The Game Theory of How Algorithms Can Drive Up Prices (www.quantamagazine.org)

Researchers studying algorithmic pricing used game theory and simulations to show a new, subtle route to high consumer prices that doesn’t rely on explicit collusion or retaliatory “threats.” Building on prior work showing learning algorithms can tacitly collude, Collina, Arunachaleswaran, Roth and colleagues proved that when a so‑called no‑swap‑regret algorithm (one that guarantees you couldn’t have benefited by consistently swapping one action for another) faces a simple “nonresponsive” opponent that randomizes with unusually high probability on expensive prices, the pair can settle into an equilibrium with persistently high prices. The nonresponsive strategy never reacts to the opponent (so it can’t be flagged as collusive), yet it nudges the learning algorithm to raise prices; neither side then has an incentive to deviate, and many randomized mixes produce the same bad outcome. The result matters because it undermines regulation strategies that look for explicit agreement or retaliatory behavior and suggests that apparently benign algorithms can produce consumer harm. Remedies are thorny: one proposal is to require sellers to use no‑swap‑regret algorithms (Hartline et al. have methods to test for that property without inspecting code), but that won’t eliminate all failure modes and may be impractical. The work highlights deep open questions about defining and policing “unfair” algorithmic pricing when harmful equilibria can arise without overt collusion.