Bot For Hire, Ethics Of AI In Recruitment

AI has now arrived in recruitment and HR, from automatically filtering job applicants to conducting interviews. This raises many new ethical questions. Some thoughts after the symposium on “The use and ethics of AI” at Greenwich University that took place yesterday. This article does not aim to be a summary of the symposium, just some thoughts along the way.

A quick flashback first: A long time ago when I was in school we were taught how to write job applications. It was kind of cute, because young as we were we had little to tell about ourselves, so most of the focus was on the formal correctness (those were typed on a typewriter, some even handwritten), as well as training for job interviews. I do not remember that aspects like discrimination or privacy were brought up back then, we were expected to answer any question, and certainly not told to question the recruiters decision.

Times have changed (not only have a grown much older and have not written a proper job application myself in a long time), now I am someone conducting job interviews occasionally, though always in a startup context and never with the need to use automated or even AI powered tools.

New technology, more challenges

Recruitment is and has always been plagued by bias and discrimination: gender, ethnic and age bias, to name some. Luckily since the time I went to school, a lot has been done both legally and in awareness to guarantee applicants a fair and equal treatment, to various success.

Now new challenges arise: In the endless strive for “efficiency” (I will come back to that) we now have automated pre-screening systems, AI powered or AI analysed job interviews.

• The question remains who make the assessment if an AI tool should be used at all in HR, what the acceptable error rate is, and if companies silently accept a “rate of discrimination”
• Have the automated systems ever been double-blind tested against human-only recruitment? This would require setting up two HR departments in parallel dealing with the same applicants.
• If there are any studies about how many people stop the application process because they do jot want to work for a company that uses AI in that way, and if those applicants may be some of the best. This would hint at a form of survivor bias when analysing HR systems, as those who dropped out in the first place are not included in evaluation.
• Students, applicants trying to game the system with help of AI, ChatGPT or even faked personas.
• While applicants have legal rights to contest decisions they perceive as discriminating, how is that possible if AI systems are used for deniability or the decision process can not be reconstructed in detail?
• If pre-filtering large number of applicants is based on non-discriminating aspects like experience, education and skill sets, is AI even the correct tool to use? Would a deterministic system not perform better and more transparently. As one audience member pointed out, LLMs are unsuitable if the context (the job applications, usually a few pages of text) is small.
• The use of AI to analyse an applicant’s performance in a job interview, via facial or emotion recognition, raises question of privacy intrusion, possibly leading companies to evade the banned use of “polygraph-like” devices by AI.¹

Ethical dilemma

While researchers have developed a good grasp of bias in automated decision-making, all ethical questions fall flat if the motivation behind the use of AI in recruitment and HR is in fact simply to find the “best applicants for the lowest cost” and evade accountability with a layer of deniability. It raises the question of who in a business has a real interest in a fair recruitment process:

The recruiter won’t care as the AI provides deniability (“Well, the AI told us..”), middle management doesn’t care as long as cost in down, and upper management levels mostly will only take interest in very senior roles, where no AI tools are involved in the recruitment process.

As Thomas Ferretti put it: Balanced use of AI is not going to come from the private sector. The incentive is not there.

What may happen though, as Guido Conaldi expressed, is that the system collapses in an arms race between applicants that game the system with help of AI and an extremely complex recruitment system that takes too much data into account and leads to no improvements in efficiency.

Regulation

Using automated systems in recruitment does not relieve employers from regulation and laws against discrimination. In fact, it adds a second layer employers have to ensure to be fair. A recent class action lawsuit against Workday², one that could affect “hundreds of millions of people” who were rejected for employment through Workday shows that the fallout of a malfunctioning system could be gigantic.

While not applicable directly in the United Kingdom, most analysts agree that recruitment, pre-filtering applicants and HR systems overall fall into the “high-risk category” of the EU Ai Act³, so employers and system providers better look carefully into the resulting obligations.

In how far the newly updated EU Product Liability Directive (PLD), which extends liability to psychological harm caused by AI services, applies in this context will be interesting.

The broad scopes of both the EU AI Act and the PLD mean that EU regulation would certainly apply for companies recruiting from the EU, as it seeks to protect EU citizens independently of the employers' location.

However, Article 22 of the GDPR, which requires that everyone has the right not to be subject to a decision based solely on automated processing, including profiling that produces legal or similarly significant effects on applicants, certainly applies in the UK (for now).

If AI in recruitment reduces the chances for discrimination is at least questionable, certainly it’s not proven. From an Ethics point of view “efficiency gains” should not be a consideration. At minimum, we are adding another layer to introduce bias and privacy issues, a layer more complex than the human one and harder to understand.

Special thanks to Guido Conaldi, Thomas Ferretti and Haining Wang for the inspiring talk. And to anyone else who made the symposium possible.

¹ The use of AI for emotion recognition and quasi polygraph testing, and connected pseudo-science, is something I will explore in a future article

² Workday AI lawsuit receives the greenlight to proceed as a collective action — Jesika Silva Blanco

³ The EU AI Act from an HR perspective — TaylorWessing

Article image by Wolfgang Hauptfleisch (with help of Microsoft AI Image Creator, heavily modified)