Researchers working with Professor Michael Neugart from the Public Finance and Economic Policy Group within the Department of Law and Economics at TU Darmstadt investigated digital work platforms such as ‘Amazon Mechanical Turk’. Such platforms act as intermediaries for work and services, for example in the fields of software development, multimedia, translation, data entry or marketing.

In Germany, crowdworking – a form of work in which companies award contracts to external workers via digital platforms – is not yet a widespread phenomenon. It is primarily used as a source of supplementary income. At the same time, its use has increased significantly in recent years: according to a survey by the Centre for European Economic Research (ZEW), around 8.2 per cent of companies in the German information industry were using crowdworking in 2020.

In the model, companies leave the decision on the wages offered to so-called deep Q-networks (DQN). These are self-learning algorithms trained using reinforcement learning. They utilise neural networks that can recognise complex relationships based on data and derive decisions from them.

For their study, the researchers combined economic modelling with machine learning methods and extensive computer simulations. The central question was whether the algorithms could independently learn to set wages below the level that would result from normal competition between companies.

Low wages without any explicit agreement

The simulations showed that, under certain conditions, self-learning algorithms can produce wages that are significantly below the competitive level. If only a few firms are competing with one another, wages may approach a level that would result from collusion between the firms – that is, an agreement that restricts competition.

cst