BLOOMINGTON, Ind. — Job performance evaluations can be stressful. You’ve put in your time, you’ve worked hard for the company, and now comes the analysis, one that could decide what kind of vacation you’ll be able to afford or whether you’ll go on one at all.
But, more and more, there is a question of who, or what, should be doing the evaluation to begin with: your boss or artificial intelligence.
In a new paper, “Using Artificial Intelligence to Evaluate Employees: The Effects on Recruitment, Effort, and Retention,” three professors at the Indiana University Kelley School of Business sought to assess reaction to the knowledge of whom or what will be evaluating job performance.
“AI is an emerging technology that’s really being used in business — in performance evaluations — today. And we don’t really understand how people perceive the information that they’re getting from those evaluation systems” said Ashley Sauciuc, assistant professor of accounting at the Kelley School. “It’s incredibly important to understand how people are internalizing and perceiving the feedback that they’re getting from this AI technology and how that might actually affect real business decisions, since perceptions influence real action.”
In the first of two experiments, 700 participants were given a scenario in which they could leave a job for a position at another firm with $10,000 less in salary but also an opportunity to earn up to $20,000 more in performance-based bonuses. They were either told their annual review may be conducted by a senior manager or handled by an artificial intelligence system. A second experiment involved 539 participant observations and tested the interaction between evaluator type and demographic group on employee effort and retention.
One of the key findings of the paper was that employees from underrepresented minority groups provided greater effort when a firm used AI instead of a traditional manager for employee evaluations. In contrast, among white males, the opposite was found to be true.
“We found effort was higher for underrepresented groups under AI versus a traditional evaluation, but we found little difference for more traditional employees,” said Jason Brown, associate professor of accounting. “So that leveling of the playing field I think was important because you’re able to let the AI give you a boost for your underrepresented employees that puts them on equal footing as your more traditionally represented employees.”
When a traditional manager is in place for evaluations, there is a gap between underrepresented minority groups and traditionally represented employees when it comes to their attraction to the company, but when AI is in place employees are equally attracted to the company, regardless of demographic group, said Joseph Burke, an assistant professor of accounting at Kelley.
“How do we know that the research results reflect a greater trust in AI and not a concern of needing to overcome the failings of it? Our models tell us that the difference in employee behavior between that of traditional managers versus AI is driven by expected biases,” Burke said.
The researchers also found that, regardless of demographics, employees hold a general belief that AI is less biased.
“The general consensus among practitioners and academics is that AI systems can be more or less biased, based on the parameters in which they were set up. But we’re not really speaking to actual bias,” Burke said concerning those who think AI is biased when it comes to evaluations. “We are studying employees’ perceptions of bias. An underlying assumption in our research is that when firms do use AI, it will be implemented using best practices to minimize any sort of actual bias in their systems.”
Collectively, the paper says its results provide important cost-benefit information for firms considering the use of AI within their performance evaluation system in terms of the firm’s attractiveness to potential employees and the effort they will receive from under-represented minority groups versus more traditionally represented employees.
The researchers intend to do more in-depth field study to see if the perceptions actually result in real effort changes or real recruitment changes. Burke thinks it could be an expansion into the effects of different implementation parameters.
“Our study focused on how employees perceive and react to AI-driven performance evaluations. However, our experiment provided relatively little detail regarding the specific parameters of the AI system,” Burke said. “I think the next step is taking a deeper dive into how employee responses to AI-driven systems may depend on the specific parameters used to develop, train, and/or implement the AI system.”