dynamics between model reasoning and generation: while rationales consistently mention fairness-related concepts, the final generated personas follow the aforementioned diverse distributional changes. These findings illustrate a picture of how models interpret geopolitical contexts, while suggesting that they process fairness and adjust in varied ways; there is no consistent, direct translation of fairness concepts into representative outcomes.

Large Language Models (LLMs) are deployed in high-stakes settings but can show demographic, gender, and geographic biases that undermine fairness and trust. Prior debiasing methods, including embedding-space projections, prompt-based steering, and causal interventions, often act at a single stage of the pipeline, resulting in incomplete mitigation and brittle utility trade-offs under distribution shifts. We propose CatRAG Debiasing, a dual-pronged framework that integrates functor with Retrieval-Augmented Generation (RAG) guided structural debiasing. The functor component leverages category-theoretic structure to induce a principled, structure-preserving projection that suppresses bias-associated directions in the embedding space while retaining task-relevant semantics. On the Bias Benchmark for Question Answering (BBQ) across three open-source LLMs (Meta Llama-3, OpenAI GPT-OSS, and Google Gemma-3), CatRAG achieves state-of-the-art results, improving accuracy by up to 40% over the corresponding base models and by more than 10% over prior debiasing methods, while reducing bias scores to near zero (from 60% for the base models) across gender, nationality, race, and intersectional subgroups.

Large language models are increasingly used to annotate texts, but their outputs reflect some human perspectives better than others. Existing methods for correcting LLM annotation error assume a single ground truth. However, this assumption fails in subjective tasks where disagreement across demographic groups is meaningful. Here we introduce Perspective-Driven Inference, a method that treats the distribution of annotations across groups as the quantity of interest, and estimates it using a small human annotation budget. We contribute an adaptive sampling strategy that concentrates human annotation effort on groups where LLM proxies are least accurate. We evaluate on politeness and offensiveness rating tasks, showing targeted improvements for harder-to-model demographic groups relative to uniform sampling baselines, while maintaining coverage.

Large language models (LLMs) exhibit pronounced social biases. Output-level or data-optimization--based debiasing methods cannot fully resolve these biases, and many prior works have shown that biases are embedded in internal representations. We propose Unified Graph Isomorphism for Debiasing large language models (UGID), an internal-representation--level debiasing framework for large language models that models the Transformer as a structured computational graph, where attention mechanisms define the routing edges of the graph and hidden states define the graph nodes. Specifically, debiasing is formulated as enforcing invariance of the graph structure across counterfactual inputs, with differences allowed only on sensitive attributes. UGID jointly constrains attention routing and hidden representations in bias-sensitive regions, effectively preventing bias migration across architectural components. To achieve effective behavioral alignment without degrading general capabilities, we introduce a log-space constraint on sensitive logits and a selective anchor-based objective to preserve definitional semantics. Extensive experiments on large language models demonstrate that UGID effectively reduces bias under both in-distribution and out-of-distribution settings, significantly reduces internal structural discrepancies, and preserves model safety and utility.

Large language models (LLMs) are increasingly used for automated text annotation in tasks ranging from academic research to content moderation and hiring. Across 19 LLMs and two experiments totaling more than 4 million annotation judgments, we show that subtle identity cues embedded in text systematically bias annotation outcomes in ways that mirror racial stereotypes. In a names-based experiment spanning 39 annotation tasks, texts containing names associated with Black individuals are rated as more aggressive by 18 of 19 models and more gossipy by 18 of 19. Asian names produce a bamboo-ceiling profile: 17 of 19 models rate individuals as more intelligent, while 18 of 19 rate them as less confident and less sociable. Arab names elicit cognitive elevation alongside interpersonal devaluation, and all four minority groups are consistently rated as less self-disciplined. In a matched dialect experiment, the same sentence is judged significantly less professional (all 19 models, mean gap $-0.774$), less indicative of an educated speaker ($-0.688$), more toxic (18/19), and more angry (19/19) when written in African American Vernacular English rather than Standard American English. A notable exception occurs for name-based hireability, where fine-tuning appears to overcorrect, systematically favoring minority-named applicants. These findings suggest that using LLMs as automated annotators can embed socially patterned biases directly into the datasets and measurements that increasingly underpin research, governance, and decision-making.

Effective personalized feedback is critical to students' literacy development. Though LLM-powered tools now promise to automate such feedback at scale, LLMs are not language-neutral: they privilege standard academic English and reproduce social stereotypes, raising concerns about how "personalization" shapes the feedback students receive. We examine how four widely used LLMs (GPT-4o, GPT-3.5-turbo, Llama-3.3 70B, Llama-3.1 8B) adapt written feedback in response to student attributes. Using 600 eighth-grade persuasive essays from the PERSUADE dataset, we generated feedback under prompt conditions embedding gender, race/ethnicity, learning needs, achievement, and motivation. We analyze lexical shifts across model outputs by adapting the Marked Words framework. Our results reveal systematic, stereotype-aligned shifts in feedback conditioned on presumed student attributes--even when essay content was identical. Feedback for students marked by race, language, or disability often exhibited positive feedback bias and feedback withholding bias--overuse of praise, less substantive critique, and assumptions of limited ability. Across attributes, models tailored not only what content was emphasized but also how writing was judged and how students were addressed. We term these instructional orientations Marked Pedagogies and highlight the need for transparency and accountability in automated feedback tools.

Bias in financial language models constitutes a major obstacle to their adoption in real-world applications. Detecting such bias is challenging, as it requires identifying inputs whose predictions change when varying properties unrelated to the decision, such as demographic attributes. Existing approaches typically rely on exhaustive mutation and pairwise prediction analysis over large corpora, which is effective but computationally expensive-particularly for large language models and can become impractical in continuous retraining and releasing processes. Aiming at reducing this cost, we conduct a large-scale study of bias in five financial language models, examining similarities in their bias tendencies across protected attributes and exploring cross-model-guided bias detection to identify bias-revealing inputs earlier. Our study uses approximately 17k real financial news sentences, mutated to construct over 125k original-mutant pairs. Results show that all models exhibit bias under both atomic (0.58%-6.05%) and intersectional (0.75%-5.97%) settings. Moreover, we observe consistent patterns in bias-revealing inputs across models, enabling substantial reuse and cost reduction in bias detection. For example, up to 73% of FinMA's biased behaviours can be uncovered using only 20% of the input pairs when guided by properties derived from DistilRoBERTa outputs.

Biases in large language models (LLMs) often manifest as systematic distortions in associations between demographic attributes and professional or social roles, reinforcing harmful stereotypes across gender, ethnicity, and geography. This position paper advocates for addressing demographic and gender biases in LLMs through a dual-pronged methodology, integrating category-theoretic transformations and retrieval-augmented generation (RAG). Category theory provides a rigorous, structure-preserving mathematical framework that maps biased semantic domains to unbiased canonical forms via functors, ensuring bias elimination while preserving semantic integrity. Complementing this, RAG dynamically injects diverse, up-to-date external knowledge during inference, directly countering ingrained biases within model parameters. By combining structural debiasing through functor-based mappings and contextual grounding via RAG, we outline a comprehensive framework capable of delivering equitable and fair model outputs. Our synthesis of the current literature validates the efficacy of each approach individually, while addressing potential critiques demonstrates the robustness of this integrated strategy. Ensuring fairness in LLMs, therefore, demands both the mathematical rigor of category-theoretic transformations and the adaptability of retrieval augmentation.

Reliable deployment of Vision-Language Models (VLMs) in radiology requires validation metrics that go beyond surface-level text similarity to ensure clinical fidelity and demographic fairness. This paper investigates a critical blind spot in current model evaluation: the use of decoding strategies that lead to high aggregate token-overlap scores despite succumbing to template collapse, in which models generate only repetitive, safe generic text and omit clinical terminology. Unaddressed, this blind spot can lead to metric gaming, where models that perform well on benchmarks prove clinically uninformative. Instead, we advocate for lexical diversity measures to check model generations for clinical specificity. We introduce Clinical Association Displacement (CAD), a vocabulary-level framework that quantifies shifts in demographic-based word associations in generated reports. Weighted Association Erasure (WAE) aggregates these shifts to measure the clinical signal loss across demographic groups. We show that deterministic decoding produces high levels of semantic erasure, while stochastic sampling generates diverse outputs but risks introducing new bias, motivating a fundamental rethink of how "optimal" reporting is defined.

As artificial intelligence (AI) is increasingly deployed across domains, ensuring fairness has become a core challenge. However, the field faces a "Tower of Babel'' dilemma: fairness metrics abound, yet their underlying philosophical assumptions often conflict, hindering unified paradigms-particularly in unified Multimodal Large Language Models (UMLLMs), where biases propagate systemically across tasks. To address this, we introduce the IRIS Benchmark, to our knowledge the first benchmark designed to synchronously evaluate the fairness of both understanding and generation tasks in UMLLMs. Enabled by our demographic classifier, ARES, and four supporting large-scale datasets, the benchmark is designed to normalize and aggregate arbitrary metrics into a high-dimensional "fairness space'', integrating 60 granular metrics across three dimensions-Ideal Fairness, Real-world Fidelity, and Bias Inertia & Steerability (IRIS). Through this benchmark, our evaluation of leading UMLLMs uncovers systemic phenomena such as the "generation gap'', individual inconsistencies like "personality splits'', and the "counter-stereotype reward'', while offering diagnostics to guide the optimization of their fairness capabilities. With its novel and extensible framework, the IRIS benchmark is capable of integrating evolving fairness metrics, ultimately helping to resolve the "Tower of Babel'' impasse. Project Page: https://iris-benchmark-web.vercel.app/