An external stability audit framework to test the validity of personality prediction in AI hiring

Abstract

Automated hiring systems are among the fastest-developing of all high-stakes AI systems. Among these are algorithmic personality tests that use insights from psychometric testing, and promise to surface personality traits indicative of future success based on job seekers’ resumes or social media profiles. We interrogate the validity of such systems using stability of the outputs they produce, noting that reliability is a necessary, but not a sufficient, condition for validity. Crucially, rather than challenging or affirming the assumptions made in psychometric testing — that personality is a meaningful and measurable construct, and that personality traits are indicative of future success on the job — we frame our audit methodology around testing the underlying assumptions made by the vendors of the algorithmic personality tests themselves. Our main contribution is the development of a socio-technical framework for auditing the stability of algorithmic systems. This contribution is supplemented with an open-source software library that implements the technical components of the audit, and can be used to conduct similar stability audits of algorithmic systems. We instantiate our framework with the audit of two real-world personality prediction systems, namely, Humantic AI and Crystal. The application of our audit framework demonstrates that both these systems show substantial instability with respect to key facets of measurement, and hence cannot be considered valid testing instruments.

Auditing of AI: Legal, Ethical and Technical Approaches

“…Others have developed auditing procedures for more targeted purposes, e.g., to test the accuracy of personality prediction in AI systems used for recruitment (Rhea et al, 2022), evaluating the capabilities of language models (Goel et al, 2021;Mökander et al, 2023), providing explanations for black-box AI systems (Pedreschi et al, 2018), and conducting audits of clinical decision support systems (Panigutti et al, 2021). Again, what links all these procedures is that they audit AI systems against predefined technical, functionality and reliability standards.…”

Gender equity in hiring: examining the effectiveness of a personality-based algorithm

“…Recent research has examined the accuracy of personality prediction in AI-based hiring systems and found that certain tools demonstrate significant instability in measuring key facets. Consequently, these tools cannot be considered valid assessment instruments ( Rhea et al, 2022 ). However, it is still uncertain whether alternative personality-based hiring algorithms, designed to predict job performance based on personality facets, could potentially result in adverse impacts or biases.…”

Examining the assumptions of AI hiring assessments and their impact on job seekers’ autonomy over self-representation

“…It is certainly true that the types of stimuli, measurements, and the technical methods of mapping those measurements to knowledge in the form of assessment scores can be very different when comparing traditional assessments and AI-based assessments (Liem et al 2018). To be clear, here we want to focus on the underlying epistemological and ontological assumptions of the psychometric paradigm, which are shared by both traditional and algorithmic hiring assessments (Rhea et al 2022). This shared set of underlying assumptions implies that their merits and limitations-which have been explored in academic literature over multiple decades-also apply to algorithmic hiring assessments.…”

“…Some industrial and organizational (I-O) psychologists have expressed concerns about the scarcity of available evidence supporting the validity, reliability, and fairness of these tools (Gonzalez et al 2019;Tippins et al 2021). In fact, a recent study auditing two personalityassessing algorithms used in hiring concluded that both tools failed to exhibit sufficient reliability and therefore cannot be considered as valid assessments (Rhea et al 2022). Furthermore, some algorithmic assessments measure features for which there is no established and scrutinized theory relating them to job seeker attributes or job performance (e.g., features like tone of voice and facial expressions) (Ajunwa 2021;Hinkle 2021;Stark and Hutson 2021;Tippins et al 2021).…”