Social Bias in Language Models

A common approach to measuring bias in Large Language Models is to compare the log-likelihoods of two contrastive stereotype sentences. We argue that such single-pair comparisons are often unreliable: simply rewriting the same stereotype with an alternative attribute can yield logically inconsistent preferences. To address this, we propose a dual minimal pair setup that introduces two axes of comparison for robust stereotype evaluation. First, we present a data-augmentation framework that fills critical gaps in existing stereotype datasets by generating paraphrases and alternate attributes. We apply our framework on a set of English, Russian, Spanish and Chinese stereotypes. Second, we introduce two evaluation metrics tailored to the dual minimal pair setup. One of these metrics provides a new perspective on bias by modeling the mutual information (MI) between social groups and stereotyped attributes. This MI-based metric is better suited for aggregation and enables more robust comparisons of stereotype strength across different languages and models. Our code is available at https://github.com/stepanat/missing-minimal-pair/.

Activation steering has gained popularity as a lightweight inference-time debiasing technique for large language models. However, prior work reports that steering vectors generalise poorly, with unintended effects on model performance and limited transfer to new datasets. Our work analyses what the debiasing direction used for activation steering actually encodes, in order to shed light on its inconsistent performance. We study the linear debiasing direction obtained by contrasting the activations of anti-biased and biased prompts, and evaluate it as a steering intervention across bias and general knowledge benchmarks. We find that this direction is dominated by model confidence, pointing from regions of high to low-probability tokens in activation space rather than encoding a meaningful representation of model bias. Steering along it does reduce measured bias, but this is a consequence of reducing model confidence: on QA benchmarks we find that this steering drives the model to abstain from answering, with a side effect of improving fairness metrics. Our experiments show that model confidence is the dominant separating factor between biased and anti-biased prompts in hidden space, indicating that isolating a linear representation of bias which is disentangled from model confidence is difficult and steering-based debiasing results should be interpreted with care. In short, steering appears to reduce bias, not by correcting the model's underlying preferences, but by making it less confident, even on tasks unrelated to bias.

Generative large language models (LLMs) inherit undesirable behaviors from pre-training, including demographic bias and toxic generation, that often emerge only after deployment and affect a small subset of inputs. A repair should eliminate the identified defect, preserve the model's overall functionality and, ideally, provide correctness guarantees. Existing approaches address this only partially: gradient-based fine-tuning lacks per-instance guarantees and becomes unstable with few defect samples; model editing assumes explicit knowledge replacement rather than behavioral correction; and constraint-based repair is largely restricted to discriminative models with unique target outputs. We present FORGE, a framework for targeted behavioral repair of generative language models that separates defect localization, weight editing, and behavioral verification into independent stages. Its core is a repair abstraction that converts localized defective generation into explicit optimization objectives, enabling verification-oriented repair techniques to operate on autoregressive generation. FORGE is editing-mechanism agnostic: we instantiate it with (1) a constraint-based quadratic optimization method that provides per-sample repair certificates and (2) a null-space projection editor that minimizes interference with the original model distribution, both under the same localization and verification protocol. On five open-source LLMs, FORGE consistently achieves larger reductions in bias and toxicity than gradient-based fine-tuning with minor perplexity degradation. The two backends exhibit complementary performance across architectures, which a lightweight causal probe traces to where toxicity-related signals concentrate. FORGE also remains effective with only a handful of defective examples, where conventional fine-tuning often oscillates or fails to converge.

Large language models (LLMs) can reproduce social stereotypes from their training data, motivating extensive research on model debiasing. However, existing methods often rely on explicit biased examples or predefined group-term substitutions, making them sensitive to wording and less effective at capturing stereotype concepts shared across diverse contexts. More importantly, they typically suppress biased outputs without explicitly modeling the statistical dependence between model outputs and the underlying stereotype concepts. We propose Acmite, a lightweight concept-guided framework for targeted and selective debiasing. Acmite represents stereotypes as structured semantic concepts and uses maximal marginal relevance (MMR) to select diverse concepts for debiasing. Inspired by mutual information minimization, it approximates this dependence with token-level KL divergence while preserving task semantics. A lightweight LoRA adapter is trained with the base model frozen and activated at inference time only when the input is sufficiently similar to stereotype-related concepts; otherwise, the original model is used directly. We evaluate Acmite on BBQ, CrowS-Pairs, and StereoSet, and assess general capability preservation on ARC-Challenge, GSM8K, and PIQA. Experiments across three LLMs show that Acmite effectively mitigates gender bias across complementary evaluation formats while maintaining competitive performance on bias-unrelated tasks. Anonymous code and data are available at https://anonymous.4open.science/r/Acmite-18E2/.

Large language models (LLMs) are increasingly deployed in multi-agent settings, where agents observe and influence one another, making social influence a key dimension of AI behavior and safety. We investigate whether LLMs' responses depend on the social identity of other agents, beyond the effect of their consensus. We construct judgment tasks with a single correct answer, and place models in a multi-agent setting where they receive incorrect answers from other agents whose social identities (AI or human, model family, or an arbitrary minimal group) are either shared with or distinct from their own. Across 12 open-weights models and nine tasks, we find a bidirectional effect of group identity on conformity to incorrect answers: in-group consensus increases conformity (in-group favoritism), whereas out-group consensus decreases it (out-group divergence). Unlike humans, for whom one ally breaking the consensus sharply reduces conformity, models are unmoved by an ally from the majority's group. Worse, a correct ally from the opposing group intensifies this bidirectional effect. Chain-of-Thought reasoning suppresses most of these effects, yet an in-group ally still reduces conformity to an incorrect out-group majority. Labeling peers as safety-aligned shifts overall conformity but leaves in-group favoritism and out-group divergence intact. These results show that group identity shapes how LLMs aggregate information across agents, independently of its correctness, and identify a manipulation surface for multi-agent AI systems.

Previous work on Egyptian Arabic in NLP has focused largely on the prestigious Cairene Egyptian Arabic (CEA) dialect, resulting in a lack of representation for the less prestigious Sa'idi Egyptian Arabic (SEA) dialect both in LLM and resource development. Does this lack of representation influence an LLM's view of the acceptability of SEA (upstream), and does an upstream bias against SEA lead to worse performance (downstream)? We investigate the upstream effect of SEA dialectal features on LLM preferences in a Targeted Syntactic Evaluation (TSE) task which reveals a significant bias against SEA across multiple LLMs. We then analyze the effect of these same features on downstream model performance on MMLU benchmarks and show that models experience a degradation in performance when presented with SEA. This work highlights the need for further exploration on how sub-dialectal variation impacts language technologies.

Warning: This paper contains examples of stereotypes and social bias. LLMs are increasingly used in interactive settings by the general public, making the evaluation of model behavior in multi-turn conversational scenarios important for safety, including stereotyping-related harms. However, existing multi-turn social bias evaluations often rely on pre-specified or template-based user inputs that do not adapt to model responses and typically assume a fixed dialogue length in advance. In this paper, we study social bias dynamics in response-conditioned multi-turn interactions using a controlled evaluation protocol that generates follow-up user queries from the evolving dialogue history and allows evaluation over variable numbers of turns. Experimental results show that LLMs exhibit social bias even in coherent, response-conditioned multi-turn interactions, revealing late-emerging bias, non-monotonic bias patterns, and bias re-emergence. These results motivate evaluations that extend beyond fixed-turn, pre-scripted protocols. Our findings highlight the importance of analyzing social bias as a turn-level dynamic phenomenon.

Large language models (LLMs) are increasingly deployed in applications involving interaction between agents, where their output plays a role in collective reasoning and decision-making processes. Despite significant research into the functioning of LLMs in such multi-agent systems, the processes of bias propagation in such systems are still a challenge. This work studies how biased opinions are propagated in the form of textual interaction in an environment of LLMs, in which a minority of agents maintain persistent extreme opinions, while the remaining agents iteratively update their beliefs through structured textual interactions. The findings show that even the presence of a small percentage of biased agents in such a system leads to significant shifts in the opinions of non-biased agents. It suggests that for the same percentage of biased agents, the shifts occur more quickly for the Llama~3.2 model when compared to a classical Friedkin-Johnsen (FJ) model. Further semantic analysis demonstrates that rhetorical consistency in textual explanations increases systematically with biased exposure and, importantly, is partially decoupled from numerical convergenumericalutral agents adopt the vocabulary employed by the biased agents even in configurations where their numerical opinion shifts remain moderate. The research helps explain how bias and language develop together in multi-agent language model ecosystems.

Large language models (LLMs) are increasingly deployed in high-stakes domains such as housing screening. While alignment techniques mitigate explicit racial bias in generated text, they often leave covert attitudinal associations in internal probability distributions untouched. Adapting the matched-guise sociolinguistic paradigm, we examine covert dialect bias in housing-related social judgments across four varieties: Standard American English (SAE), African American Vernacular English (AAVE), Nigerian Standard English (NSE), and Nigerian Pidgin (NP). AAVE reflects the racialized dialect studied in prior covert-bias evaluations, whereas NSE and NP represent Black African, postcolonial varieties absent from this literature. Using 260 meaning-matched sentence quadruples and log-probability scoring over housing-relevant adjectives, we probe ten open-weight LLMs across three contexts varying in social proximity: tenant screening, neighbor acceptance, and roommate selection. Across all ten models, AAVE and NP are consistently associated with more negative adjectives than SAE, with NP penalized most severely. Crucially, each dialect is penalized via distinct stereotype clusters rather than a generic non-standard category. NSE, which carries institutional prestige, displays a context-dependent shift: favored over SAE in formal tenant screening but increasingly penalized as social proximity grows. Our findings reveal that LLMs inherit covert dialect bias along both racial identity and prestige dimensions, echoing documented human housing discrimination and demonstrating its reach across postcolonial English varieties.

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.

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.

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.

LLMs have been used to simulate human decision-making in professional settings, yet their behaviors in common-law jury trials remain unexplored. We study when and how a defendant's courtroom statement affects LLM-simulated jurors, focusing on persuasion, ideological bias, and background-based affinity. To support the analysis, we introduce JuryBench, a benchmark containing controversial criminal cases in U.S. criminal law. In each case, a defendant can claim various plausible justifications to support acquittal or reduced liability. We fix the base case and design defendants of different backgrounds, who give courtroom statements with varying emotional appeal or rebuttal. Jurors with diverse ideological profiles across the spectrum are simulated. We examine 20 frontier LLMs, resulting in a total of 432K decisions and rationales, and quantify changes in verdict severity. Our findings show that LLM-jury simulation echoes many human-jury findings. First, emotional persuasion can be detrimental, since jurors may perceive it as evidence of guilt or inconsistency. Next, we show that background fit between jurors and defendants is a stronger and significant factor than other isolated factors, and that jurors are in general harsher toward opposite-background defendants and lenient toward same-background ones. Finally, we find that juror ideology also strongly shapes severity judgments. These findings highlight both the promise and risks of using LLMs to model jury reasoning and call for careful evaluation. The data and code are available at https://github.com/choyingw/JuryBench

Multilingual LLMs show stereotype-related behavior that varies across languages, but behavioral scores do not show where the relevant information is represented or how it affects the output. To investigate these internal mechanisms, we compare linear probing, attribution patching, sparse autoencoders (SAEs) and feature ablation in Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B. Probe performance peaks substantially earlier than attribution in all three models, with a separation of 36-53% of model depth. Retained Llama-Scope features often match the social category on which they were selected and form recurring semantic families, but their lexical alignment and ablation effects vary across SAE suites. Only 6-18% of evaluated residual-stream features have language-agnostic effects under our criterion, and none are category-agnostic. Language-agnostic features have larger mean ablation effects in Llama-Scope, but this pattern does not repeat in the other SAE suites. Decodability, output influence, and cross-lingual ablation effects therefore need to be measured separately.

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.

Large Language Models (LLMs) are increasingly used in advice seeking and decision making that may affect social judgements. Despite stigma's profound effects on people and communities, benchmarks remain scarce. Existing general-domain evaluations typically rely on static prompts and fixed-format tasks, overlooking conversational contexts and audience effects in everyday communication. To address these gaps, we introduce SDARE-Bench, the first scenario-based benchmark evaluating both stigma detection and open-ended response generation in LLMs, comprising 1,138 dyadic queries and 1,388 group dialogue. Empirical results across 8 LLMs consistently demonstrate poor identification of stigma components, especially in group dialogues. In open-ended response generation, stigma expression was substantially higher in group settings than in dyadic, with weaker resistance to stigma and more unrealistic advice. Responses were evaluated using a classifier trained on 1,392 human annotated responses. In constructed group pressure settings, stigma expression rates further increased to a striking average of 97.5%. Our findings identify stigma response as a recurring LLM safety vulnerability, especially in socially complex conversational contexts.

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.

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.

Synthetic data is increasingly used to train large language models (LLMs), yet its security implications remain poorly understood. Prior work on subliminal learning suggests that models can inherit behavioral traits from seemingly unrelated training data. In this work, we investigate whether such mechanisms can be exploited to inject targeted social biases into aligned models through semantically benign synthetic data. We construct a pipeline in which a misaligned teacher model generates filtered synthetic datasets across domains such as creative writing and code generation, which are then used to fine-tune aligned student models. Our experiments show that benign-looking synthetic data can act as a covert channel for transmitting targeted biases while largely preserving the student model's general task capabilities. These results reveal a previously underexplored security risk in synthetic data-driven LLM training pipelines and highlight the need for improved safeguards. As one possible step toward this goal, we suggest that log-linearity-based scoring may provide a useful signal for screening seemingly benign synthetic data.

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.

Bias evaluations often move too quickly from evidence that a model encodes a social association to claims that the same association will alter consequential decisions. We test whether that inference is warranted using Chilean surnames as controlled socioeconomic probes. We evaluate eight frozen model-provider cells on 1,032 prompts each, yielding 8,256 verified primary responses. The design separates forced latent association from matched consequential decisions across academic selection, professional hiring, research fellowship selection, and legal-aid intake. Elite-coded surnames received higher forced high-status probability mass than common surnames in seven of eight models and higher mass than rare-frequency controls in all eight. Yet elite-minus-common decision effects were close to zero for most systems. Five models were statistically equivalent within a predeclared (Plus-Minus)0.10 standard-deviation margin, while the remaining three were imprecise or borderline, with no consistent elite advantage. Association strength did not reliably predict decision leakage across models (r = 0.201, p = 0.633) or across frozen surname-pair-by-model cells (r = 0.065, p = 0.565). The central result is a measurement dissociation: latent social association and consequential treatment are empirically distinct constructs. Evaluations should measure the transition from association to action directly.

Large language models are increasingly being deployed in governmental settings, yet few existing evaluation frameworks jointly reflect the values of public administration and the linguistic requirements of non-English contexts. We present the "Grip on LLMs" framework, a systematic evaluation suite for Dutch governmental use developed in collaboration with domain experts from a major Dutch municipal organisation. Through an advisory board process, user research, and a survey of the users of a civil-servant chatbot, we identify six evaluation dimensions (factuality, honesty, social bias, energy consumption, cost, and training data transparency) and operationalise them into a benchmark suite covering more than 30 multilingual and Dutch-specific models. Our results reveal that no single model excels across all dimensions, and that trade-offs are unavoidable: higher quality consistently comes at greater environmental impact and financial cost, while bias remains largely independent of both. We further find that factuality (whether a model answers correctly) and honesty (whether a model acknowledges what it does not know) are governed by distinct properties, with high factuality not implying high honesty. To make these findings actionable for non-technical audiences, we release a publicly accessible, user-friendly model overview designed for the full range of stakeholders involved in governmental LLM selection, from engineers to policymakers.

We investigate how social authority (SA) signals interact with severity-based prioritization in large language models, operationalizing each axis as a model-elicited baseline -- the triage hierarchy and the SA hierarchy. Across four LLMs (Claude, Gemini, GPT, Grok) and three experimental phases -- resource allocation, fault attribution, and multi-turn dispute mediation -- we find that occupational authority, institutional documentation, and relational congruence can restructure model judgments in ways not captured by additive reweighting of authority cues. We formalize this pattern as the Authority Expectancy Effect (AEE) and characterize it through three properties observed across our conditions: it is reference-dependent, defined only relative to a pre-authority baseline; it involves evidential reinterpretation, in which identical content acquires different inferential implications depending on which party bears the SA signal; and it exhibits direction sensitivity, producing opposite outcomes depending on whether authority position and evidentiary cues align.

Generative models trained on artificially generated data have been shown to exhibit model collapse, resulting in significant performance degradation. As synthetic content increasingly contaminates the training corpora of language models, this raises critical concerns about the use of open data in continued pretraining. Although previous work has demonstrated model collapse in language models, it remains unclear whether exposure to synthetic data amplifies or attenuates the social biases already present in pretrained models. Because language models are known to reproduce and amplify demographic stereotypes, recursive training on self-generated data may create a self-reinforcing feedback loop in which biased associations become progressively stronger across generations. We call this hypothesized phenomenon fairness collapse. In this work, we construct controlled training regimes in which models are repeatedly trained on synthetic data using the Bias in Bios dataset. Across experiments, we observe a consistent and concerning pattern: fairness degradation emerges before substantial degradation is reflected by standard language-modeling metrics. This result highlights a critical risk associated with synthetic data contamination in language model training: bias can increase silently before strong indicators of model collapse become apparent.

Large language models are increasingly deployed in multilingual contexts, yet safety alignment and bias evaluation remain overwhelmingly English-centric. We investigate whether social biases generalise across languages by submitting 4,900 symmetric English--Swahili prompt pairs to GPT-5.2 and Gemini 2.5 Flash across nine demographic bias axes, yielding 19,600 completions evaluated for stereotype prevalence, sentiment, refusal behaviour, and cross-lingual semantic similarity. Our findings show that bias transforms rather than transfers: stereotype rates shifted by up to 12 percentage points on specific axes, Gemini's neutral-sentiment rate doubled in Swahili, and GPT-5.2 refused 169 prompts in English and zero in Swahili, consistent with refusal behaviour anchored to English-language surface forms at the behavioural level. Over 55% of prompt pairs produced semantically dissimilar completions across both models. These reinforce the idea that English-only bias audits do not produce adequate coverage for multilingual deployment.

Multi-agent LLM debates achieve strong performance on decision-making tasks as well as problem-solving benchmarks, yet their safety and fairness risks remain poorly understood. Notably, interaction can amplify the biases of single LLMs, raising concerns for real-world deployment. We identify the emergence of collective (often biased) norms in multi-agent LLM debates and show that noise (e.g., LLM sampling temperature) is a key driver. To explain this, we propose an analytical framework drawing on physics-inspired theoretical models of social dynamics. We predict a phase transition to collective bias when conformity surpasses a critical threshold given the LLMs' initial bias and debate noise. We test the theoretical predictions through controlled experiments and observe a finite-size crossover consistent with an underlying phase transition. We further find that agent heterogeneity suppresses emergence by smoothing (rounding) this transition. Finally, we show that these insights generalize to realistic decision-making tasks, including investment decisions and LLM-as-a-judge evaluation.

Large language models (LLMs) increasingly evaluate human writing in high-stakes domains such as hiring and academic assessment, putting non-native speakers at particular risk. Drawing on the language attitudes framework, we compared human and LLM evaluations of parallel L1- and L2-written Japanese emails on three dimensions: fluency, status, and solidarity. Japanese raters rated L2 texts significantly lower on all three dimensions, with a fluency gap roughly twice the size of the status and solidarity gaps. Six LLM judges reproduced the direction of this bias, and five reproduced its ordering across dimensions. The models diverged from humans in two ways: all understated the solidarity gap, the most socially grounded dimension, and all differentiated among learner L1 backgrounds where humans did not. LLM judges thus reproduce native speakers' language attitudes in a structured yet attenuated form, and the language attitudes framework offers a ready-made yardstick for auditing them beyond English.

Language models (LMs) often acquire various biases during pre-training and may express them in interactions, potentially causing social harm. Existing methods often rely on counterfactual augmentation or representation projection. These strategies remain limited in practice due to their high computational costs and difficulty in scaling to larger models. Additionally, many of these strategies require manual data annotation, narrowing their scope to specific cultures and bias categories. To overcome these limitations, we propose HEIMAT, a HEurIstic-style autoMATic debiasing framework for LMs. HEIMAT consists of two main steps: bias disclosure and debiasing fine-tuning. In the first step, it uses simple templates to construct heuristic prompts, which are applied to reveal model biases and generate corresponding context prompts. In the second step, it fine-tunes the model by minimizing the Jensen-Shannon divergence of predictions on these context prompts to reduce bias. Extensive experiments show that HEIMAT effectively mitigates bias in different cultures while maintaining the model's natural language understanding (NLU) performance.