Algorithmic Competition in the Gig Economy

Many workers are stressed about their professional future given the progress of Artificial Intelligence (AI) and Machine Learning (ML). One of the main concerns is that many of the tasks currently carried out by humans will be replaced in the near future by machines. 

Another concern, related to the above, is that Machine Learning is going to be used by people to better manage other people. 

Well, the future is already here and we can get a glimpse of it via the gig economy. 

As I have mentioned before, gig economy platforms use algorithms to make almost every worker-related decision, from wages and promotions, to job assignment or termination. Algorithms have yielded two different reactions: some try to “hack” and game them, while others (or maybe the same group) complain about their arbitrariness, primarily when it comes to job termination, and try to find ways to regulate them, with the EU leading the way. 

A week ago, the European Commission proposed measures to improve working conditions for gig economy workers.

“The new rules will ensure that people working through digital labour platforms can enjoy the labour rights and social benefits they are entitled to. They will also receive additional protection as regards the use of algorithmic management (i.e. automated systems that support or replace managerial functions at work).”

The EU legislators have many interesting ideas on how and when to classify these workers as employees; ideas which seem to be in line with more labor-friendly suggestions.

But the most interesting part is the requirement for more transparency in the use of algorithmic management. 

In particular, the European Commission wants gig work platforms to be more transparent about the algorithms they use to manage workers. According to the proposal, workers should be able to understand how jobs are assigned to them and how their pay is set. The proposal also suggests that people should oversee the algorithms, and workers should be able to appeal any automated decisions.

The central assumption is that there’s a battle between workers and algorithms. 

But what if I told you that the main issue is not how ML is used to manage people, but rather how ML is used to compete with other firms (which are run by people) that are also using ML to manage the same gig workers. 

In other words, what we are seeing is that those who work on behalf of the gig platforms (e.g., managers and data scientists) develop algorithms that then compete, in their favor, on “utilizing” other people (the gig workers).

Ultimately, drivers don’t have to outsmart the algorithms; they can just let the algorithms (e.g., the ones used by Uber or Lyft) outcompete each other on them. 

And indeed, it's clear that algorithms compete more and more. 

Algorithmic Competition

You are probably wondering whether it will be beneficial to have different algorithms competing with each other and who will ultimately benefit from this competition. 

The answer is not all that trivial. While different projects deal with how workers make decisions, there are only a few that I know of which try to address how the world will look like when these algorithms compete among themselves. 

Specifically, as achieving optimal pricing becomes increasingly difficult, and algorithms become increasingly complex (to the point that most of those who use algorithms don’t fully understand them and some of those who develop them can’t fully explain their results), several researchers have started looking into the question of algorithmic competition.

Last week, Chamsi Hssaine, a Ph.D. student from Cornell and a prominent job candidate in operations management, visited Wharton and gave an excellent talk on studying the impact of competing algorithms. Her study explores the consumer side, but I will also try to discuss the implications for gig workers. 

Chamsi and her co-authors study a theoretical model of competition among ride-hailing firms, and begin by modeling how consumers make decisions in this market: Customers sample firms’ prices in their preferred order until they find a firm with pricing levels below their WTP (willingness to pay). 

The research paper relies on the fact that in order to optimize such a complex system, firms must use optimization algorithms, and the team specifically studies the impact of firms using distributed online gradient descent algorithms.

For those who are not familiar, imagine that the algorithm is trying to find the bottom of the “mountain” by choosing the steepest descent (or ascent if it’s trying to find the peak). I am clearly oversimplifying (and apologies to those who already know this). 

Chamsi and her co-authors show that the price dynamics resulting from utilizing such an algorithm may often converge to an undesirable outcome. 

Specifically, the study shows that these “games” are frequently plagued by a local Nash equilibrium: The price of the firm with a smaller market share is only a “local” solution rather than the “best” solution. 

“Our results thus suggest that algorithms that adaptively rely on local information to learn good pricing decisions, {despite faring well in monopolistic settings}, may suffer from serious drawbacks in competitive environments.”

In other words, the way algorithms compete with each other results in getting stuck in what is known as Local Nash Equilibrium. Nash equilibrium is a concept used to describe the anticipated solution of a “game” in which no one has a profitable deviation, assuming everyone plays their expected strategy. Local Nash Equilibrium is a slightly refined concept where no one has a profitable deviation when searching for the optimal solution “locally.” What does locally mean? Small perturbations of the current pricing. 

But why is it necessary for algorithms to search locally? 

Finding optimal price levels is a highly complex problem and firms must react fast to a constantly changing environment. Carrying out an exhaustive search of all possible solutions is expensive and slow. 

The immediate implication is that algorithms get stuck if they compete. 

It seems that firms are worse off if they continue to use local search and standard algorithms that work well in a non-competitive setting. 

As I mentioned before, Chamsi’s research explores the consumer side. 

But I don’t think there’s a big difference since gig economy platforms use similar algorithms to compete on the worker side. 

So if firms are not better off, then who is? 

At a first glance, drivers and customers seem to be the ones benefiting. 

The paper also outlines ways that firms can get out of this local Nash equilibrium. But even under the real Nash equilibrium, I expect that competition will favor drivers rather than firms. 

Algorithmic Pricing and Collusion

But algorithmic prices may also work against customers. Algorithms may collude with each other to benefit other sides of the market. 

In a working paper with Ken Moon and Amandeep Singh, our initial results show that when Airbnb offered ‘owners’ the ability to use the Smart Pricing algorithm and better price their apartments, the algorithms’ market behavior was as if these owners actually colluded with each other, rather than competed. 

What’s the difference: In the gig economy, we expect firms to compete with others and use better algorithms to do so. In the case of Airbnb, the same firm utilizes the same algorithm to help the same side of the market (the apartment owners) compete with itself. In all these cases, the lack of transparency and precise mapping between actions and outcomes make it hard to understand who is actually hurting and who is benefiting. 

But this also highlights another issue: any attempt to allow drivers to collude and bargain collectively may actually create a loophole that allows others to also collude, under the guise of algorithmic pricing. 

This has been one of the main issues brought up by legislators when allowing gig workers to bargain collectively. 

Finally, it’s important to say that algorithms are not only used to generate pricing but also for sanctioning workers. And while in many ways this is beneficial to consumers, the arbitrariness is concerning. The future is quite complex, thus transparency is key. 

To conclude, this is an extremely competitive and complex marketplace

Gig work platforms use algorithms to compete among themselves.

Drivers compete among themselves. 

Customers compete among themselves. 

So who is gaining from all this competition? 

Probably those using the algorithms.

It seems the EU Commission is taking a step in the right direction, but it’s only a start.