This research formalizes bias in machine learning systems as a symmetry-breaking operation, defining fairness as invariance under counterfactual sensitive attribute switching. It implements loss-based regularization as a symmetry-restoring mechanism, achieving over 90% bias violation reduction with approximately 5% accuracy cost.
This article explores a critical perspective on artificial intelligence, arguing that current models and their developers, particularly concerning ethics and bias, will not be the ultimate solution to the challenges we face. It draws parallels with historical concepts to analyze the inherent limitations and ideological influences within AI technology.
The article discusses how AI age estimation technology used at borders frequently misidentifies child asylum seekers as adults, leading to severe consequences for vulnerable individuals. It highlights the ethical failures and inherent biases in these systems, advocating for a more humane approach that prioritizes children's safety and well-being over flawed algorithmic assessments.
Researchers state that AI chatbots, including Claude and ChatGPT, exhibit a bias towards Catholicism. This observation raises questions about the neutrality and inherent prejudices within advanced language models.
This paper investigates how Large Language Models (LLMs) represent disability by simulating the perspectives of individuals with disabilities in generating social media posts. These posts are then compared with those written by real people with disabilities to analyze the perpetuation or overcorrection of biases.
This study investigates how conversational context affects LLM answers, especially in high-stakes scenarios. It demonstrates that conversation topics are the main predictors of LLM-generated advice, influencing outcome disparities.
SocioEval é um framework baseado em templates para avaliar sistematicamente o viés de status socioeconômico em modelos de fundação, incluindo LLMs, uma área pouco explorada. A pesquisa avaliou 13 LLMs e revelou variações substanciais nas taxas de viés (0,42% a 33,75%), manifestando-se de forma diferente em vários temas.
This research paper demonstrates that embedded numeric anchors on images systematically bias Vision-Language Model quality judgments across multiple VLMs. Layer-wise probing reveals that optimal layers for quality prediction are deeper than where anchor classification saturates, establishing a causal account of visual anchoring bias.
Large language models (LLMs) exhibit consistent asymmetries when advising on religious conversions, favoring some religions like Catholic, Baháʼí, and Sikh, while subtly discouraging others such as Atheists and Jehovah's Witnesses. These biases vary by model and provider, with Grok 4.20 showing the strongest asymmetries, identified through an LLM-as-a-judge framework.
This paper investigates counterfactual unfairness in LLMs by analyzing how their responses to humor change when swapping speaker and addressee identities. Experiments reveal consistent relational disparities, where jokes told by privileged speakers are more often refused or judged as malicious by the models.
This research systematically evaluates the relationship between differential privacy (DP) and social bias in large language models (LLMs). It compares a DP-trained LLM against non-DP baselines across various tasks, finding that DP reduces bias in sentence scoring but not universally, and reveals a discrepancy between logit-level and output-level bias.
The article addresses the critical challenge of ensuring reliability in AI-assisted research, where the bottleneck is no longer information access but the accuracy of AI outputs. It proposes a three-layer model—retrieval integrity, reasoning fidelity, and output verifiability—to evaluate AI tools for research.
arXiv paper 2605.11145 introduces DPAA, a debiasing framework for GNN-based collaborative filtering that applies adaptive weighting during message passing. It effectively counters popularity amplification and outperforms previous debiasing methods on real-world datasets.
This article discusses a 2026 paper by Kumaran et al. identifying two critical, asymmetric biases in LLMs: a choice-supportive bias where models gain confidence in their prior answers, and a hypersensitivity to contradiction causing them to over-adjust when challenged. These findings have significant implications for developers building on top of LLMs, influencing how we interact with AI.
This paper identifies and formalizes
This study proposes a structured framework to improve LLM reasoning when analyzing long documents, addressing issues like contextual bias and omission error. It combines parallel chunk-level processing with evidence-anchored consolidation to generate more robust and bias-resilient conceptual abstractions.
This paper evaluates open-weight and domain-adapted Large Language Models (LLMs) for conflict event classification in West Africa. The study reveals that open-weight models exhibit a "False Illegitimation" bias, while domain-adapted models achieve directional neutrality but retain an actor-based selection bias.
This study investigates the impact of post-training quantization on Large Language Models (LLMs) quality, revealing that compression can lead to bias emergence. 3-bit quantization caused 6-21% of previously unbiased items to develop new stereotypical behaviors in models like Qwen2.5-7B, Mistral-7B, and Phi-3.5-mini. This follows a clear dose-response pattern across various precision levels.
This paper introduces a causal framework to study rationalization bias in LLMs used as automatic judges for summarization and dialogue evaluation. It investigates whether LLM judges' rankings and explanations remain stable when non-evidential cues are perturbed, proposing cue interventions and anchoring metrics.
Este trabalho propõe um novo modelo para treinar Redes Neurais Gráficas (GNNs) sensíveis à justiça, aprimorando o framework CAF. A abordagem utiliza uma estratégia de treinamento em duas fases, editando o grafo para ajustar a homofilia e integrando perdas contrastivas e ambientais modificadas para melhorar a predição e a justiça.