Developing Bias Auditing and Mitigation Tools for Self-Assessment of AI Conformity with the EU AI Act through Statistical Matching

Introduction

The vision behind Act.AI is to utilize statistical matching for mitigating and auditing bias in Artificial Intelligence (AI) models. AI has been rapidly growing in various industries, from financial services to healthcare, education, and job recruitment.

Concerns About AI Fairness

However, as AI algorithms have become increasingly sophisticated and pervasive in decision-making processes, concerns have arisen about their fairness and compliance with regulations. In particular, the EU AI Act requires that AI providers in high-risk applications -- such as employment, credit, or healthcare -- identify (and thereby address) discrimination by their algorithms against certain demographics of people.

Challenges for AI Startups

Ensuring compliance with the Act can be challenging, particularly for AI startups that may not have the resources or expertise to fully understand and implement the Act's requirements.

Addressing Disconnects

Addressing existing disconnects between AI fairness toolkits' capabilities and current practitioner needs, the Act.AI tool can be easily integrated into any AI workflow, in a plug-and-play fashion, to continuously monitor and improve its fairness.

Key Features of Act.AI

A key aspect of Act.AI is the ability to operate with different types of data, including:

• Tabular data
• Images
• Text

It functions in a variety of contexts, such as:

1. Binary classification
2. Multiclass classification
3. Regression

Additionally, it is able to match datasets in different domains, including out-of-distribution data, even if these datasets have different numbers of variables or features.

Stakeholder Integration

To ensure usability of Act.AI, it will integrate feedback from relevant stakeholders from two immediate target markets:

• Financial services
• Healthcare