Hyperbole isn’t helpful

Headlines like this are one of the reasons there is so much hyperbole about AI in hiring versus sensible, measured debate.

'Largest study of AI hiring algorithms to date finds clear racial disparities, over 25% of Black applicants tainted by bias'

The study reviews recommendation ratios from a single vendor. Regardless of sample size, and regardless of how concerning the findings might be, that is not a 'large study' of AI hiring algorithms. It is a study of a single vendor.

The 'AI in hiring' label flattens important distinctions. To my knowledge, the vendor was an early provider of games-based assessment, built on machine learning models trained against incumbent performance data, with human review of final algorithms. Machine learning is recognised as a sub-set of AI but it is categorically different from the LLM-driven AI interviewers calling candidates and scoring answers that most people picture when they hear 'AI in hiring'. This study focused on a very specific implementation of AI in hiring, not the multitude of applications of AI in hiring.

The central point of the study, behind the headline, is hugely important as it calls out 'algorithmic monocultures' - the idea that the algorithms barely vary across employers, meaning the same applicant is very likely to be rejected from any role that is being assessed by a single vendor. However, it is perhaps reductive to frame this as a problem unique to AI in hiring. 'Traditional' hiring methods such as cognitive ability tests, personality assessments, even structured interviews are often keyed to reward the same constructs across employers. We've just never been able to systematically measure it. The same old majority group behaviours being rewarded in hiring and in-job performance is wider problem to solve, the hiring mechanism is simply the delivery.

As for whether the bias audits discussed in the study are adequate or not, you will have to join me and for our webinar at 3.30pm on 18 June to hear us unpack the myth of 'bias-free' assessment. Link to register in comments. Link to article and the original paper also in comments.

Register here: https://zoom.us/meeting/register/8AQCVBvCRUyV_tdas9TfdA