Algorithms in the workplace
Forget driverless cars. The real concern is when algorithms start making decisions about employees.
Psychologists who study the workplace have long recognised that the quirkiness of employment - where it is hard to measure performance, where tasks sometimes change quickly - is best managed if the employees care about doing a good job. This is unlike engaging a contractor where we can spell out exactly what we want them to do, what it should look like when they are done, and otherwise leave them alone.
To get employees to care about their job involves a lot of work on the part of the employer, with the most important aspects performed by our immediate supervisors. If we think our boss is looking out for us, if we think our boss is fair, then we are much more likely to look after our job in return. The aphorism that people quit bosses more than their job reflects this fact.
How do supervisors make that happen? They control things that matter to us: They hire us, they create our work schedules, they can give us a break when we are having problems. When the organisation has a problem, they ask us to return the favour: can we cover for the employee who is out today, can you handle this rush project, and so forth. As both researchers and experienced managers have long known, this exchange of favours is what makes the workplace function.
Now we come to the challenge and opportunity associated with algorithms. These are data science-driven formulas that are built around what has worked in the past, that tell us how to make decisions. The most popular algorithms right now include assessing the characteristics of job applicants that tell us which ones to hire, the tasks for which given employees are best suited, what work schedules will minimise labour time, and so forth. We don’t always make these decisions well right now, so there is a lot of room for improvement.
The challenge comes when the algorithms - rather than the supervisor - start making these workplace decisions. Consider, for example, the task of scheduling employees. In the past, if someone was scheduled to travel next week and had an important family conflict come up, they could go to their supervisor and at least get a sympathetic ear. If they were scheduled to work two weekends in a row, they could go to their manager and complain that this wasn’t fair. The supervisor might well change the schedule, or at least explain the reason and possibly offer to do something else that made up for it.
Algorithmic scheduling can be quite different. Let’s say that the algorithm schedules an employee to work two weekends in a row, and they go to my supervisor to complain. They ask why this happened, and what can be done about it. Will the supervisor hand their employee the email of the software programmer in Silicon Valley who wrote the code? Unlikely. It’s easy to assume that supervisors could intervene and make exceptions, but it is extremely tricky for algorithmic formulas to keep working when we don’t follow the outcomes. Maybe the employer can pretend that a manager made the decision to hire someone, but at some point the employee will learn that it was a score on the algorithm that did it just as a score on a different algorithm determined that they got the expat assignment or the promotion.
There are compelling arguments for making greater use of algorithms in all areas of employment. They reduce sharply at least the explicit biases that occur in the final decision processes, and if they are constructed properly, they lead to better matches between tasks and the attributes of people. All that is great, especially if the organisation operates smoothly and consistently. But the workplace rarely operates like that. Businesses need change, sometimes very quickly. People get sick or quit, complicating schedules and plans. Contrary to some contemporary views, the most important flexibility needed to adjust to these persistent challenges does not come from using contractors or hiring and firing workers. It comes from within the existing workplace where people adjust and go beyond their standard roles, taking on new tasks, and rearranging their own efforts to accommodate these ups and downs.
A fundamental problem with algorithms is that they take decisions and power away from supervisors who in the past used them in part to build the relationships that made the workplace go. A second basic problem is that we cannot explain how they work. Why is it that this candidate got hired rather than that one? Yes they had a higher score, but what about them led to that higher score? There could be dozens of factors in the algorithm, and the machine learning-based equation on which it is built has complicated non-linear relationships in it that are hard to pull out, let alone to explain. There is no doubt on paper that this is far more efficient than using simple rules like seniority to make decisions, but seniority at least is understandable and therefore seems more fair.
Whether algorithms actually prove to be more effective once we start observing employee responses to them, and to the diminution of supervisor control that comes with them, is an open question. Whether it will be possible to strike some kind of balance between the optimisation norms of algorithms and the fairness concerns of employees remains open as well. Only time will tell.
