Bias Amplification: Language Models as Increasingly Biased Media

As Large Language Models (LLMs) become increasingly integrated into various facets of society, a significant portion of online text consequently become synthetic. This raises concerns about bias amplification, a phenomenon where models trained on synthetic data amplify the pre-existing biases over successive training iterations. Previous literature seldom discusses bias amplification as an independent issue from model collapse. In this work, we address the gap in understanding the bias amplification of LLMs with four main contributions. Firstly, we propose a theoretical framework, defining the necessary and sufficient conditions for its occurrence, and emphasizing that it occurs independently of model collapse. Using statistical simulations with weighted maximum likelihood estimation, we demonstrate the framework and show how bias amplification arises without the sampling and functional form issues that typically drive model collapse. Secondly, we conduct experiments with GPT-2 to empirically demonstrate bias amplification, specifically examining open-ended generational political bias with a benchmark we developed. We observe that GPT-2 exhibits a right-leaning bias in sentence continuation tasks and that the bias progressively increases with iterative fine-tuning on synthetic data generated by previous iterations. Thirdly, we explore three potential mitigation strategies: Overfitting, Preservation, and Accumulation. We find that both Preservation and Accumulation effectively mitigate bias amplification and model collapse. Finally, using novel mechanistic interpretation techniques, we demonstrate that in the GPT-2 experiments, bias amplification and model collapse are driven by distinct sets of neurons, which aligns with our theoretical framework.

Assessing Bias in Metric Models for LLM Open-Ended Generation Bias Benchmarks

Open-generation bias benchmarks evaluate social biases in Large Language Models (LLMs) by analyzing their outputs. However, the classifiers used in analysis often have inherent biases, leading to unfair conclusions. This study examines such biases in open-generation benchmarks like BOLD and SAGED. Using the MGSD dataset, we conduct two experiments. The first uses counterfactuals to measure prediction variations across demographic groups by altering stereotype-related prefixes. The second applies explainability tools (SHAP) to validate that the observed biases stem from these counterfactuals. Results reveal unequal treatment of demographic descriptors, calling for more robust bias metric models.

HEARTS: A Holistic Framework for Explainable, Sustainable and Robust Text Stereotype Detection

Stereotypes are generalised assumptions about societal groups, and even state-of-the-art LLMs using in-context learning struggle to identify them accurately. Due to the subjective nature of stereotypes, where what constitutes a stereotype can vary widely depending on cultural, social, and individual perspectives, robust explainability is crucial. Explainable models ensure that these nuanced judgments can be understood and validated by human users, promoting trust and accountability. We address these challenges by introducing HEARTS (Holistic Framework for Explainable, Sustainable, and Robust Text Stereotype Detection), a framework that enhances model performance, minimises carbon footprint, and provides transparent, interpretable explanations. We establish the Expanded Multi-Grain Stereotype Dataset (EMGSD), comprising 57,201 labeled texts across six groups, including under-represented demographics like LGBTQ+ and regional stereotypes. Ablation studies confirm that BERT models fine-tuned on EMGSD outperform those trained on individual components. We then analyse a fine-tuned, carbon-efficient ALBERT-V2 model using SHAP to generate token-level importance values, ensuring alignment with human understanding, and calculate explainability confidence scores by comparing SHAP and LIME outputs. Finally, HEARTS is applied to assess stereotypical bias in 12 LLM outputs, revealing a gradual reduction in bias over time within model families.

THaMES: An End-to-End Tool for Hallucination Mitigation and Evaluation in Large Language Models

Hallucination, the generation of factually incorrect content, is a growing challenge in Large Language Models (LLMs). Existing detection and mitigation methods are often isolated and insufficient for domain-specific needs, lacking a standardized pipeline. This paper introduces THaMES (Tool for Hallucination Mitigations and EvaluationS), an integrated framework and library addressing this gap. THaMES offers an end-to-end solution for evaluating and mitigating hallucinations in LLMs, featuring automated test set generation, multifaceted benchmarking, and adaptable mitigation strategies. It automates test set creation from any corpus, ensuring high data quality, diversity, and cost-efficiency through techniques like batch processing, weighted sampling, and counterfactual validation. THaMES assesses a model's ability to detect and reduce hallucinations across various tasks, including text generation and binary classification, applying optimal mitigation strategies like In-Context Learning (ICL), Retrieval Augmented Generation (RAG), and Parameter-Efficient Fine-tuning (PEFT). Evaluations of state-of-the-art LLMs using a knowledge base of academic papers, political news, and Wikipedia reveal that commercial models like GPT-4o benefit more from RAG than ICL, while open-weight models like Llama-3.1-8B-Instruct and Mistral-Nemo gain more from ICL. Additionally, PEFT significantly enhances the performance of Llama-3.1-8B-Instruct in both evaluation tasks.

SAGED: A Holistic Bias-Benchmarking Pipeline for Language Models with Customisable Fairness Calibration

The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(-Bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that assessment tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. (vice-/former-) presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models.

From Text to Emoji: How PEFT-Driven Personality Manipulation Unleashes the Emoji Potential in LLMs

As the demand for human-like interactions with LLMs continues to grow, so does the interest in manipulating their personality traits, which has emerged as a key area of research. Methods like prompt-based In-Context Knowledge Editing (IKE) and gradient-based Model Editor Networks (MEND) have been explored but show irregularity and variability. IKE depends on the prompt, leading to variability and sensitivity, while MEND yields inconsistent and gibberish outputs. To address this, we employed Opinion QA Based Parameter-Efficient Fine-Tuning (PEFT), specifically Quantized Low-Rank Adaptation (QLORA), to manipulate the Big Five personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. After PEFT, models such as Mistral-7B-Instruct and Llama-2-7B-chat began generating emojis, despite their absence in the PEFT data. For instance, Llama-2-7B-chat generated emojis in 99.5% of extraversion-related test instances, while Mistral-8B-Instruct did so in 92.5% of openness-related test instances. Explainability analysis indicated that the LLMs used emojis intentionally to express these traits. This paper provides a number of novel contributions. First, introducing an Opinion QA dataset for PEFT-driven personality manipulation; second, developing metric models to benchmark LLM personality traits; third, demonstrating PEFT's superiority over IKE in personality manipulation; and finally, analyzing and validating emoji usage through explainability methods such as mechanistic interpretability and in-context learning explainability methods.

JobFair: A Framework for Benchmarking Gender Hiring Bias in Large Language Models

This paper presents a novel framework for benchmarking hierarchical gender hiring bias in Large Language Models (LLMs) for resume scoring, revealing significant issues of reverse bias and overdebiasing. Our contributions are fourfold: First, we introduce a framework using a real, anonymized resume dataset from the Healthcare, Finance, and Construction industries, meticulously used to avoid confounding factors. It evaluates gender hiring biases across hierarchical levels, including Level bias, Spread bias, Taste-based bias, and Statistical bias. This framework can be generalized to other social traits and tasks easily. Second, we propose novel statistical and computational hiring bias metrics based on a counterfactual approach, including Rank After Scoring (RAS), Rank-based Impact Ratio, Permutation Test-Based Metrics, and Fixed Effects Model-based Metrics. These metrics, rooted in labor economics, NLP, and law, enable holistic evaluation of hiring biases. Third, we analyze hiring biases in ten state-of-the-art LLMs. Six out of ten LLMs show significant biases against males in healthcare and finance. An industry-effect regression reveals that the healthcare industry is the most biased against males. GPT-4o and GPT-3.5 are the most biased models, showing significant bias in all three industries. Conversely, Gemini-1.5-Pro, Llama3-8b-Instruct, and Llama3-70b-Instruct are the least biased. The hiring bias of all LLMs, except for Llama3-8b-Instruct and Claude-3-Sonnet, remains consistent regardless of random expansion or reduction of resume content. Finally, we offer a user-friendly demo to facilitate adoption and practical application of the framework.

Intersectional Fairness: A Fractal Approach

The issue of fairness in AI has received an increasing amount of attention in recent years. The problem can be approached by looking at different protected attributes (e.g., ethnicity, gender, etc) independently, but fairness for individual protected attributes does not imply intersectional fairness. In this work, we frame the problem of intersectional fairness within a geometrical setting. We project our data onto a hypercube, and split the analysis of fairness by levels, where each level encodes the number of protected attributes we are intersecting over. We prove mathematically that, while fairness does not propagate "down" the levels, it does propagate "up" the levels. This means that ensuring fairness for all subgroups at the lowest intersectional level (e.g., black women, white women, black men and white men), will necessarily result in fairness for all the above levels, including each of the protected attributes (e.g., ethnicity and gender) taken independently. We also derive a formula describing the variance of the set of estimated success rates on each level, under the assumption of perfect fairness. Using this theoretical finding as a benchmark, we define a family of metrics which capture overall intersectional bias. Finally, we propose that fairness can be metaphorically thought of as a "fractal" problem. In fractals, patterns at the smallest scale repeat at a larger scale. We see from this example that tackling the problem at the lowest possible level, in a bottom-up manner, leads to the natural emergence of fair AI. We suggest that trustworthiness is necessarily an emergent, fractal and relational property of the AI system.

Local and Global Explainability Metrics for Machine Learning Predictions

Rapid advancements in artificial intelligence (AI) technology have brought about a plethora of new challenges in terms of governance and regulation. AI systems are being integrated into various industries and sectors, creating a demand from decision-makers to possess a comprehensive and nuanced understanding of the capabilities and limitations of these systems. One critical aspect of this demand is the ability to explain the results of machine learning models, which is crucial to promoting transparency and trust in AI systems, as well as fundamental in helping machine learning models to be trained ethically. In this paper, we present novel quantitative metrics frameworks for interpreting the predictions of classifier and regressor models. The proposed metrics are model agnostic and are defined in order to be able to quantify: i. the interpretability factors based on global and local feature importance distributions; ii. the variability of feature impact on the model output; and iii. the complexity of feature interactions within model decisions. We employ publicly available datasets to apply our proposed metrics to various machine learning models focused on predicting customers' credit risk (classification task) and real estate price valuation (regression task). The results expose how these metrics can provide a more comprehensive understanding of model predictions and facilitate better communication between decision-makers and stakeholders, thereby increasing the overall transparency and accountability of AI systems.

Uncovering Bias in Face Generation Models

Recent advancements in GANs and diffusion models have enabled the creation of high-resolution, hyper-realistic images. However, these models may misrepresent certain social groups and present bias. Understanding bias in these models remains an important research question, especially for tasks that support critical decision-making and could affect minorities. The contribution of this work is a novel analysis covering architectures and embedding spaces for fine-grained understanding of bias over three approaches: generators, attribute modifier, and post-processing bias mitigators. This work shows that generators suffer from bias across all social groups with attribute preferences such as between 75%-85% for whiteness and 60%-80% for the female gender (for all trained CelebA models) and low probabilities of generating children and older men. Modifier and mitigators work as post-processor and change the generator performance. For instance, attribute channel perturbation strategies modify the embedding spaces. We quantify the influence of this change on group fairness by measuring the impact on image quality and group features. Specifically, we use the Fr\'echet Inception Distance (FID), the Face Matching Error and the Self-Similarity score. For Interfacegan, we analyze one and two attribute channel perturbations and examine the effect on the fairness distribution and the quality of the image. Finally, we analyzed the post-processing bias mitigators, which are the fastest and most computationally efficient way to mitigate bias. We find that these mitigation techniques show similar results on KL divergence and FID score, however, self-similarity scores show a different feature concentration on the new groups of the data distribution. The weaknesses and ongoing challenges described in this work must be considered in the pursuit of creating fair and unbiased face generation models.