Hybrid social learning in human-algorithm cultural transmission

Abstract

Humans are impressive social learners. Researchers of cultural evolution have studied the many biases that enable solutions and behaviours to spread socially from one human to the next, selecting from whom we copy and what we copy. In a digital society, algorithmic and human agents both contribute to transmission of knowledge. One hypothesis is that machines may influence the patterns of social transmission not only by providing a means for spreading human behavior but also by providing novel behaviors themselves. We propose that certain algorithms might show (either by learning or by design) different behaviors, biases and problem-solving abilities than their human counterparts. This may in turn foster better decisions in environments where diversity in problem-solving strategies is beneficial. In this study, we ask whether machines with complementary biases to humans could boost cultural evolution in a lab-based planning task, where humans show suboptimal biases. We conducted a large behavioral study and an agent-based simulation to test the performance of transmission chains with human and machine players. In half of the chains, an algorithmic bot replaced a human participant. We show that the bot boosts the performance of immediately following participants in the chain, but this gain is lost for participants further down the transmission chain. Our findings suggest that machines can potentially improve performance, but human bias can hinder machine solutions from being preserved, especially under conditions of uncertainty or high cognitive load. Our results suggest that the conditions for hybrid social learning and cultural evolution may be limited by task environment and human biases.

Algorithm-Mediated Social Learning in Online Social Networks

“…Designers of content algorithms can take inspiration from contexts where algorithms appear to augment functional human social learning (Box 1). Several studies examining complex but well-defined game solutions (e.g., Go, Chess) have demonstrated that algorithm-mediated social learning is mutually beneficial for humans and algorithms when algorithms develop complementary learning biases that increase the diversity of strategies to solve the problem (Brinkmann et al, 2022;Strittmatter et al, 2020). These findings support the general idea that a diverse set of problem solving strategies reduces error and bias over time (Hong & Page, 2004;Yaniv, 2011).…”

From the origin of life to pandemics: emergent phenomena in complex systems

“…Brinkmann and coauthors shed light on the role that the interaction between humans and algorithms might play in shaping the emergent properties of cultural evolution. In Hybrid social learning in human-algorithm cultural transmission [178] they propose a set of six hypotheses, related to the improvement of collective performance tasks via (hybrid) social learning, that are tested in an experimental set-up. Their empirical findings highlight the importance of biases: even if an algorithm aims to aid humans, the provided information can be quickly lost in successive human-human interactions due to, precisely, human biases.…”

From the origin of life to pandemics: Emergent phenomena in complex systems

“…Brinkmann and coauthors shed light on the role that the interaction between humans and algorithms might play in shaping the emergent properties of cultural evolution. In Hybrid social learning in human-algorithm cultural transmission [178] they propose a set of 6 hypothesis, related to the improvement of collective performance tasks via (hybrid) social learning, that are tested in an experimental set-up. Their empirical findings highlight the importance of biases: even if an algorithm aims to aid humans, the provided information can be quickly lost in successive human-human interactions due to, precisely, human biases.…”