Warden AI provides independent, tech-led AI bias auditing, designed for both HR Tech platforms and enterprises deploying AI solutions in HR. As the adoption of AI in recruitment and HR processes grows, concerns around fairness have intensified. With the advent of regulations such as NYC Local Law 144 and the EU AI Act, organisations are under increasing pressure to demonstrate compliance and fairness.

This document addresses bias in relation to AI systems, especially with regards to AI-aided decision-making.

Diagnose bias in LLMs (Large Language Models) from various points of views, allowing users to choose the most appropriate LLM.

Practical recommendations for mitigating bias in automated speaker recognition, and outline future research directions.

Repository of algorithmic bias related metrics and measures to enable researchers and practitioners to leverage their use. The repository is also intended to allow researchers to expand their research to identify more metrics that may be relevant and appropriate to specific context.

This bias detection tool identifies potentially unfairly treated groups of similar users by a binary algorithmic classifier. The tool identifies clusters of users that face a higher misclassification rate compared to the rest of the data set. Clustering is an unsupervised ML method, so no data is needed is required on protected attributes of users. The metric by which bias is defined can be manually chosen in advance: False Negative Rate (FNR), False Positive Rate (FPR), or Accuracy (Acc).

Biaslyze is a python package that helps to get started with the analysis of bias within NLP models and offers a concrete entry point for further impact assessments and mitigation measures.

[ICML 2021] Towards Understanding and Mitigating Social Biases in Language Models

The report steers its way to identify the challenges often faced in the process of bias mitigation. The challenges have been categorized into organizational levels, industry-wide levels and societal levels and includes limitations such as lack of domain knowledge and accountability, lack of regulations and actionable guidance, persistence of black box algorithms and outdated education approaches for data and scientists.

bt4vt is a python library to diagnose performance discrepancies (i.e. bias) in automatic speech processing models.

This course explains what is meant by bias in the context of machine learning algorithms, how bias can be present in the training data, and how it can be unintentionally introduced in the learning phase. Learners will gain an appreciation of why this is a concern and why it needs to be addressed. It is is divided into two sections: milestones 1-2 address the conceptual background and context, milestones 3-5 explore practical applications.

A fully independent and impartial audit to ensure compliance with the bias audit requirements for New York City Local Law 144 and upcoming regulations.

Machine learning approach to detect algorithmic bias in all types of binary AI classifiers. No protected user characteristics needed. Identifies unfairly treated groups characterized by mixture of features, detects intersectional bias. Model-agnostic, open-source web application, easy to use for the entire AI auditing community, e.g., journalists, data scientists, policy makers etc. Finalist of Stanford's AI Audit Competition.

Open Sourced Bias Testing for Generalized Machine Learning Applications audit-AI is a Python library built on top of pandas and sklearn that implements fairness-aware machine learning algorithms. audit-AI was developed by the Data Science team at pymetrics

HireVue recognizes the impact that their software can have on individuals and on society, and they act upon this responsibility with deep commitment. The following principles guide their thoughts and actions as they develop artificial intelligence (AI) technology and incorporate it into their products and technologies. 1. They are committed to benefiting society 2. They […]

Aequitas is an open source bias and fairness audit toolkit that is an intuitive and easy to use addition to the machine learning workflow, enabling users to seamlessly test models for several bias and fairness metrics in relation to multiple population sub-groups. Aequitas facilitates informed and equitable decisions around developing and deploying algorithmic decision making […]

Using crowd-worker evaluation as well as standard benchmarks, we empirically demonstrate that our algorithms significantly reduce gender bias in embeddings while preserving the its useful properties such as the ability to cluster related concepts and to solve analogy tasks. The resulting embeddings can be used in applications without amplifying gender bias.

This paper presents RAN-Debias, a gender de-biasing methodology which eliminates the bias present in a word vector and alters the spatial distribution of its neighboring vectors, achieving a bias-free setting while maintaining minimal semantic offset. This paper also proposes a new bias evaluation metric – Gender-based Illicit Proximity Estimate (GIPE), which measures the extent of […]

” ”