AI & FAIRNESS

Facial Recognition & Recruitment

While the use of facial recognition technology could help recruiters streamline some processes, it is hard to believe that the long-standing issues of recruiting could be solved through algorithms knowing that existing datasets could perpetuate or even worsen the current biases that a human recruiter can have.

By Gabriel Obando-Chacon & Yannick Stadtfeld & Betül Çolak

August 27, 2021

Imagine you are applying for your dream job at your dream company. You write a cover letter to explain why you are the perfect fit for the job, polish up your resume and line up people to be references. You submit your application and some days later you receive an interview invitation. But you are not interviewed by a recruiter, you have to use the camera of your smartphone or laptop to record your answers to several questions. Your answers will be assessed by an algorithm that judges your voice, your body posture, your enthusiasm, or some other feature that you do not know about. This may sound futuristic but depending on where you live its reality.

Facial recognition technology (“FRT”) is increasingly used in the recruitment processes, particularly with the rise of video interviews during the Covid-19 pandemic by both public and private sector. Around 700 companies, including Vodafone, Hilton and Urban Outfitters have tried it out. For example, Unilever uses AI systems both taking job applications and making pre-interview. So FRT is mainly used for candidate screening. If an applicant would be successful then a human recruiter has an interview with the candidate. 

The market for recruiting-technology and hiring-solutions is growing exponentially (Source 1, Source 2), as many companies try to leverage predictive analytics in their recruiting process. Countless start-ups and corporations compete for this market. They advertise their solutions as fair and ethical, but many lack transparency of their systems and independent audits. 

Standards for recruiting

There are several standards for psychological and aptitude testing available which can inform the recruitment process, but none sets definite criteria. The ISO norm 10667 is as close as it comes to a global standard. It defines “Procedures and methods to assess people in work and organizational settings”. It formulates requirements and guidelines for tests and methods of job-related aptitude assessment of individuals, groups, and organizations. The norm demands an evidence-based approach underpinning the decision for the chosen aptitude diagnostic procedure and the procedure needs to be checked regularly for their purpose. It also discusses questions about the combination of data and the decision rules. (The second edition of this norm which was released last year explicitly includes AI systems but has, to our knowledge, not been used to audit an AI system up to now.) 

No organization is forced to comply with these standards, which is also evident even if no AI systems are used. Some organizations make obscure hiring decisions, which leads to discontent applicants and the organization misses out on recruiting the best person for the job.  The increasing adaptation of facial recognition in recruitment leads to concerns that this lack of standardization and best practices will be built into FRT systems used for recruiting.

The Implementation and Concerns of The Use of Facial Recognition in Recruitment 

Thanks to FRT, many employers save their time, eliminate repetitive tasks and find the most appropriate candidates. However, FRT may have bias in its data sets. The reasons behind that can depend on data sets, developers or the failure of AI systems to take account of changing circumstances. In the system, race, gender and age-related bias can lead to discrimination and unfair consequences among candidates regardless of their cause or type. This poses a great risk for employers to violate the obligations imposed under human rights and equality legislation. On the other hand, a human employer has bias against candidates too. For instance, black workers are twice as likely to be unemployed as white workers overall (6.4% vs. 3.1%) in the US. 

Furthermore, FRT is not always unambiguous. The use of complex AI systems in FRT puts it at a disadvantage in terms of transparency or interpretability. Especially in FRTs where deep neural network algorithms are used, which are the main actors of the black box problem, it is generally not possible to explain what cause bias, injustice and discrimination. However, one of the important components of ethical AI is the “explainability”. This is mentioned also in the draft Regulation on artificial intelligence of the European Commission by stating that the processes of high-risk artificial intelligence systems should be transparent (Art. 13). Failure of companies to explain the problems arising from FRT will cause them to lose their reputation in the face of the importance attached to the issue.

Another issue about FRT is whether the candidates can adapt their actions regarding the criteria. The technology analyzes things like keywords, intonation, and body language, and makes notes on them for the hiring manager.  It is an effective way to find an ideal candidate. However, that may lead to the rejection of talented and innovative people. For example, a candidate who does not smile at the right moment or cannot adjust his/her tone of voice correctly for a moment may be eliminated by the system. Well, can candidates adapt their behavior to avoid being eliminated by the system? Considering that the systems used in the queries are quite less likely to be mistaken, it can be said that it is not possible for the candidates to deceive the system.

The added value question

Now that we have analyzed the pros and cons of implementing FRT as a tool in personnel selection, we should take a look at its added value. Just because the technology exists doesn’t mean that it will necessarily improve the HR recruitment process. The bottom line is: what do recruiters gain by using FRT? Like every cutting-edge technology, it’s not easy to answer this question right now, as its deployment is limited, and examples are scarce. Nevertheless, in an attempt to answer this question, we can make some broad inferences based on its current uses. 

FRT systems will likely create added value for recruiters that (i) have the capacity to determine exactly what they want to assess through facial recognition and (ii) have the means to collect data sets and tailor the algorithms to their specific needs. On one hand, recruiters have to know what they are looking for with the use of FRT technologies. They need to determine whether they want to assess the behavior of the prospective employee in a hypothetical scenario or whether they want to assess the degree of truthfulness of the responses, or any other output, otherwise the exercise becomes futile. The algorithm design process will force recruiters to think about what they want as output, to program the systems to obtain the desired result. 

On the other hand, recruiters must have the capacity to work hand in hand with engineers to build and tailor the algorithms around the existing data sets. Most recruiters do not collect biometric information (in some jurisdictions it could even be illegal) that is required to run a facial recognition algorithm, but those who can collect these data sets still need to evaluate if the information is enough to meet the desired output. This will require assembling a team of experts, including data analysts and engineers that gap the data with the desired output. 

Taking into account the above, it is not surprising that Big Tech has taken the forefront in the use of FRT systems for recruitment purposes. For example, Google and Amazon have in-house engineering teams that can work hand-in-hand with the HR departments and are able to make adjustments on the fly. On the contrary, medium-sized and small recruiters will likely not turn to facial FRT any time soon.

Wrapping it up!

In conclusion, facial recognition is making its way into recruitment processes. It is likely that recruiters turn to collecting more information (including biometric data) from prospective employees to use it in the design, building and training of facial recognition algorithms in the future. While the use of these technologies sounds exciting and could help recruiters streamline some processes, we are wary of some of the implications of using this type of technology without the proper standards of care. It is hard to believe that the long-standing issues of recruiting could be solved through algorithms knowing that existing datasets could perpetuate or even worsen the current biases that a human recruiter can have.

Nevertheless, the use of FRT is at the same time forcing recruiters to become aware of their flaws and follow standards for profound assessment methods which have long been in existence. While it is still early, there is no reason not to see facial recognition as an opportunity to ensure fair and equal recruitment processes for employees in the future. 

Yannick completed his bachelor’s degree in Psychology at Trier University. Currently he is pursuing his master’s degree in Psychology (Human Performance in Socio-Technical-Systems) at the Technical University of Dresden with a focus on Human Factors and Economic Psychology. As a research assistant in various research groups he gained experience in the application of qualitative and quantitative research methods. Yannick interned at the German Aerospace Center (DLR) where he conducted research on group behavior in aerospace teams. He is currently interested in the application of computational social sciences methods.

Gabriel recently graduated from McGill University, where he pursued an LLM. His research focused on privacy concerns related to the collection of data by governmental agencies to implement judicial AI. Before that, he obtained an LLL from the University of Costa Rica with a concentration in Human Rights Law. He has worked in BigLaw and boutique firms in Costa Rica. Most recently, he joined Nelson Champagne Avocats in Montreal where he's part of the class actions team. He's fluent in English and Spanish, and proficient in French.

Betül Çolak holds a bachelor’s degree in Law from the University of Bahçeşehir. She, also, studied law at University of Jean Monnet in France within the Erasmus program for a semester. She worked as an intern on data protection law, immigration law and the corporate law in the United States and in Turkey. She is a lawyer specialized on data protection law and intellectual property law in Turkey. She is a member of the Artificial Intelligence Working Group at Istanbul Bar Association. She writes news and articles on data protection regulation and AI. Her academic interest lies in the regulation of data-driven technologies, in particular, the challenges they impose to democracy, data privacy and human right. She is fluent in Turkish, English and can communicate in French.