Research conducted by institutions including TU Darmstadt, Bielefeld University, and the University of Côte d’Azur has revealed that artificial intelligence (AI) can generate extremely low wage levels on digital working platforms without any explicit communication or agreement between the companies involved. These findings, published in the journal “Labour Economics,” stem from complex computer simulations.
The scientists analyzed digital labor platforms, such as Amazon Mechanical Turk, which facilitate work and services in fields ranging from software development and multimedia to data entry and marketing. While crowdworking is currently mainly utilized in Germany as a supplementary income source, the study provides context by noting that roughly 8.2% of German information technology companies adopted crowdworking in 2020, according to the European Centre for Economic Research.
In their model, the researchers allowed various companies to set the wages offered by what are called Deep Q-Networks. These are self-learning algorithms trained using reinforcement learning and utilizing neural networks. The simulations demonstrated a key insight: if only a few firms compete, wages can approach the level that would result from a deliberate collusion among those companies.
This research directly contributes to the field of algorithmic collusion-a domain that has traditionally focused on elevated, collusion-like prices in product markets.
According to the universities involved, the results hold significant relevance for labor market and competition policies, as well as the regulation of platforms and AI systems. Michael Neugart, from the Faculty of Law and Business Sciences at TU Darmstadt, commented on the implications, stating that the study highlights a problem that could become increasingly prominent as the automation of economic decisions advances. He noted that companies might not need to explicitly coordinate on wages or prices; under specific conditions, learning algorithms can independently develop behaviors whose market outcome mirrors that of a formal agreement. For competition authorities, understanding these characteristics of modern machine-learning algorithms is crucial to identify when they favor non-competitive market outcomes.
