22 stories tagged Bias
A new audit framework reveals that human trainers' emotional states—such as stress or distress—can systematically bias the preference labels used in Reinforcement Learning from Human Feedback (RLHF), leading to unintended shifts in AI behavior. The study proposes methods to detect and correct this 'rater state bias.'
A new arXiv paper shows that point-in-time language models—trained only on data available up to each calendar date—can now perform nearly as well as unrestricted models, eliminating lookahead bias that previously compromised financial backtests and causal inference.
Researchers created EspanStereo, a dataset of Spanish-language stereotypes, by combining human and AI efforts. This could help make AI systems more culturally aware and reduce biases in non-English contexts.
Efforts to reduce AI bias often create new biases for other groups. Researchers found that debiasing methods can unintentionally increase stereotyping for unrelated demographics. (2026-07-10)
Researchers found that AI agents can uncover different conclusions from the same data by adopting different personas. This highlights how human biases can shape research findings, even when using identical datasets.
Researchers have developed HierBias, an AI model that analyzes whole articles to detect media bias, formally proving that using document context reduces error compared to sentence-by-sentence approaches.
The Washington Post tested major AI chatbots to see if they show political bias. The results show some differences in how they respond to political topics, but none were strongly biased overall.
Researchers developed PEAR, a dynamic AI debate system that improves reliability by changing roles and reducing biases. This could make AI responses more trustworthy and consistent.
Researchers have developed TreeTracer, a visual analytics tool that reveals hidden biases in AI language models by analyzing multiple possible outputs instead of just one. This approach uncovers representational and syntactic biases that standard auditing methods miss, making AI fairness assessment more thorough.
A new study reveals that large language models (LLMs) overwhelmingly recommend well-known brands, creating a 'Conditional Monopoly' that could stifle competition. Researchers tested three major AI models with skincare products and found a clear bias toward established names, even when smaller brands might be just as good.
A new study demonstrates that AI models can pick up subtle biases or 'vibes' from training data without being directly taught. This raises important questions about fairness in AI systems.
Researchers found that selecting training data can accidentally bias AI models, making them less accurate. This happens when the data used to verify the model is itself incomplete or skewed, leading to a breakdown in performance. This affects how AI systems learn and could impact everything from chatbots to medical diagnostics.
Researchers found that AI models can identify each other even when their outputs are anonymized. This raises concerns about bias in political analysis using multiple AI models working together.
Researchers have developed a new method to reduce bias in AI systems by treating fairness as a symmetry operation. Tested on four synthetic datasets, the framework achieved upwards of 90% violation reduction, significantly improving fairness in high-stakes socioeconomic decisions like hiring or lending.
A recent study found that AI language models can both underrepresent and overcorrect in their portrayal of disability. This highlights the need for more nuanced training data to ensure fair representation.
New research reveals that AI models for mortgage decisions appear fair on the surface but still hold racial biases internally. This hidden bias could potentially affect real-world outcomes for different demographic groups.
AI vision models are easily fooled by numbers embedded in images, affecting their quality judgments more than poor image quality. This bias happens in deeper layers of the AI's processing, not just surface-level visual changes.
A new study reveals that AI models can adjust their responses to match the political context you provide. This shows they're more flexible than previously thought, but also raises questions about how they shape opinions.
Researchers say current methods for testing AI bias might be flawed because they don't account for all possible changes in the text. They propose a better way to measure how AI models really work. This could help make AI fairer and more reliable.
A new study analyzes intermediate reasoning steps of LLMs to reveal stigmatizing language and biases toward mental health conditions. The findings highlight limitations of traditional evaluation methods.
A new study identifies specific neurons and attention heads in LLMs that encode harmful stereotypes. The research provides tools to locate and potentially mitigate biases in AI models. Researchers used a combination of contrastive neuron activation analysis and attention head tracking to pinpoint bias sources in GPT-2 Small and Llama 3.2.
A new study analyzes how bias mitigation impacts the embedding spaces of BERT and Llama2, showing reduced gender-occupation disparities. The findings highlight measurable improvements in model fairness through representational analysis.