Setting and adjusting product prices has always been a big challenge in retail. With the emergence of e-commerce, mobile commerce, and other factors, the problem is now bigger. Retailers now make use of sophisticated pricing engines to make frequent price adjustments.
Can AI Help Pricing Algorithms Learn to Collude?
In previous eras, pricing was a simple process, you set prices according to your location, target customers and your competition in that area. Now, prices are constantly changed in response to various factors. Ecommerce giants like Amazon change prices on thousands of products several times each day.
Initially, simple rule-based pricing algorithms were used. These would monitor prices across various competitors, and market trends. They would then adjust prices within the parameters set by the retailer.
Machine Learning Pricing Algorithms
Next came smart pricing systems that could learn as they made changes. These pricing systems are constantly making changes against each other. Repricing is a good thing for both buyers and sellers, it helps keep prices in control, and it also helps sellers avoid a race to the bottom.
However, these new smart systems can learn from their previous moves and the result of each action. The concern now is, can these systems learn to collude and keep prices at deliberately high levels, a disadvantage to customers, but beneficial to sellers?
The Antitrust Issues
Concerns regarding AI Pricing systems were first expressed in a paper in 2015 presented by Ezrachi and Stucke. Pricing algorithms use the same AI technology as those used in highly complex strategic games like chess. There, AI systems have already demonstrated their ability to learn and evolve and even beat humans.
The concern is that these systems can learn from trial, and error that the best way to increase profits would be to avoid price wars and keep prices at an artificial high. If all sellers sold at the same or similar high prices, customers would be left with no choice but to buy the products at these high rates.
If this sort of price changes were initiated by humans, they would be subject to antitrust laws and be declared illegal. However, there is no law that can apply to smart pricing systems
Here, these AI-powered systems are not specifically instructed to keep prices high or even to collide. However, they may learn to collude on their own with time.
The Recent Study
There are arguments for and against the theory of collusion. Some say that there is not enough evidence to confirm collusion. They point out that colluding without active communication is difficult for even the smartest of AI systems. They say that incidences of over-pricing in simulation could merely indicate that the algorithms have not yet learnt to achieve competitive equilibrium.
So, to shed further light on the debate, a new study was conducted in 2018 by Emilio Caverio et al a Bologna University, Italy. They constructed AI pricing algorithms and let them make constant price adjustments. In one instance, they forced one algorithm to plunge downwards. They found that the other algorithm also went down, canceling any price advantage the first one may have achieved. It then moved back up to the original convergent price again. While the deviations settled below monopoly price, the researchers found evidence of definite and persistent collusion among different pricing algorithms. However, they do not leave any trace behind of cooperative effort. They merely make adjustments through trial and error.
The concern now is for more assertive evidence to prove or disprove the tendency for collusion in AI-based pricing engines. This can help decide if the issue needs to be addressed by law.