AI in Recruiting. Multi-agent Systems Architecture for Ethical and Legal Auditing

María del Carmen Fernández Martínez, Alberto FernándezIJCAI 2019 · DOI

TL;DR

This paper proposes a novel multi-agent system architecture designed to ethically and legally audit AI-driven recruitment tools, particularly those analyzing video interviews. By automating parts of this auditing process, the authors aim to ensure fairness, accuracy, and reduce potential discrimination in the job market.

What Problem Does This Paper Address?

The increasing adoption of AI in recruitment, especially for analyzing video interviews, presents significant ethical and legal challenges. While AI can streamline processes like resume scanning, video analysis introduces new risks. These include potential biases embedded in datasets, leading to unfair pre-selection based on characteristics like age, gender, or sexual orientation, which can have serious legal consequences in many jurisdictions. Furthermore, the lack of comprehensive regulation for these AI systems necessitates robust, neutral auditing mechanisms to ensure their responsible deployment.

Methods

The core contribution of this paper is the proposal of a Multi-agent System (MAS) architecture for auditing HR processes. This architecture aims to automate parts of the legal and ethical auditing of AI recruitment tools. The system is designed with distributed agents, each potentially responsible for specific tasks related to auditing. While the paper focuses on the architectural proposal, it also mentions the creation of a prototype for a "legal rules engine" and the definition of a common lexicon. The approach involves formalizing legal rules into a computable format, with approximately 50 current legal rules considered so far. The integration of ontologies is also highlighted as a key aspect for domain specification and interoperability within the MAS.

Datasets

Not discussed in the paper. The authors mention relying on prototyping with simulated corporate scenarios due to the lack of accessible real-world data for testing.

Key Insights and Ideas

  • The Need for Proactive Auditing: The paper strongly emphasizes that the rapid adoption of AI in recruitment, particularly video analysis, outpaces regulatory frameworks, making proactive and automated auditing crucial.
  • Multi-Agent Systems as a Solution: A MAS architecture is proposed as a flexible and distributed approach to tackle the complexity of auditing diverse AI recruitment tools.
  • Formalizing Legal Rules is Key: The authors recognize that translating complex legal statutes into a formal, machine-readable format is a significant but necessary step for automated legal auditing.
  • Bias is Pervasive: The paper highlights how AI algorithms can inherit biases from training data (e.g., Affectiva's emotion dataset) and how even seemingly neutral features can be used for discriminatory purposes.
  • Ethical Implications Beyond Bias: Beyond bias, the automation of HR processes can strip away human elements, reducing candidates to mere descriptors, which raises broader ethical concerns.

Key Results

The paper primarily presents an architectural proposal and a conceptual framework. The key "result" is the design of the Multi-Agent System architecture for auditing HR, illustrated in Figure 1. Prototyping of a "legal rules engine" has also been initiated. The authors have begun formalizing approximately 50 legal rules.

Limitations and Future Work

The authors acknowledge that they are currently in the implementation and prototyping phase, focusing on defining agent flows and interactions. A significant challenge is the potential use of different ontologies across agents, which could hinder interoperability. The paper also notes the difficulty of fully formalizing labor law into a set of rules, especially if the number of rules becomes very large. Future work will involve the complete implementation and testing of this complex system, addressing the challenges of formalizing labor law and potentially the detailed architectural design for a PhD dissertation. The authors also point out the limitation of not having access to affordable and limited testing opportunities of real corporate applications in their academic setting, necessitating the use of simulated scenarios.

Summary updated

BibTeX
@inproceedings{martinez2019ai,
  author = {María del Carmen Fernández Martínez and Alberto Fernández},
  title = {{AI in Recruiting. Multi-agent Systems Architecture for Ethical and Legal Auditing}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/903},
}
Figure