A facial recognition system deployed by the Metropolitan police at Oxford Circus on 13 May in London. A facial recognition system deployed by the Metropolitan police at Oxford Circus on 13 May in London. 'Urgent clarity' sought over racial bias in UK police facial recognition technology The UK's data protection watchdog has asked the Home Office for "urgent clarity" over racial bias in police facial recognition technology before considering its next steps. The Home Office has admitted that the technology was "more likely to incorrectly include some demographic groups in its search results", after testing by the National Physical Laboratory (NPL) of its application within the police national database. The report revealed that the technology, which is intended to be used to catch serious offenders, is more likely to incorrectly match black and Asian people than their white counterparts.

LLMs are increasingly being used in healthcare. This promises to free physicians from drudgery, enabling better care to be delivered at scale. But the use of LLMs in this space also brings risks; for example, such models may worsen existing biases. How can we spot when LLMs are (spuriously) relying on patient race to inform predictions? In this work we assess the degree to which Sparse Autoencoders (SAEs) can reveal (and control) associations the model has made between race and stigmatizing concepts. We first identify SAE latents in Gemma-2 models which appear to correlate with Black individuals. We find that this latent activates on reasonable input sequences (e.g., "African American") but also problematic words like "incarceration". We then show that we can use this latent to steer models to generate outputs about Black patients, and further that this can induce problematic associations in model outputs as a result. For example, activating the Black latent increases the risk assigned to the probability that a patient will become "belligerent". We evaluate the degree to which such steering via latents might be useful for mitigating bias. We find that this offers improvements in simple settings, but is less successful for more realistic and complex clinical tasks. Overall, our results suggest that: SAEs may offer a useful tool in clinical applications of LLMs to identify problematic reliance on demographics but mitigating bias via SAE steering appears to be of marginal utility for realistic tasks.

Generative models are popular for medical imaging tasks such as anomaly detection, feature extraction, data visualization, or image generation. Since they are parameterized by deep learning models, they are often sensitive to distribution shifts and unreliable when applied to out-of-distribution data, creating a risk of, e.g. underrepresentation bias. This behavior can be flagged using uncertainty quantification methods for generative models, but their availability remains limited. We propose SLUG: A new UQ method for VAEs that combines recent advances in Laplace approximations with stochastic trace estimators to scale gracefully with image dimensionality. We show that our UQ score -- unlike the VAE's encoder variances -- correlates strongly with reconstruction error and racial underrepresentation bias for dermatological images. We also show how pixel-wise uncertainty can detect out-of-distribution image content such as ink, rulers, and patches, which is known to induce learning shortcuts in predictive models.

Individuals from minority groups, even with equivalent qualifications, consistently receive fewer opportunities in critical areas such as employment, education, and healthcare. Yet, empirically demonstrating the existence of such pervasive bias, let alone measuring the extent of bias or correcting it, remains a significant challenge. Over several decades, researchers have utilized a range of experimental methodologies to test for biases in real-life situations (Bertrand and Duflo 2017). Audit studies, among the earliest of such methods, match two individuals who are similar in all respects except for sensitive characteristics like race, to test decision-makers' biases (Ayres and Siegelman 1995). A significant limitation of this method, however, is the inherent impossibility of achieving an exact match between two individuals, precluding perfect comparability (Heckman 1998). Correspondence studies have emerged as a predominant experimental approach for measuring biases (Guryan and Charles 2013, Bertrand and Mullainathan 2004). They create identical fictional profiles with manipulated attributes like race to assess differential treatment. However, these studies traditionally manipulate solely textual information, which may not reflect contemporary decision-making scenarios increasingly influenced by visual cues like facial images, as seen in recent hiring processes (Acquisti and Fong 2020, Ruffle and Shtudiner 2015). This reliance on text limits their effectiveness, as modern contexts often involve multimedia elements, making it challenging to measure real-world biases accurately or correct them based on such incomplete information (Armbruster et al. 2015).

We employ model pruning to examine how LLMs conceptualize racial biases, and whether a generalizable mitigation strategy for such biases appears feasible. Our analysis yields several novel insights. We find that pruning can be an effective method to reduce bias without significantly increasing anomalous model behavior. Neuron-based pruning strategies generally yield better results than approaches pruning entire attention heads. However, our results also show that the effectiveness of either approach quickly deteriorates as pruning strategies become more generalized. For instance, a model that is trained on removing racial biases in the context of financial decision-making poorly generalizes to biases in commercial transactions. Overall, our analysis suggests that racial biases are only partially represented as a general concept within language models. The other part of these biases is highly context-specific, suggesting that generalizable mitigation strategies may be of limited effectiveness. Our findings have important implications for legal frameworks surrounding AI. In particular, they suggest that an effective mitigation strategy should include the allocation of legal responsibility on those that deploy models in a specific use case.

Large Language Models(LLMs) have revolutionized various applications in natural language processing (NLP) by providing unprecedented text generation, translation, and comprehension capabilities. However, their widespread deployment has brought to light significant concerns regarding biases embedded within these models. This paper presents a comprehensive survey of biases in LLMs, aiming to provide an extensive review of the types, sources, impacts, and mitigation strategies related to these biases. We systematically categorize biases into several dimensions. Our survey synthesizes current research findings and discusses the implications of biases in real-world applications. Additionally, we critically assess existing bias mitigation techniques and propose future research directions to enhance fairness and equity in LLMs. This survey serves as a foundational resource for researchers, practitioners, and policymakers concerned with addressing and understanding biases in LLMs.

Demographic bias is one of the major challenges for face recognition systems. The majority of existing studies on demographic biases are heavily dependent on specific demographic groups or demographic classifier, making it difficult to address performance for unrecognised groups. This paper introduces ``LabellessFace'', a novel framework that improves demographic bias in face recognition without requiring demographic group labeling typically required for fairness considerations. We propose a novel fairness enhancement metric called the class favoritism level, which assesses the extent of favoritism towards specific classes across the dataset. Leveraging this metric, we introduce the fair class margin penalty, an extension of existing margin-based metric learning. This method dynamically adjusts learning parameters based on class favoritism levels, promoting fairness across all attributes. By treating each class as an individual in facial recognition systems, we facilitate learning that minimizes biases in authentication accuracy among individuals. Comprehensive experiments have demonstrated that our proposed method is effective for enhancing fairness while maintaining authentication accuracy.

This study addresses the racial biases in facial expression recognition (FER) systems within Large Multimodal Foundation Models (LMFMs). Despite advances in deep learning and the availability of diverse datasets, FER systems often exhibit higher error rates for individuals with darker skin tones. Existing research predominantly focuses on traditional FER models (CNNs, RNNs, ViTs), leaving a gap in understanding racial biases in LMFMs. We benchmark four leading LMFMs: GPT-4o, PaliGemma, Gemini, and CLIP to assess their performance in facial emotion detection across different racial demographics. A linear classifier trained on CLIP embeddings obtains accuracies of 95.9\% for RADIATE, 90.3\% for Tarr, and 99.5\% for Chicago Face. Furthermore, we identify that Anger is misclassified as Disgust 2.1 times more often in Black Females than White Females. This study highlights the need for fairer FER systems and establishes a foundation for developing unbiased, accurate FER technologies. Visit https://kvjvhub.github.io/FERRacialBias/ for further information regarding the biases within facial expression recognition.

This study evaluates whether large language models (LLMs) exhibit biases towards medical professionals. Fictitious candidate resumes were created to control for identity factors while maintaining consistent qualifications. Three LLMs (GPT-4, Claude-3-haiku, and Mistral-Large) were tested using a standardized prompt to evaluate resumes for specific residency programs. Explicit bias was tested by changing gender and race information, while implicit bias was tested by changing names while hiding race and gender. Physician data from the Association of American Medical Colleges was used to compare with real-world demographics. 900,000 resumes were evaluated. All LLMs exhibited significant gender and racial biases across medical specialties. Gender preferences varied, favoring male candidates in surgery and orthopedics, while preferring females in dermatology, family medicine, obstetrics and gynecology, pediatrics, and psychiatry. Claude-3 and Mistral-Large generally favored Asian candidates, while GPT-4 preferred Black and Hispanic candidates in several specialties. Tests revealed strong preferences towards Hispanic females and Asian males in various specialties. Compared to real-world data, LLMs consistently chose higher proportions of female and underrepresented racial candidates than their actual representation in the medical workforce. GPT-4, Claude-3, and Mistral-Large showed significant gender and racial biases when evaluating medical professionals for residency selection. These findings highlight the potential for LLMs to perpetuate biases and compromise healthcare workforce diversity if used without proper bias mitigation strategies.

Just weeks after Google was forced to shut down its'woke' AI, another tech giant faces criticism over its bot's racial bias. Meta's AI image generator has been accused of being'racist' after users discovered it was unable to imagine an Asian man with a white woman. The AI tool, created by Facebook's parent company, is able to take almost any written prompt and convert it into a shockingly realistic image within seconds. However, users found the AI was unable to create images showing mixed-race couples, despite the fact that Meta CEO Mark Zuckerberg is himself married to an Asian woman. On social media, commenters have criticised this as an example of the AI's racial bias, with one describing the AI as'racist software made by racist engineers'.