Can two AI agents fix prices without ever agreeing to? Tanise Brandão and Carlos Neves of Brazi’'s competition agency, CADE, join Thibault Schrepel and Alba Ribera Martínez to walk through their paper on algorithmic collusion.
They put two large language models in a repeated Bertrand duopoly, let them price against each other for 400 rounds, and watch supra-competitive prices emerge on their own. The agents coordinate even when they cannot communicate, and some learn to cover their tracks, with one proposing that they steer clear of red flags.
The conversation turns to why the choice of model and language moves the result, and why the behavior looks less like textbook rationality than a complex dynamical system. It closes on enforcement, where collusion no longer needs a meeting of minds, and Brandão and Neves make the case for shifting liability from whether firms agreed to what they chose to expose themselves to.
Podden och tillhörande omslagsbild på den här sidan tillhör
Thibault Schrepel. Innehållet i podden är skapat av Thibault Schrepel och inte av,
eller tillsammans med, Poddtoppen.
