Guidelines for explainability and transparency
Participants in our Participatory Action Research (PAR) sessions ( → see chapter #3) frequently expressed frustration with the opacity of hiring processes. Candidates often do not know whether their application is being reviewed by a human or an algorithm, especially when online tools provide little or no explanation of whether and how decisions are made by an algorithm. On the other hand, recruiters often lack clarity on the criteria behind system recommendations or decisions. Last but not least, understanding algorithmic decision-making is also essential for auditors to ensure compliance with ethical, legal, non-discrimination and privacy standards. In summary, transparency is key for all stakeholders involved, to make AI hiring processes fairer and more comprehensible. It essentially starts with providing explanations on the decision-logic of a system.
To design an explainable system that supports diverse stakeholders in the recruitment process, two approaches are possible:
i) develop an explanation mechanism that generates post-hoc explanations from a recommendation or ranking system, making the system’s outcomes transparent; or
ii) create a recommendation or ranking system that is interpretable by design, where the algorithm inherently provides clear reasons and justifications for its behaviour and recommendations.
In either case, the explainable system should meet certain requirements (see recommendations below).
Explainability of algorithmic hiring systems and their outputs fosters stakeholder trust, enables verification of the algorithm’s expected behaviour, uncovers system vulnerabilities, supports legal compliance (e.g. GDPR transparency requirements) and helps to identify potential biases and patterns of discrimination. The Court of Justice of the EU1 emphasised the importance of transparently informing individuals on how their data influences decisions.
Recommendations:
To ensure explainability and transparency in the hiring system, it should:
- transparently inform candidates, recruiters and auditors about how the system is used, including what kind of data it uses and how, the way decisions are made and whether they are reviewed by a human. → see recommendations in chapter #6
- disclose the algorithm’s decision logic to enable external audits and ensure ethical compliance.
- provide tailored, actionable, motivating and fair explanations to each stakeholder (candidates, recruiters, auditors).
- ensure that explainability also accounts for intersectional fairness by analysing how bias may affect individuals at the intersection of multiple categories, not just isolated ones.
- help candidates improve their profile by providing suggestions based on counterfactual examples.
- support recruiters to improve job descriptions by suggesting keywords and skills based on successful past job postings.
→ see more on explainability in FINDHR Software Development Guide, section 2.6
→ see more on Multi-Stakeholder actionable interpretability in FINDHR Research Informed Technical Implementation Guidelines
- CJEU, Case C-203/22, Dun & Bradstreet.
