Language Model Bias Evaluation

Powerful technologies labeled ``AI''-without sufficient epistemic caution-are already reshaping political and private life, bringing both new dangers and new opportunities for citizen participation. These range from electoral and legislative engagement to the most ambitious form: political co-creation through citizens' assemblies. Large Language Models (LLMs) could support such processes through moderation, translation, facilitation, summarization, and writing assistance. But this potential remains largely unrealized. The Democratic Commons project takes a fundamentally interdisciplinary approach-from philosophy to computer science-to evaluate LLMs against five proposed democratic principles. At its core, the project is driven by the question of political bias: under what conditions can LLMs be used democratically within forms of citizen participation that are them- selves still largely experimental? Addressing these socio-technical questions requires grounding in political theory and, more broadly, in philosophy-disciplines that provide the normative frameworks without which the democratic evaluation of AI systems cannot be mean- ingfully conducted.

Generative AI writing assistants and the Large Language Models (LLMs) that power them are increasingly part of how voters gather information before elections. With growing evidence that they influence users' opinions, it is increasingly important to understand the views and positions of these tools. To better understand these views, we examine the stances supplied by six LLMs on a variety of Swedish-language writing tasks ahead of the 2026 Swedish parliamentary election. We cross 107 policy propositions with 77 writing templates and neutral, positive, and negative prompt framings, producing 24,717 prompts per model and 148,302 responses. To study these, we look at the models' default stance tendencies, compare how they respond to similar issues, and compare their responses with those of each of Sweden's eight parliamentary parties on the same issue. We find that Claude, DeepSeek, Gemini, and Mistral have similar profiles; ChatGPT more often supplies neutral or ambivalent text; and Grok differs most on topics such as migration, crime, and gender. When comparing the political parties, we find that the Social Democrats are closest to all six models. Still, after correcting for multiple comparisons, none of the within-model differences in party distances remains significant. Overall, we find that no model has a clear preference, nor a clear preference for a party, but that this depends on the specific issue or task the user asks about.

LLMs are increasingly used as automated judges for model training and evaluation, yet individual judges exhibit systematic biases that undermine reliability. Much of prior work has studied biases in pairwise LLM-as-a-judge settings; in this paper, we focus on absolute scoring tasks, which mirror more realistic use cases. Across four benchmarks and six models (36 judge-examinee pairs), we show that a model's task accuracy strongly predicts its judging accuracy (Pearson $r \geq 0.90$ on most models) and inversely predicts its directional bias ($r \leq -0.83$), but that accuracy alone does not ensure fair evaluation: more capable examinee models consistently receive more lenient judgments from all judges ($r \geq 0.83$). To address this, we propose calibrated weighted majority voting (WMV), an ensemble evaluation method that aggregates multiple LLM judges weighted by online estimates of their false-positive and false-negative rates. We introduce a disagreement-based estimator that derives these error rates purely from inter-judge agreement patterns, requiring no ground-truth labels or task metadata. In a simulated experiment with shifting task distributions, our label-free WMV tracks an oracle with perfect error-rate knowledge to within 0.5 percentage points on average, outperforming both individual judges and unweighted majority voting. These results demonstrate that principled multi-judge calibration can simultaneously improve accuracy and correct for systematic leniency without requiring labeled data, offering a scalable path to reliable automated evaluation as model capabilities increase.

LLM-as-a-judge protocols are commonly debiased by instructing judges to ignore presentation cues such as citation formatting, source labels, and evidence-display style. We show that this intervention can suppress bias while damaging the resolution of the measurement instrument. We introduce TraceJudgeBench, a diagnostic benchmark for auditing citation-like artifacts in RAG and agent-workflow evaluation, covering content-equivalent pairs, citation ablations, correctness conflicts, human-validated soft and moderate quality gaps, prompt-strength ladders, decoupled judging, and a controlled workflow-ranking probe. Across GPT-5.5, Claude Sonnet 4.6, and DeepSeek V4-Flash, stronger anti-citation prompts reduce worse-cited wins from up to 50.5% to 0%; yet some operating points already convert validated moderate-gap decisions into Tie before the strict stress-test endpoint, while correctness-conflict accuracy remains at or above 93.0%. A second, 50-pair FinQA moderate-gap construction reproduces the qualitative frontier, and open-weight Qwen2.5-14B-Instruct-AWQ and Gemma-3-12B-IT runs reproduce the central HotpotQA frontier. TRACE-style decoupling recovers 96.5-100.0% better-plain resolution across the reported settings. Human validation separates three meanings of Tie: correct equivalence Tie, calibrated soft-boundary Tie, and resolution-destroying Tie on validated quality gaps. We frame debiasing as a measurement intervention whose bias suppression, resolution retention, Tie cost, and protocol cost must be reported jointly. The supplementary artifact contains benchmark splits, prompts, raw judge outputs, validation summaries, and analysis.

Large language models (LLMs) are increasingly deployed as tutors, but it is unclear whether they support all students equally well. We introduce \textbf{EduFair-Bench}, a benchmark for auditing the pedagogical fairness of LLM tutors---whether tutoring quality varies systematically with student demographics. EduFair-Bench pairs a multi-domain question bank (mathematics, physics, chemistry) with a controlled simulation in which a fixed LLM student interacts with each tutor across nine demographic levels spanning four dimensions: gender, immigration background, first language, and socioeconomic status (SES). Tutoring quality is scored on five turn-level pedagogical metrics and four conversation-level dimensions, using an LLM judge validated against three-annotator consensus on 180 tutor turns. Bias is measured via paired Wilcoxon signed-rank tests and bootstrap effect-size confidence intervals. Two ablations (demographic cues conveyed through names; conflicting demographic information between tutor and student) disentangle tutor-driven from student-driven bias. Across five tutors, we find that model capability and demographic fairness are largely orthogonal: the smallest model is the most consistent while the four more capable tutors all exhibit wide demographic gaps with no clear capability-to-fairness ordering, pedagogy-specific RL training redistributes rather than removes bias, and language- and immigration-related cues produce larger gaps than gender- and SES-related cues.

Warning: This submission studies stereotypes and biases, and contains toxic and offensive examples, used for illustration purposes only. Fairness benchmarks such as BBQ have become the de facto standard for fairness evaluation across major model families. We argue that these benchmarks are too easy to support their role: training Qwen 2.5 7B Base with Group Relative Policy Optimization (GRPO) on a single BBQ example, or placing that example in context as a one-shot demonstration for in-context learning (ICL), lifts mean BBQ accuracy from 79.9% to 92.9% and 99.0%, respectively, closing 80% of the gap to its large-scale RLHF counterpart (96.1%) with GRPO, and surpassing it with ICL. These effects generalize across model families. A cross-conditioning analysis shows the improvement is carried by the reasoning traces generated by the model, and one example suffices to elicit a category-agnostic ``missing evidence'' reasoning pattern. We argue that BBQ-style multiple-choice abstention benchmarks measure a single structural cue, and a model that solves them does not thereby become fair. We call for evaluation suites that cover a broader spectrum of fairness alignment.

We post-train Qwen3.8-27B for Korean response style -- verbosity, list and markdown usage, discourse structure and register -- and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text -- observable without any reference labels.

Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.

Large language models (LLMs) are increasingly deployed in globally used assistants, yet their default choices in culturally grounded everyday situations can systematically favour some cultures over others, affecting localisation, user trust, and equitable behaviour. Existing cultural benchmarks evaluate accuracy against a single "correct" answer, making it difficult to characterise an LLM's cultural preference prior when multiple culturally grounded responses are all valid; they also conflate default preferences with context-driven adaptation. We propose DiSCo, a distribution-first forced-choice evaluation framework that isolates default cultural priors and tests steerability via a four-level context gradient (C0--C3). Using DiSCo-Bench (304 items) derived from BLEnD spanning 12 cultures, we evaluate six diverse instruction-tuned LLMs. Default priors are heavily concentrated, with UK and US together absorbing approximately 35% of all selections despite representing only 2 of 12 cultures. Most critically, prompt-based steering consistently widens the selection gap between high- and low-resource cultures, and injecting explicit cultural facts produces negligible distributional disruption, confirming that cultural preference bias cannot be resolved through prompt-based personalisation alone.

Existing bias auditing methods typically rely on model outputs, requiring costly benchmarks or judge models and potentially missing internal shifts that never appear in generated text. We propose a reference-based method that audits bias in hidden-state representations across related model variants, for example before and after fine-tuning. Because fine-tuning reshapes representation geometry, absolute hidden states are not directly comparable, so we encode each sentence by its similarities to a fixed set of anchor sentences, yielding relative representations in a shared comparison space. There we measure how target groups shift in their association with positive and negative attributes, a quantity we call the Representational Bias Shift $ΔB$. Across three model families and the WildGuardMix, DecodingTrust and ToxiGen benchmarks, $ΔB$ correlates with output-level bias change in 15 of the 18 settings we test, reaching $|r| = 0.84$ ($p < 0.001$) under full fine-tuning and becoming more model-dependent under parameter-efficient adaptation. Thresholding $ΔB$ detects checkpoints whose bias increased with ROC AUC between $0.65$ and $0.99$, and on WildGuardMix and DecodingTrust it separates them better than a SEAT-based baseline for all three families. $ΔB$ is also stable under changes to the anchor set, attribute sets and target templates. Our method requires no task-specific evaluation data and audits a model in about three minutes, using $3$-$50\times$ less compute than the output-level benchmarks considered here. We view it as complementary to output-based auditing rather than a replacement for it.

Large language models often repeatedly select the same answer even when many alternatives are plausible. Prior work treats this concentration as bias, but it does not distinguish stable model preferences from responses that depend on a particular prompt wording. We introduce a bias depth score that measures both how strongly a model prefers its top answer under direct prompting and whether that answer survives scenario reframing. Across 4,442 opinion prompts and four large language models, only about a quarter of the concentrated preferences survive reframing. We call these persistent cases Deep biases, and the remaining prompt-dependent cases Shallow biases. Our results show that Deep biases are more often inherited from pretraining and preserved through SFT. Under both continued fine-tuning and prompt-based debiasing for diversity, Deep biases are consistently harder to remove than Shallow biases. Bias depth therefore separates stable learned biases from prompt-wording artifacts that single-prompt metrics conflate. Code, models, and data are available at deepbias.github.io.

Whether a language model looks demographically biased can depend on how the audit asks its question. A charitable-aid benchmark reports that the same models favor minority applicants when rating requests one at a time and penalize some when ranking side by side. We test whether that reversal generalizes to hiring, lending, and medical triage: 40,726 requests to five models, applications differing only in the applicant's name, and a primary test fixed before collection. It does not. None of 36 planned contrasts survives correction. The rating advantage keeps its sign at roughly half the published size, and a precision extension bounds any hiring ranking penalty below the published effect, though the lending and triage ranking floors sit above that margin, so the exclusion is conclusive for hiring ranking and for rating in all three domains only. Planted disparities tracking their injected sizes and a directional replication on the original aid materials bound these nulls. The audit is livelier than the demographics: models recognize transparent audits nearly always, tie every identical-content comparison whether the varying detail is race or a hobby, and reward first-listed candidates as much as any demographic effect we measure. Audit verdicts reflect audit construction more than demographic bias.

Large Language Models are increasingly deployed as information intermediaries, yet measuring their political behavior remains fragile because questionnaire results mix model dispositions with measurement artifacts and response-elicitation biases. We introduce a robust Political Compass Test evaluation framework that samples 300 configurations across an eight-dimensional perturbation space varying language, framing, instructions, answer format, option order, and persona wording. We evaluate eight Gemma 3 and Qwen 3 models across 14 languages and three quantization levels, obtaining design-averaged political coordinates with quantified uncertainty. Most models lean Libertarian-Left on average, but instruction phrasing, language, and answer format significantly affect recovered coordinates. Cross-lingual differences primarily reflect coordinate drift rather than distinct cultural reasoning. Reverse-engineering the test also exposes axis-weighting imbalances and the collapse of degenerate responses toward the center, so near-origin estimates for the smallest models can reflect weak signal rather than centrism. Free-text reasoning and chat-then-classify elicitation alter recovered coordinates, and larger models show clearer persona separation, with a specific failure of the Authoritarian-Left persona to move most models in the intended social direction. In downstream tasks, persona effects are modest relative to model size and target group for hate-speech detection, while base and centrist prompts give the highest agreement for topic-level sentiment. Political role prompting therefore has measurable but task- and dataset-specific downstream effects.

We introduce FramingQA, a benchmark that measures the model sensitivity to question framing across law, medicine, finance, and robotic simulations. Large language models (LLMs) often change their responses to subtle rephrasings that align with an implied stance by users. This can leave users with advice tainted by how they happened to phrase a question rather than by the underlying facts, and the consequences are highly costly in high-stakes domains. Because in the realistic scenarios, both expert practitioners and non-expert users frequently ask LLMs questions containing incomplete or misleading assumptions, models are highly susceptible to those framings. To test this, we inject the framing bias across three nested levels: a framing-biased question phrasing (root), an injected framing-biased premise prepended to a neutral question (propositional), and a premise paired with a framing-biased question (global). Evaluating nine open models (3.8B-70B) across four families, we find that strong per-variant accuracy does not guarantee the robustness across differently phrased questions under the fixed factual information.

While English language models have been widely examined for anti-queer bias, Dutch models remain understudied. To address this gap, we developed a culturally and linguistically adapted Dutch dataset based on the English WinoQueer benchmark, containing pairs of stereotypical and counter-stereotypical sentences. To validate and expand it, we conducted an online survey with 43 Dutch queer participants, confirming 145 of 171 stereotypes as culturally relevant and identifying 22 new biases through free-text responses. The final released dataset, comprising 42,906 sentences, was evaluated using a range of Dutch-specific and multilingual models, including both masked language models (MLMs) and autoregressive language models (ARLMs), with bias measured via a score comparing log-likelihoods of stereotypical versus counter-stereotypical sentences. While the mean bias score across models appeared neutral (~50%), closer analysis revealed significant disparities: some models favored stereotypical sentences up to 97% of the time for transgender identities, but only 6% of the time for gay-related pairs, with transgender and non-binary identities consistently receiving the highest bias scores. Our findings highlight the importance of culturally grounded datasets for evaluating and mitigating biases that disproportionately impact marginalized groups in Dutch language models.

When a question has valid answers under different normative frameworks, a language model must decide which framework to use and whether it can answer correctly within it. We call this setting normative pluralism and study it in Islamic finance using a four-choice taxonomy that separates framework selection from within-framework correctness. This separation reveals the stereotype trap: a cultural cue steers a model toward one framework, but the model selects an incorrect answer within that framework. Across twelve models, two languages, and fifty demographic signals, cultural cues change framework selection and reveal substantial differences in accuracy, especially among non-frontier models. Under the strongest signal, large open-weight models select the Islamic framework 97% of the time. A two-choice evaluation would report near-perfect alignment, although 57--66% of those selections are incorrect. These findings motivate, but do not directly test, the competence-conditioned routing hypothesis: models may favor frameworks where they are more accurate, while cultural cues may expose framework-specific competence gaps.

Large Language Models (LLMs) are increasingly used as conversational search engines for academic literature, yet whether they judge papers on content or on authority signals has not been tested causally. We investigate authority bias: systematic preference for papers based on author prestige, venue, and citations rather than content. Holding title and abstract constant, we vary authority metadata across three counterfactual conditions (original, flipped, boosted) over eight LLMs (five open-weight and three frontier closed-weight) in an in-context, single-turn, top-1 recommendation setting. Our experiments show that authority bias is substantial and directional, varies markedly across models, and is only partially addressable through prompt-level debiasing. We further document a say-do gap: debiasing instructions suppress authority mentions far faster than authority-driven flips, so surface auditing systematically underestimates behavioral bias.

Large language models (LLMs) are increasingly used as judges for subjective tasks, where annotators disagree and the relevant question is not only how accurate a judge is, but whose judgments it reproduces. Sociodemographic prompting conditions the judge on an annotator's demographic profile to align its judgments with the corresponding group's. We test whether this alignment emerges distributionally, comparing the predicted label distributions of 23 open-weight LLMs on three subjective tasks against those of real annotator groups, under three conditions: no demographic information, single-attribute profiles, and intersectional profiles over gender, age, race, and education. Three findings emerge. First, a judge prompted with no demographics is not perspective-neutral: models best reproduce the judgments of White, college-educated annotators. Second, demographic conditioning is asymmetric: it moves the judge toward majority groups and away from minority groups, most strongly on offensiveness, where intersectional profiles amplify the harm. Third, by comparing base and instruct models we identify instruction-tuning as a possible source of the asymmetry. Demographic conditioning should therefore be used with caution to estimate group judgments: conditioning moves predictions away from the reference distributions of the minority groups the method is often invoked to serve.

Public trust in Autonomous Vehicles (AVs) may depend not only on technical success but also on the fairness of their decision making. While a recent trend in AV research involves using general purpose "common sense" models to guide AV decision making, the degree to which these inherit human biases in driving is still understudied. Given that psychology studies have shown human driver biases exist, such as lower pedestrian-yielding rates to Black pedestrians in the US, we argue that analyses of model bias should also be part of AV evaluation. Concretely, in this paper we propose two new bias testing methodologies for Large Language Models (LLMs) and Visual-Language Models (VLMs)-"All Else Being Equal" tests and "Self-Consistency" tests-in order to assess bias in pedestrian-yielding decisions. Our findings show that both LLMs and VLMs make yielding decisions which are influenced by pedestrian gender, ethnicity, religion, disability, age, skin tone and socio-economic status. While the type and degree of bias is different from model to model, we highlight common patterns-and raise questions about the "common sense" model paradigm, particularly the need to either revise the paradigm or address issues of downstream bias.

Efficient evaluation changes the protocol used to support claims about model behavior, yet it is rarely tested whether those claims remain stable after the evaluation itself is made cheaper. We stress-test conclusion robustness in responsible-AI benchmarking by evaluating three dense and mixture-of-experts models on BBQ and BBQ-V under seven conditions spanning batching, quantization, benchmark reduction, and their combinations. Rather than treating preserved aggregate accuracy as sufficient, we compare accuracy, bias severity and prevalence, reasoning quality, subgroup behavior, subset-membership stability, runtime, and measured GPU energy against a full-benchmark BF16 baseline. Larger batching keeps accuracy within 0.35 percentage points of baseline and produces comparatively small subgroup changes, while reducing energy in five of six model--dataset settings. INT8 largely preserves quality but uses 1.79--4.26$\times$ baseline energy. INT4 causes larger, model- and context-dependent changes. Reduced benchmarks provide the most consistent savings, but very small subsets are substantially more sensitive to which items are retained. Efficient evaluation should therefore be treated as a measurement intervention whose validity must be checked across the conclusions the benchmark is intended to support. Our project website is https://vectorinstitute.github.io/sustainable-rai-evaluation/ and the code is available at https://github.com/VectorInstitute/sustainable-rai-evaluation.

While gender and racial biases in language models have been widely studied, anti-LGBTQ biases remain underexplored, particularly beyond English. Existing benchmarks often do not capture cultural and linguistic variation and rely on gender representations. This paper introduces a multilingual German-English benchmark dataset for the evaluation of anti-LGBTQ biases in language models. It combines community-sourced stereotypes from German-speaking queer individuals with a German translation of WinoQueer. The data is used to evaluate eight language models across sizes and architectures and explore mitigation through fine-tuning on community and progressive media content. Results show that language models reproduce anti-queer stereotypes, with variation across identities and models. Differences between the translated and community-based data highlight the importance of cultural adaptation for multilingual bias evaluation. Fine-tuning reduces bias on average, but not consistently across models and identities. Warning: This text contains examples of anti-queer hateful language and stereotypes.

Language models are increasingly mediating information access to end users, urging a systematic evaluation of their responses for a fair and reliable information ecosystem. Existing evaluations, however, are often topic-specific or synthetic, limiting their ability to capture the complexity of "in the wild" information-seeking queries and the risks present in model responses. To address this gap, we introduce WildSEEK, a manually annotated dataset of 3k information-seeking queries from real user interactions, and an evaluation framework for LLM-generated responses. WildSEEK includes annotations for risk-sensitive domains (e.g. health and financial information), and distinguishes factoid queries from analytical queries which seek responses beyond facts. We train classifiers on WildSEEK to analyze more than 1.8M realistic user queries. We find that over a third of information-seeking queries are high-risk and more often analytical. Our findings show that LLM responses fail more often in four criteria: sycophantic behavior, overreliance, a default US-centric perspective, and poor handling of vulnerable populations -- with failure rates being mostly higher for analytical queries. By providing methods to monitor the reliability, safety, and fairness of LLM behavior, our dataset and evaluation framework offer an empirical foundation for the broader question of how these systems should behave as they take on a growing role in information access.

Multi-Agent Debate (MAD) has been widely adopted to improve LLM-based evaluation by prompting multiple agents to negotiate and reach a consensus. However, for subjective rubric-based scoring, inter-agent agreement does not guarantee alignment with human judgments. In this paper, we compare a single-judge baseline against a consensus-based MAD protocol on subjective evaluation tasks and design three ablations to isolate the impact of role prompting, multi-round interaction, and explicit score sharing. Evaluations across six LLMs show that the single-judge baseline achieves the strongest human alignment on average across six judge models, whereas MAD shows degradation in human alignment on both tasks. Our ablations demonstrate that this performance drop stems primarily from asymmetric role prompting rather than the interaction itself. Specifically, assigning a strict judge role introduces a systematic downward bias that the consensus process fails to correct. The central finding is that this bias reflects strict-stance dominance beyond averaging: the consensus score falls well beyond the arithmetic midpoint of the standalone strict and lenient conditions, rather than averaging them out. Removing role asymmetry (Symmetric MAD) largely recovers baseline performance, while masking peer scores widens inter-agent disagreement on average and worsens average human alignment. These findings demonstrate that multi-agent consensus can enforce artificial agreement at the expense of true human alignment, revealing a structural limitation in consensus-style, role-specialized MAD protocols for subjective scoring.

Bias in human-agent interaction can manifest not only through hostile language but also as benevolent bias, whereby unequal treatment hides behind a warm, positive tone. To make it detectable, we operationalise benevolent bias along two dimensions, tone and treatment, yielding three classes: neutral support, overt bias, and benevolent bias. Building on these definitions, we construct BENEVDIAL, a class-balanced corpus of 362,880 multi-turn support dialogues spanning user and agent demographics, roles, and generators, to support controlled evaluation. We then test two detector families on it: off-the-shelf safety detectors and prompted large language model (LLM) judges. Our findings reveal a notable detection gap: off-the-shelf detectors reliably flag overt bias yet largely fail to identify benevolent bias. LLM judges improve sensitivity when guided by explicit detection criteria, but this comes at the cost of increased misclassification of neutral supportive statements as benevolent bias, a tendency that is further exacerbated by the presence of demographic context. These findings suggest that fair monitoring of human-agent dialogue must look beyond surface cues to whether the agent's treatment is disparate.

Professional communication is increasingly mediated by LLMs - but do these models serve all users equally? We show that when prompts contain linguistic features more commonly used by women (hedges, tag questions, collective reference), they systematically elicit shorter, less sophisticated, and less formal responses across three document types and four models. These effects persist after controlling for prompt complexity and feature carry-over. Explicit gender cues like sign-off names are encoded in the same representational space as linguistic dialect - suggesting shared underlying mechanisms - yet linguistic register is far more influential, producing large, consistent effects where names produce none. Our results further reveal that post-hoc mitigation is challenging: because these patterns are culturally embedded and outside conscious control, users cannot easily avoid them through strategic self-presentation, and mechanistic analysis reveals that linguistic features are encoded in early transformer layers and entangled with other features. Our work calls for upstream consideration of the influences of linguistic variation to mitigate disparate impacts of LLM-mediated workplace communication.

Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with an LLM matcher. However, eliminating positional influence entirely still does not reliably yield accuracy gains, while cyclic permutation often improves them. For two-stage prompting, an aggregate measure of recall imbalance and a direct per-question measure of order sensitivity both fail to show reliable debiasing.

As users increasingly turn to Large Language Models (LLMs) for information and advice on political matters, particularly during election periods, the political preferences expressed by these systems have become a matter of public interest. Prior research has shown that interactions with LLMs can influence users' political attitudes and choices, raising questions about how these models themselves evaluate political actors. In this paper, we investigate whether and how LLMs express preferences toward political parties and political leaders. We introduce a systematic and reproducible auditing framework in which multiple LLMs are prompted to evaluate parties and leaders across nine criteria. Rather than attempting to infer the models' "true" political beliefs, we focus on their observable behavior, examining consistency across evaluations, differences between models, refusal rates, and sensitivity to prompt formulation. We further investigate how these evaluations vary when models are instructed to adopt different personas. We demonstrate the framework through an Italian case study, providing a systematic analysis of LLM-generated political evaluations on italian parties and leaders.

Group alignment adapts a language model to a demographic group to produce responses that reflect the group's opinions, values, and preferences. Sycophancy, a well-documented by-product of alignment, causes the model to over-agree with the user regardless of factual and objective information. However, existing group alignment methods and evaluations focus only on how closely the model matches the group's opinions, overlooking the induced change in sycophantic behaviour. To bridge this gap, we introduce \textbf{G}roup \textbf{A}lignment-induced \textbf{S}ycophancy (GAS) and systematically evaluate alignment across 3 methods, 4 models and 13 demographic groups, on both the intended gain in opinion alignment and the unintended shift in sycophancy. We find that gain and shift are non-uniform across groups: under an identical budget, some groups receive larger gains in opinion alignment than others, and the induced sycophancy shift forms a group-specific profile rather than a single-dimensional change. These results suggest that group alignment should be reported as a two-sided, multi-dimensional profile rather than a single fit score that accounts for per-group differences when adapting LLMs to diverse populations.

Political stance detection in LLMs has long been dominated by closed-ended, multiple-choice political survey questions---originally designed for humans, and thus lacks the realism and nuance of human-AI interactions in the wild, while also being susceptible to sandbagging. The recent IssueBench framework substantially mitigates these limitations with templated prompts anchored in real-world chat logs. Given the rise in non-work-related use of GenAI assistants, we extend IssueBench beyond writing assistance to include two additional tasks, information seeking and opinion sharing. We argue that templated prompts still lack the nuance of real ones, especially for open-ended tasks, and remain recognisable as evaluation artefacts. We propose the use of fully synthetic (LLM-generated) prompts, produced under detailed instructions with real prompts as seeds. We assess the ecological validity of real, templated, and LLM-generated prompts in a small-scale study covering 3 highly contested policy issues and 3 recent geopolitical conflicts. Human and LLM annotators rank LLM-generated prompts as no less realistic than real ones and clearly more realistic than templated ones, and find that they carry their intended intent and stance more clearly; the LLMs separate templated prompts from the other two far more sharply than the humans do. In a case study, templated and LLM-generated prompts yield systematically different stance estimates for the same model, most visibly under neutral framings, where templated prompts overstate the model's leaning in the direction encoded by the topic-and-stance text (filler) slotted into their templates.

As Large Language Models (LLMs) are increasingly deployed as autonomous agents, accurately evaluating their latent values and biases is critical. The NLP community typically evaluates models using large, unstructured benchmarks. While effective for general capabilities, these datasets fundamentally conflate causal mechanisms: even when an aggregate bias is detected, unstructured evaluations cannot disentangle whether it stems from baseline traits, contextual confounders, or complex interactions. To address this, we introduce an analytically exact framework for the controlled behavioral evaluation of LLMs. We bridge human psychometrics with LLM mechanics by resolving gaps in design, measurement, and analysis. First, we replace unstructured prompting with fully crossed factorial experiments to systematically isolate causal main and interaction effects. Second, we eliminate Monte Carlo text sampling noise by operating directly on exact, token-level Probability Mass Functions (PMFs). Third, we derive a multivariate ordinal consensus metric and a distributional ANOVA to process these PMFs analytically. We validate our framework with a case study on consumer ethnocentrism across five LLMs, demonstrating how our approach isolates systemic country-of-origin biases that aggregate benchmarks otherwise obscure.