Introduction

Hiring is expensive and time-consuming. Employers often deal with hundreds of applicants per job opening and are often under great time pressure. Applicant Tracking Systems (ATS) streamline the application process by allowing candidates to submit an application and enabling recruiters to manage and review it. Increasingly, these systems incorporate AI-driven features that automate tasks such as filtering, pre-sorting and ranking candidates. These recruitment systems leverage search, recommendation and ranking algorithms to identify and suggest candidates to HR personnel based on specific criteria, returning a ranked list of candidates, where the most promising candidates (according to the system) are in the top positions. In this Toolkit we focus on ATS that include subsystems to return a ranked list of candidates. ATS can have other automated features and be supported by different AI-assisted tools, which can save time for recruiters but also reinforce existing patterns of discrimination or create new ones – often without the knowledge of the people using them. This can also include intersectional forms of discrimination which take place when (due to societal inequalities) various personal characteristics such as race, ethnicity, age, gender, religion or sexual orientation interplay and lead to or reinforce discrimination.

Recruiting, with or without technological aids, can be discriminatory. Against this background, algorithmic recruitment systems are sometimes mistakenly believed to lead to more ‘objective’ decisions that are free from stereotypes and preconceptions about an individual’s personal characteristics.¹ From this perspective, there is a misplaced hope that these systems could be less susceptible to bias than human recruiters. Instead, basically all applications of AI – including those in hiring – have been found to involve risks of discrimination.² The data used to train algorithms reflect pre-existing societal and organisational biases, which are reproduced and often amplified by algorithmic systems. Developers’ and deployers’ choices about data, models, criteria and how the system is used in general, have an impact on it, which means that recruitment systems cannot simply be regarded as neutral and objective and require close oversight by HR professionals once deployed. Lastly, recruiters and candidates can use these tools in ways that were not originally intended and their behaviour cannot be fully anticipated. This makes the real-world impact of these systems difficult to predict.

Yet, because of the black box nature of many hiring algorithms, while it is often hard to tell when, how and why such discrimination happens, one thing is clear: discrimination is a lose-lose scenario. First, the individual or groups affected have one of their most fundamental rights violated and are less favourably treated (based on protected characteristics). Second, there are high costs to companies: not only reputational risks but also the possibility of missing talented and suitable candidates if AI tools introduce biases. This might also affect the diversity in the company’s workforce as well as its performance. And finally, discriminatory systems don’t just harm users – they can also backfire on the companies that developed them, damaging their credibility and trust. Fighting against intersectional discrimination can help create more inclusive and equitable hiring processes, attract a wider talent pool and ultimately build a more innovative and successful workforce. Reducing algorithmic discrimination in recruitment is therefore crucial on legal, ethical, social and economic grounds. This document tackles some of the technical challenges and dilemmas in detecting and mitigating bias and discriminatory effects of algorithmic hiring tools, but there are also challenges in countering algorithmic discrimination at the socio-political and legal levels. At the same time, there is also a growing set of procedures and tools to detect and mitigate biases. Within FINDHR we have developed methods, algorithms and training that enable developing, implementing and maintaining algorithmic hiring tools to be less discriminatory – with a special focus on mitigating intersectional discrimination.

If you are a software developer or product manager working on algorithmic hiring tools, countering discrimination in these systems starts with you! By designing more inclusive, transparent and accessible hiring systems you are taking crucial steps towards making hiring more fair.

We hope that this Toolkit will motivate you to take them and provide you with concrete ideas on where to start.

Find out what you can do to fight algorithmic discrimination in hiring – and act!