AI-powered recruiting isn’t always a smart move

One of the trends that the covid-19 pandemic promoted in the world of work is the recruitment with artificial intelligence (AI). As has happened with technological advances in each era, the opinion about them can go towards suspicion or towards total dedication. But there are several points to take into account before making a decision.

Some algorithmic personality tests cannot be considered as “valid testing instruments” to decide whether or not to hire a person, says a study from the University of New York (UNY), in which a team of researchers analyzed a couple of of platforms and found basic flaws.

According to Atento, a company dedicated to the outsourcing of various care services, “eight out of 10 candidates interviewed through solutions based on Artificial Intelligence (AI) are approved to fill the requested vacancies, which represents an increase of 30% in the rate of approval compared to the traditional method”.

For Miranda Bogen, Head of AI Policy at Meta, it is expected that “algorithms help humans avoid their own biases by adding consistency to the hiring process. But the algorithms introduce new risks of their own”, he wrote in the Harvard Business Review, a magazine of Harvard University.

The UNY team analyzed the Humantic AI and Crystal platforms and concluded that “exhibit instability in the rank order with respect to the context of the source and the time of the participant”.

The results were published in the academic article Resume Format, LinkedIn URLs, and Other Unexpected Influences on Hiring AI Personality Prediction: Audit Results.

evidence is lacking

According to the study, “Crystal frequently calculates different personality scores if the same resume is provided in PDF and raw text, violating the assumption that the result of an algorithmic personality test is stable across irrelevant input variations for the job”.

We found “evidence of persistent data linkage, and often wrongby Humantic AI”.

Crystal’s text prediction tool “is deterministic,” the results for Humantic AI “were not perfectly reproducible, even immediately,” the report says. That could explain why they found score changes in several participants, that is, in people who wanted the job for which they applied.

Humantic AI uses information from LinkedIn profiles. However, “we discovered a substantial instability” of time, because sometimes have not been updated and they do not match the professional background and experience of job applicants.

“It is essential to obtain more evidence of domain-specific validity to support the use of algorithmic testings of personality in hiring,” say the authors and the authors of the UNY publication.

Research is by Alene Rhea, Kelsey Markey, Lauren D’arinzo, Hilke Schellmann, Mona Sloane, Pablo Escuderos, and Julia Stoyanovich. They are specialists in computer science and data science, sociology, industrial psychology and investigative journalism.

“Our methodology can be used by employers to make informed purchase and use decisions, by legislators to guide regulation, and by job seekers to make informed decisions about disclosing their information to prospective employers,” they say in the document.

We cannot leave everything to technology

AI recruiting platforms “can replicate institutional biases and historical, amplifying the disadvantages” such as schooling or performance evaluation, says Miranda Bogen.

“Even if algorithms remove some subjectivity from the hiring process, humans are still heavily involved in final hiring decisions.” For example, they can deny that the apps or platforms are wrong.

In research with Aaron Rieke, Bogen found “that most hiring algorithms skew toward bias by default.” The results of that study were published in the report Help wanted. An examination of hiring algorithms, fairness and bias.

The potential of the platforms should not be ruled out precisely to avoid the interpersonal bias of those in charge of recruiting, he says. But these tools must be created with the goal, he adds, of “proactively addressing the deepest disparities” to promote equity, “rather than erode it.”