Bias in Foundation Models (FMs) - trained on vast datasets spanning societal and historical knowledge - poses significant challenges for fairness and equity across fields such as healthcare, education, and finance. These biases, rooted in the overrepresentation of stereotypes and societal inequalities in training data, exacerbate real-world discrimination, reinforce harmful stereotypes, and erode trust in AI systems. To address this, we introduce Trident Probe Testing (TriProTesting), a systematic testing method that detects explicit and implicit biases using semantically designed probes. Here we show that FMs, including CLIP, ALIGN, BridgeTower, and OWLv2, demonstrate pervasive biases across single and mixed social attributes (gender, race, age, and occupation). Notably, we uncover mixed biases when social attributes are combined, such as gender x race, gender x age, and gender x occupation, revealing deeper layers of discrimination. We further propose Adaptive Logit Adjustment (AdaLogAdjustment), a post-processing technique that dynamically redistributes probability power to mitigate these biases effectively, achieving significant improvements in fairness without retraining models. These findings highlight the urgent need for ethical AI practices and interdisciplinary solutions to address biases not only at the model level but also in societal structures. Our work provides a scalable and interpretable solution that advances fairness in AI systems while offering practical insights for future research on fair AI technologies.

Warning: This paper contains several contents that may be toxic, harmful, or offensive. In the last few years, text-to-image generative models have gained remarkable success in generating images with unprecedented quality accompanied by a breakthrough of inference speed. Despite their rapid progress, human biases that manifest in the training examples, particularly with regard to common stereotypical biases, like gender and skin tone, still have been found in these generative models. In this work, we seek to measure more complex human biases exist in the task of text-to-image generations. Inspired by the well-known Implicit Association Test (IAT) from social psychology, we propose a novel Text-to-Image Association Test (T2IAT) framework that quantifies the implicit stereotypes between concepts and valence, and those in the images. We replicate the previously documented bias tests on generative models, including morally neutral tests on flowers and insects as well as demographic stereotypical tests on diverse social attributes. The results of these experiments demonstrate the presence of complex stereotypical behaviors in image generations.

Bias research in NLP seeks to analyse models for social biases, thus helping NLP practitioners uncover, measure, and mitigate social harms. We analyse the body of work that uses prompts and templates to assess bias in language models. We draw on a measurement modelling framework to create a taxonomy of attributes that capture what a bias test aims to measure and how that measurement is carried out. By applying this taxonomy to 90 bias tests, we illustrate qualitatively and quantitatively that core aspects of bias test conceptualisations and operationalisations are frequently unstated or ambiguous, carry implicit assumptions, or be mismatched. Our analysis illuminates the scope of possible bias types the field is able to measure, and reveals types that are as yet under-researched. We offer guidance to enable the community to explore a wider section of the possible bias space, and to better close the gap between desired outcomes and experimental design, both for bias and for evaluating language models more broadly.

Due to their similarity-based learning objectives, pretrained sentence encoders often internalize stereotypical assumptions that reflect the social biases that exist within their training corpora. In this paper, we describe several kinds of stereotypes concerning different communities that are present in popular sentence representation models, including pretrained next sentence prediction and contrastive sentence representation models. We compare such models to textual entailment models that learn language logic for a variety of downstream language understanding tasks. By comparing strong pretrained models based on text similarity with textual entailment learning, we conclude that the explicit logic learning with textual entailment can significantly reduce bias and improve the recognition of social communities, without an explicit de-biasing process

The awareness and mitigation of biases are of fundamental importance for the fair and transparent use of contextual language models, yet they crucially depend on the accurate detection of biases as a precursor. Consequently, numerous bias detection methods have been proposed, which vary in their approach, the considered type of bias, and the data used for evaluation. However, while most detection methods are derived from the word embedding association test for static word embeddings, the reported results are heterogeneous, inconsistent, and ultimately inconclusive. To address this issue, we conduct a rigorous analysis and comparison of bias detection methods for contextual language models. Our results show that minor design and implementation decisions (or errors) have a substantial and often significant impact on the derived bias scores. Overall, we find the state of the field to be both worse than previously acknowledged due to systematic and propagated errors in implementations, yet better than anticipated since divergent results in the literature homogenize after accounting for implementation errors. Based on our findings, we conclude with a discussion of paths towards more robust and consistent bias detection methods.

We generalize the notion of social biases from language embeddings to grounded vision and language embeddings. Biases are present in grounded embeddings, and indeed seem to be equally or more significant than for ungrounded embeddings. This is despite the fact that vision and language can suffer from different biases, which one might hope could attenuate the biases in both. Multiple ways exist to generalize metrics measuring bias in word embeddings to this new setting. We introduce the space of generalizations (Grounded-WEAT and Grounded-SEAT) and demonstrate that three generalizations answer different yet important questions about how biases, language, and vision interact. These metrics are used on a new dataset, the first for grounded bias, created by augmenting extending standard linguistic bias benchmarks with 10,228 images from COCO, Conceptual Captions, and Google Images. Dataset construction is challenging because vision datasets are themselves very biased. The presence of these biases in systems will begin to have real-world consequences as they are deployed, making carefully measuring bias and then mitigating it critical to building a fair society.

A new approach for testing whether algorithms contain hidden biases aims to prevent automated systems from perpetuating human discrimination. Machine learning is increasingly being used to make sensitive decisions, says Matt Kusner at the Alan Turing Institute in London. In some US states, judges make sentencing decisions and set bail conditions using algorithms that calculate the likelihood that someone will reoffend. Other algorithms assess whether a person should be offered a loan or a job interview. But it is often unclear how these systems come to their conclusions, which makes it impossible to tell if they are fair ones. An algorithm might conclude that people from a certain demographic are less likely to pay back a loan, for example, if it is trained on a data set in which loans were unfairly distributed in the first place.