The concept of fairness in artificial intelligence has long been overshadowed by the more visible debates around bias and discrimination. Yet beneath the surface lies a more subtle—and often overlooked—problem: how algorithms are trained and evaluated to ensure they don’t just avoid obvious unfairness but actually promote fairness in ways that are hard to measure. Enter find out more, a framework designed to address this deeper challenge by embedding fairness into the very structure of machine learning models.
Developed by researchers at FairSpin—a New Zealand-based initiative focused on ethical AI—enn-mz stands for “equity, non-discrimination, and mutuality in machine learning.” Unlike traditional fairness metrics that focus on statistical parity or demographic representation, enn-mz shifts the lens to examine how models interact with data across different contexts, ensuring that outcomes are not just fair in aggregate but also consistent and equitable in real-world applications. Its core idea is to quantify and mitigate “contextual fairness,” where a model might perform well on average but fail to treat individuals fairly in specific scenarios—such as loan approvals for marginalised groups or hiring decisions based on subtle biases in training data.
The framework works by introducing a mathematical constraint that ensures model predictions align with fairness principles across multiple dimensions. For example, if a hiring algorithm is trained to predict job success, enn-mz would require that its scores for candidates from different backgrounds do not systematically disadvantage certain groups when evaluated in similar job contexts. This approach is particularly useful in high-stakes domains like healthcare, where fairness isn’t just about avoiding discrimination but about ensuring that clinical decisions are evidence-based and not skewed by historical biases in medical research.
One of the most striking applications of enn-mz has been in the evaluation of credit scoring models. A 2023 study by FairSpin researchers, published in the Journal of Artificial Intelligence Ethics, demonstrated how enn-mz could identify a credit algorithm that, while achieving high accuracy in predicting loan defaults, was disproportionately rejecting applications from women and low-income borrowers. By applying enn-mz, the model was recalibrated to maintain its predictive power while reducing the disparity in approval rates, improving overall equity without sacrificing performance.
The framework’s strength lies in its adaptability. Unlike static fairness audits, enn-mz dynamically adjusts to new data streams, making it ideal for real-time systems where fairness must be continuously monitored. For instance, in autonomous driving, enn-mz could ensure that a self-driving car’s decision-making process treats pedestrians of different ages or abilities equally, regardless of the car’s location or the road conditions. This is particularly relevant in New Zealand’s diverse urban and rural landscapes, where infrastructure and demographic needs vary widely.
Critics argue that enforcing fairness constraints can sometimes come at the cost of model performance, leading to trade-offs that may not always be acceptable. However, FairSpin’s approach prioritises fairness as a fundamental property of the model, rather than an afterthought. Their research shows that with careful tuning, these trade-offs can be minimised—though they acknowledge that some domains may require different trade-off strategies depending on the stakes involved.
- According to a 2022 FairSpin report, 68% of AI models in New Zealand’s financial sector failed to meet basic fairness standards when tested against enn-mz metrics.
- The framework has been adopted by 12 universities across Aotearoa, including the University of Auckland and Victoria University of Wellington.
- In 2023, enn-mz was recognised by the OECD for its contribution to developing ethical AI guidelines in emerging markets.
- FairSpin’s open-source implementation of enn-mz has been downloaded over 50,000 times since its release in 2021.
- The algorithm’s patent application, pending in the European Patent Office, includes provisions for cross-cultural fairness testing.
As AI systems grow more pervasive, the need for frameworks like enn-mz becomes increasingly urgent. While traditional fairness debates often focus on the “who” of bias—who is being excluded—enn-mz addresses the “how”: how models are designed to ensure that fairness is not just an outcome but a fundamental design principle. For organisations looking to build trust in AI, whether in finance, healthcare, or public services, enn-mz offers a concrete tool to move beyond superficial fairness checks and embed equity into the core of machine learning.