The importance of end-to-end, socio-technical audits

To ensure the ongoing integrity and fairness of AI-based recruitment systems, continuous audits are essential. Audits provide an opportunity to examine the decision-making processes, algorithmic operations and outcomes of the system, identifying any biases or disparities that may have emerged over time. By implementing audit mechanisms throughout the development life cycle of AI-assisted hiring systems, organisations can proactively identify and mitigate biases, ensuring equity and inclusion in recruitment processes. These audits serve as fundamental tools in promoting accountability, fostering trust among candidates and regulatory bodies and upholding ethical standards.

• The FINDHR Impact Assessment and Auditing Framework proposes an End-to-End, Socio-Technical Algorithmic Audit (E2EST/AA) to inspect AI systems in context, looking at the specific data used and the constituencies impacted by its use. It is end-to-end because it ensures continuous monitoring, evaluation and development across the different phases of the algorithmic life cycle from beginning to end.
• Pre-processing refers to and involves various steps of cleaning, transforming and preparing the raw data before it is used for training and testing a model.
• In-processing refers to the techniques and methods that are applied directly during the training phase of a machine learning model.
• Lastly, post-processing includes the techniques and methods that are applied after the model has been trained.

Additionally, the E2EST/AA is socio-technical because it recognises that algorithmic systems work with data produced by complex and imperfect individuals and societies and operate and intervene in complex social and organisational contexts. An E2EST/AA addresses key issues across the system lifecycle, including commissioning, problem definition, training, testing, validation, deployment and real-world use. As AI systems are deeply socio-technical, focusing on technical issues alone would fail to incorporate both problems and possibilities for system improvement and impact testing, the core of the in-processing phase.

Most of the E2EST/AA focuses on the pre-processing and post-processing stages of the algorithmic life cycle. Models and systems that have optimal performance and accuracy rates during in-processing may perform in inefficient or harmful ways, and this may only be revealed through an end-to-end, socio-technical audit process. For example, a machine learning model might achieve impressive accuracy in predicting outcomes during testing but could struggle when faced with diverse, real-world data. This discrepancy might lead to biased decisions or unfair treatment of certain groups. Additionally, a system that operates efficiently in controlled environments may encounter unforeseen challenges in dynamic settings, leading to suboptimal performance or unintended negative consequences, such as increased resource consumption or safety risks.¹

The methodology proposed focuses on bias assessment but is not limited to it. E2EST/AA incorporates questions related to broader social impact and desirability, as well as the incorporation of end-users in the design process and the existence of recourse mechanisms for those impacted by algorithmic systems. For a system to pass an algorithmic audit, issues of impact, proportionality, participation and resources (considering human resources, hardware, software, development, time, etc.) must be tackled. Two main components of an E2EST/AA are the System Card (SC) and the System Map (SM) → see chapter #12

→ For more general details on the proposed framework, read the FINDHR Impact Assessment and Auditing Framework, section 3