TLDR: A study involving 410 German HR managers investigated how their AI literacy affects their interpretation of Explainable AI (XAI) features in recruitment dashboards. While XAI improved subjective perceptions of trust and helpfulness for users with moderate to high AI literacy, it did not consistently improve objective understanding. Complex explanations like ‘counterfactuals’ and ‘model criteria’ often hindered understanding, especially for less literate users, and even high-literacy users showed signs of overconfidence. The research highlights that XAI is not a universal solution and emphasizes the critical need for tailored explanation strategies and targeted AI literacy training in Human Resource Management.
Artificial intelligence (AI) is rapidly transforming the landscape of human resource management (HRM), especially in recruitment. AI-based recommender systems are increasingly used to identify promising candidates, match skills to projects, and accelerate hiring processes. These tools promise to reduce human bias and ground decisions in objective data, aligning with trends like HR Analytics and People Analytics. However, the growing influence of AI in decisions about individuals also brings critical concerns about transparency, fairness, and responsible adoption, particularly given regulations like the EU AI Act and GDPR.
The Challenge of Understanding AI in HR
Despite the benefits, AI systems in recruitment often operate as ‘black boxes,’ making their decisions opaque. This lack of transparency can lead to issues, especially when HR managers, who typically have non-technical backgrounds, need to make legally and ethically sound decisions. Explainable AI (XAI) aims to address this by making AI models interpretable, providing insights into why a candidate is ranked highly or why certain competencies are flagged. However, simply adding XAI features to a dashboard doesn’t guarantee understanding or effective use.
AI Literacy: The Missing Link
A crucial factor in the effectiveness of XAI is the user’s ‘AI literacy’ – their knowledge, skills, and attitudes for understanding and assessing AI systems. This study, titled “Explained, yet misunderstood: How AI Literacy shapes HR Managers’ interpretation of User Interfaces in Recruiting Recommender Systems” by Yannick Kalff and Katharina Simbeck, investigates how HR managers’ AI literacy influences both their subjective perception and objective understanding of XAI elements in recruitment dashboards. The researchers hypothesized that AI literacy would directly affect how helpful, trustworthy, and accurately understood XAI explanations are.
The Experiment: Testing XAI with HR Managers
To explore this, an online experiment was conducted with 410 German-based HR managers. Participants compared a baseline AI dashboard (without explanations) to versions enriched with three different XAI styles: ‘important features’ (showing key influencing factors), ‘counterfactuals’ (what-if scenarios), and ‘model criteria’ (general rules the model uses). The study measured participants’ subjective perceptions (trust, usability, assessment quality) and objective understanding (correctly interpreting factual statements about the dashboard’s output).
Subjective Perceptions: XAI Improves Trust, But Not for Everyone
The findings revealed that adding any of the three XAI components generally led to a modest increase in users’ subjective perception of the dashboard’s quality. However, this positive effect was primarily concentrated among HR managers with at least moderate levels of AI literacy. Low-literacy users saw only slight, non-significant increases in perceived usefulness, trust, or transparency. This suggests that a foundational understanding of AI is necessary for users to even subjectively appreciate the value of XAI explanations.
Objective Understanding: A Surprising Paradox
The impact of XAI on objective understanding was more complex and, at times, paradoxical. Only the ‘important features’ explanation style consistently improved objective understanding, and this benefit was almost exclusively for high-literacy users. For low and medium-literacy users, ‘important features’ had negligible effects.
Counterfactual explanations, which present ‘if-then’ scenarios, actually led to a statistically significant *decline* in objective understanding for low-literacy participants, suggesting they found this information confusing. Medium-literacy users showed a non-significant downward trend, while high-literacy users remained largely unchanged.
Most strikingly, ‘model criteria’ explanations, which detailed the AI’s decision rules, proved counterproductive across the board. Interpretation scores fell significantly for low, medium, and high-literacy users alike, indicating that this type of detailed explanation overwhelmed users regardless of their AI background, leading to poorer objective understanding.
The study also uncovered a potential ‘overconfidence effect’ among high-literacy users. While they were more likely to benefit subjectively from explanations, their absolute objective understanding was sometimes lower than expected, hinting that self-assessed AI literacy might not always align with actual comprehension.
Also Read:
- Human-like AI Features Influence Generative AI Adoption Among Chinese University Students
- Navigating AI Maturity: A New Framework for Responsible Engagement
Implications for HR and AI Design
These findings challenge the assumption that XAI elements are universally beneficial. The study concludes that XAI is not a one-size-fits-all remedy for making AI transparent. Its impact is conditional on users’ AI literacy, the specific type of explanation provided, and the context of use. For HR, this means:
- Explanation strategies must be carefully tailored to users’ actual literacy and cognitive needs.
- Investments in AI for HR should be coupled with targeted, domain-specific AI literacy initiatives.
- Designers should avoid overly complex explanations that can overwhelm or mislead users, even those with higher AI literacy.
The research emphasizes that effective and responsible deployment of AI in HR requires an integrated approach: robust, context-sensitive literacy training, user-centered explanation design, and ongoing evaluation of both subjective perception and objective understanding. For a deeper dive into the methodology and detailed findings, you can read the full research paper here.
