Do You Want to Have an Algorithmic Boss?
Amazon drivers are now being punished when other people cut them off. Last week’s Insider article documents complaints made by Amazon drivers regarding a new system that tracks their behavior and rewards or punishes them accordingly. But let’s start by first understanding what the system does.
“Amazon said it has seen a reduction in accidents and other safety violations since installing the Netradyne cameras in its delivery vehicles. When the cameras spot possible unsafe driving "events," these instances factor into workers' performance scores and can, in turn, hurt their chances of getting bonuses, extra pay, and prizes. They can also affect the income of the Amazon delivery service partner itself.”
Algorithms that are used to manage and make decisions about employees are not unique to Amazon. In fact, most gig economy platforms use algorithms to make most of their decisions, from pricing to matching to promotions.
Of course, since we are dealing with people, one would expect at least one of two things: some will try to “game” these algorithms, and some will protest them. So far, we have witnessed both.
On the gaming side, Bloomberg published an article on how DoorDash employees tried to game the algorithm. The article, DoorDash Drivers Game Algorithm to Increase Pay, tells the story of Dave Levy and Nikos Kanelopoulos, two DoorDash drivers, who created #DeclineNow; a 40,000-person online forum. Through this forum, they are prompting their fellow Dashers to turn down the lowest-paying deliveries so that the algorithm used for matching tasks with drivers will respond by raising pay rates.
On the protest side, two years ago, Uber Eats drivers protested in Australia, in front of the NSW Parliament, against an alleged algorithm change which they claimed cut their income by up to 50%. Similar protests against Uber and Lyft occurred around the same time across the US.
Management by Algorithm
The use of algorithms to make management more efficient is not new. Some people call it Algorithmic Management. The term is used to describe management practices that use algorithms to either support or make labor decisions. The term was initially coined in 2015 by Min Kyung Lee and her co-authors to describe the managerial role played by algorithms on the Uber and Lyft platforms.
The Data&Society explainer of these practices categorize these algorithms into five buckets:
Data collection and surveillance of workers through technology;
Support systems to real-time, data-driven management decisions;
Automated or semi-automated decision-making;
Performance evaluations via rating systems or other metrics; and
The use of “nudges” and penalties to indirectly incentivize worker behaviors.
But even the notion of an algorithmic boss is not entirely new. Its origin lies in the idea of Taylorism, or more formally, Scientific Management, which describes the concept of using data and empiricism to analyze workflows and labor productivity. It takes its name after its pioneer, Frederick Winslow Taylor, who began developing the theory in the late 19th century within manufacturing industries (in Pennsylvania).
In his work, Taylor rejected the notion (which was common at that time) that many jobs, including manufacturing, could only be performed by craft production methods and were thus immune to rigorous analysis. In his empirical studies, by observing workers, Taylor examined various jobs and how a worker’s productivity was affected. He found, for example, that workers were more efficient when labor included small rest periods (breaks), offering them time to recover from either mental or physical fatigue. Some of this may now sound trivial, but the debate was ideological rather than factual, since there was no data.
These days, the term Taylorism has a very negative connotation, and it is very uncommon to see “time and motion studies” carried out. Still, the idea of algorithmic management is a clear continuation of it.
The goal is quite simple: firms are trying to develop a scalable method of managing a massive workforce using data and scientific methods.
The Benefits of the Algorithmic Boss
At face value, this does not sound all that bad.
And indeed, the benefits for the firms and consumers are significant.
For example, how many times have you had to tell an Uber driver not to text while driving? I have done it too many times, but I know that I get a bad rating for it every time (or at least, that is what I tell myself to explain my less-than-perfect score).
The fact that Amazon tracks its employees “up to their eyeballs” ensures that these drivers are more likely to drive safely. Does it solve all problems? Absolutely not. Can it be gamed? Absolutely. Will Amazon drivers drive safer? I am sure they will. And the data already supports it.
“The company added that it has seen the following changes since installing the cameras in more than half of its US fleet: accidents decreased 48%, stop-sign and signal violations decreased 77%, following distance decreased 50%, driving without a seat belt decreased 60%, and distracted driving decreased 75%.”
The same is true for Uber tracking their drivers. Susan Athey and Juan Camilo Castillo show, in their paper Service Quality on Online Platforms: Empirical Evidence about Driving Quality at Uber, that Uber drivers perform better than UberTaxi drivers. They find that drivers respond to user preferences and nudges, such as notifications due to low ratings. Specifically, they show that informing drivers about their past behavior improves quality, especially for low-performing drivers. How did they get an objective metric of how drivers behave? They looked at telemetry data, an objective measure of quality. Very soon, Lyft will be doing the same (if it is not doing so already).
In other words: customers are better off. Employers are better off, given their ability to use processes that create leverage and power while improving safety, quality, and cost.
So Why Do Workers Hate Them?
As the article about Amazon reports, the drivers hate these systems. No one wants to be surveilled, but the reasons are more profound and should be a warning sign to all of us.
Let me outline some of these reasons:
Algorithms are arbitrary: They definitely feel random to the workers, even if the algorithms are elementary. But these algorithms will only become more arbitrary over time, given that the more sophisticated methods in machine learning, such as deep learning, are hardly interpretable even for those developing them.
Algorithms are unfair: As mentioned in the Amazon article, some of the drivers actually behave properly, but the algorithm still views their behavior as dangerous:
"It's upsetting when I didn't do anything," a Los Angeles delivery driver told Motherboard. "Every time I need to make a right-hand turn, it inevitably happens. A car cuts me off to move into my lane, and the camera, in this really dystopian dark, robotic voice, shouts at me.
Algorithms are biased: There are excellent studies about this topic. The most famous one is from ProPublica, which shows that algorithms only replicate and amplify biases rather than remove them. This is a whole area of research that I will not get into here, but you can read more about it in the book by Cathy O’Neil, Weapons of Math Destruction.
Finally, algorithms include a level of surveillance and control measures that most workers were unaware of when they signed up for this type of work.
In general, these algorithms are very dehumanizing and remove much of the accountability that these firms should shoulder for their workers.
Long-time readers of this newsletter know that I don't believe that optimal solutions exist. Only trade-offs. So, what’s the trade-off?
Are We Trying to Scale Algorithms Too Fast?
In this case, looking at the trade-offs requires us to delve deeper into these algorithms and understand the source of their scalability.
The goal is to develop labor management systems that are instantly scalable. That is, to generate significant benefits without any increase in cost as more workers are added.
The main issue is that any attempt for instant scalability will always result in the misuse of technology or in over-reliance on technology.
Technology can be significant in identifying dangerous drivers. But we all need to admit that algorithms make mistakes. The type of mistakes they make and our tolerance to these mistakes is the crux of the issue here.
The algorithm makes two types of mistakes (and let me use the Amazon example to illustrate):
False-negatives: people that drive dangerously but cannot be identified by the algorithm.
False-positive: people that drive safely but are flagged by the algorithm as driving dangerously.
We would all like algorithms that make no mistakes, but they do. The solution is to minimize one type of mistake and mitigate the impact of the other.
In this case, we want to minimize the false-negatives. We want to make sure Amazon does not employ dangerous drivers. So, we design an algorithm that is very sensitive to every behavior that may be considered hazardous.
But by doing so, we increase the number of false-positive mistakes, since the system is overly sensitive. What must we do? We must mitigate the impact of these false alerts. How? By adding a human element to the process. How? We make sure the alerts created by these systems are being reviewed and discussed by their supervisors. We create a simple appeal process. This will make the systems less dehumanizing while increasing their effectiveness over time.
The downside of adding people to the process: it's costly and slows the algorithm down.
Only trade-offs. No perfect solutions.
The Impact on You
If you think that this will not impact you, you are probably in for a surprise.
On Friday, Dan Price tweeted a series of complaints from people who work from home and received emails alerting them that their employers would be tracking their online activity. As we will reach a more “gigified” economy, and one that encourages more work from home, expect to be tracked and managed by an algorithm.
Do I want to be managed by an algorithm?
Let’s look at labor decisions in academia as an example.
When someone promising doesn't get tenure, I am often asked how that is possible, given the (usually high) number of publications they have. People are generally amazed that we do not have clear, objective criteria such as the number of overall publications, teaching evaluations (yeah right), or the number of publications in “A” journals.
And the answer is that there are very broad guidelines, but ultimately, the question is whether someone is a "thought leader in their area.” And this is something algorithms cannot assess. The line is fuzzy, and it's fuzzy on purpose; to force the decision-makers (the tenured faculty) to truly use judgment when making their decision.
Any attempt to employ merely countable criteria means that the decision has been turned over to an algorithm (counting is an algorithm). Using an algorithm, in this case, does exactly what you expect it to: it reduces the accountability of the decision-makers. And since hiring is the most critical decision we make in academia, we keep it as human-driven as possible.
The downside: you may not like people's judgment. It may be political... Subjective... Unscientific...
Only trade-offs. No perfect solutions.