Imagine a small town with two merchants selling widgets. Customers naturally prefer lower prices, prompting the merchants to compete. However, dissatisfied with their shrinking profits, they contemplate a clandestine meeting to agree on raising prices collaboratively—a practice known as collusion, which remains illegal. Ultimately, they refrain from this risky plan, allowing consumers to enjoy competitively priced widgets.
For over a century, U.S. law has focused on outlawing such covert agreements to preserve fair pricing. However, the landscape has changed with the rise of algorithms—specifically learning algorithms that adjust prices based on market data. While these algorithms tend to be less sophisticated than the AI-driven deep learning models, they are still capable of displaying unpredictable behavior.
This raises a question for regulators: How can they ensure fair pricing without clear signs of collusion? Traditional methods of regulation, which rely on proving explicit agreements, may no longer suffice in an age where intelligent software dictates market dynamics. As computer scientist Aaron Roth points out, "the algorithms definitely are not having drinks with each other."
A landmark study from 2019 indicated that algorithms could engage in tacit collusion without being specifically programmed to do so. Researchers simulated a market with two versions of a simple learning algorithm, discovering that they developed a way to retaliate against price cuts by sharply increasing prices. This mutual threat led to inflated prices, despite no formal agreement being in place.
The implications of these findings are still debated within the scientific community, with much depending on how terms like "reasonable" are defined. Still, they underscore the complexities of algorithmic pricing and the challenges it poses for regulation.
Roth and his colleagues have explored this issue further in recent research. They demonstrate that even algorithms deemed benign can produce adverse impacts, leading to high prices in seemingly reasonable ways. Economist Mallesh Pai emphasizes that without visible collusion, it becomes difficult for regulators to challenge perceived pricing injustices.
Delving into game theory—which combines economics and computer science—researchers are investigating how collusion might be replicated in controlled environments. Joseph Harrington, an economist, explains that their goal is to create collusion in a lab setting in order to figure out how to break it apart.
A familiar analogy in game theory is rock-paper-scissors, where players adapt their strategies based on opponents’ previous moves. If played correctly, participants reach an "equilibrium" where neither has incentive to alter their strategy. In a long series of rounds, players employing learning algorithms can significantly outperform random players, leading to regret if they fail to adapt.
Game theorists have developed "no-regret" algorithms that guarantee no player ends up wishing they had chosen differently. Past research indicated that if two such algorithms compete, they cannot collude and will instead maintain competitive pricing.
However, researchers revealed that other algorithmic strategies might undermine this ideal. In their experiments, they found that a learning algorithm confronting a no-swap-regret algorithm could adopt a nonresponsive pricing strategy—picking its moves randomly—and still yield substantial profits. This alarming discovery flipped expectations: these nonresponsive strategies could coax responses from their competitors without explicit collusion.
Initially, researchers worried that such strategies might not hold real-world relevance, suspecting players would quickly recognize and adapt to their competitors’ behaviors. They were surprised to find that equilibrium could sustain itself without players feeling pressured to change their strategies. This results in high prices for consumers, resembling collusion despite the absence of direct collusive behavior.
The crux of the issue is determining suitable regulations. Roth admits that challenges abound in crafting effective policies. While banning no-swap-regret algorithms might seem reasonable, such a move could paradoxically drive prices up if implemented across the board. On the other hand, the potentially harmful nonresponsive strategies might lead sellers to inadvertently inflate prices.
Proposals to address the situation vary, with some suggesting stricter conditions around algorithm types used in pricing. For instance, Hartline, a computer scientist, advocates for limiting algorithms to those that are no-swap-regret, even if not all adverse outcomes could be avoided.
Ultimately, these advancements provoke crucial questions about how to govern the increasingly sophisticated algorithms that govern market pricing. As Pai notes, understanding these mechanisms remains vital, given their centrality to modern economics.
