There is no doubt that AI presents organizations with an efficient way to streamline the hiring process. However, efficiency should not come at the expense of fairness. In fact, ethical concerns posed by the use of AI technologies in recruitment contexts have already been raised in law through the New York City Automated Employment Decision Tool Law (NYC AEDT) and the impending EU AI Act [5].
Let’s look at some of the commonly AI-enabled tasks in the recruitment process and the opportunities and risks presented from the perspective of social mobility.
Candidate Sourcing
AI tools can assist in finding candidates by scanning numerous online platforms, job boards, and social media profiles to identify individuals who match the desired criteria for a posting. In addition, generative AI, that is, the use of AI to produce new content, including text, images, audio, and even video, can be used to draft job postings.
The opportunities for social mobility:
- Recruiters can reach a greater potential pool of applicants than what manual sourcing methods permit for, which may provide individuals without access to a professional network greater visibility towards recruiters.
- Tone meters in combination with generative AI can “make suggestions on how to improve the inclusiveness of the language used” in job postings prior to publication [6].
- As generative AI capabilities mature, personalisation in job-matching is also likely to mature. This could enable opportunities for candidates to query what jobs best fit their CV’s, where gaps in their experience may be, and suggested next steps to be a competitive applicant. Increased transparency and personalised feedback can assist young people in navigating the recruitment process and understand pathways to get to the role that they want.
The risks for social mobility:
- Depending on the parameters of the AI model used in candidate sourcing, such tools may algorithmically discriminate against individuals based on personal characteristics (e.g., race, gender, socio-economic status) or data that serves as a proxy for such characteristics (as explicit gender and race data is not typically included in model design). As an example of a proxy variable, consider the use of postal or zip codes in AI-assisted credit assessments for a bank. Even if race data is not considered by the model, because “zip code is strongly correlated with race” this datapoint ends up becoming a proxy for race [7].
CV Screening
After sourcing, RMS can be leveraged to automatically screen CV’s and identify key qualifications, skills, and experiences. Using natural language processing (NLP), an AI technique, these systems can analyse the content of CV’s, match keywords, and rank candidates based on their suitability for specific roles. This can save recruiters significant time and effort by quickly narrowing down the pool of applicants for a given job posting. It is often at this stage that automated or semi-automated decision-making to progress with a candidate further is made.
The opportunities for social mobility:
- We already know that hiring led by humans can be remarkably biased, which can disadvantage vulnerable or marginalised groups. If effectively governed, tested, and monitored, AI tools could be leveraged to mitigate human biases or variations in decision-making between multiple recruiters by using a consistent evaluation criterion for all candidates.
- There’s also concerns that recruitment has historically overemphasised credentials over skills. AI tools could open the door for individuals that may have relevant work experience but not a university degree by focusing more on tangible skills [8]. This could assist young people in finding work that face financial barriers in attending post-secondary education and avoid biases that individual recruiters may have favouring specific academic institutions.
The risks for social mobility:
- Despite consistent evaluation criterion, it is important to remember that AI models do not represent objective truths about people, our society, and our value. AI models are subjective by nature, as they are created by humans that make design choices that impact the overall fairness of a tool (e.g., via feature selection, whereby the data that is to be considered by a model is selected). The data upon which models are trained can be unrepresentative or contain historical biases that affect how particular groups are treated (e.g., disproportionately selecting male candidates for managerial roles). Captured by the adage of ‘garbage in, garbage out,’ if the data on which an AI system is trained is biased or unrepresentative, it will lead to unfairness and inequality.
Video Interviewing and Assessments
For applicants that progress past the initial CV screening stage of recruitment, some companies are making use of AI-powered video interviewing platforms or recruitment games to further shortlist candidates. Oftentimes, AI interviewing tools analyse facial expressions, voice tone, and body language, in addition to what is said during an interview to provide insights on candidate behaviour and characteristics. Automated assessments and coding challenges are also used to evaluate candidates' technical skills and problem-solving abilities.
The opportunities for social mobility:
- AI interviews are often held without a human practitioner being present and, in some cases, will permit candidates to re-record responses. For those early on in their careers, the opportunity to practice responses and record multiple takes can lessen anxiety and improve candidate experience.
The risks for social mobility:
- The ways in which AI interviews are assessed may be opaque to users, particularly if employers use third party tools that they have not developed themselves. Given the asymmetric power dynamics that exist between employers and applicants, this can be discouraging, especially so for young people.
- Models that consider emotional affect, body language, and tone of voice may be discriminatory against particular groups, as what are seen as appropriate or desirable emotional responses may differ across cultures and genders. This also poses a particular concern for candidates that are neurodiverse and may not exhibit emotional responses or social behaviour in the same way as their neurotypical peers yet be perfectly capable of fulfilling work expectations.
