AI-Driven Cultural Evolution
Sample Project: Recommender Systems and Cultural Evolution (Ongoing)
Sample Project: Human–Machine Cultural Evolution (Ongoing)
How will AI accelerate the evolution of human culture and innovation?
The ability of humans to create and disseminate culture is often credited as the single most important factor in our success as a species. Culture evolves as a function of the range of available traits (variation), the ways these traits can replicate (transmission), and the success they encounter (selection). Whenever there is a major transition in any of these three processes, cultural evolution is altered. For example, in the space of roughly 200 years (1300–1500), European culture was rocked by a succession of major transitions in variation, transmission, and selection. Through the tales of Marco Polo and other explorers, Europeans were exposed to a broad set of novel cultural traits, such as paper money and coal burning (variation); the invention of the printing press made it easy to replicate cultural traits with great fidelity (transmission); and the Black Death drastically changed the social and economic opportunities of commoners, reshuffling the value of cultural traits (selection).
Modern communication technologies have scaled some of the key factors that shape human cultural evolution, enabling the instant sharing of knowledge across the globe and the rapid self-organization of online communities of interest. We believe that a particular digital technology, artificial intelligence (AI), can substantially impact the process of cultural evolution, from recommender algorithms altering the flow of knowledge to AI agents becoming participants in the generation of culture itself, from music and visual art to scientific discoveries. Our long-term goal is to map the different ways in which AI will impact—or may already be shaping—human culture, and to establish a research agenda for behavioral scientists studying hybrid cultural evolution in the digital age.
Sample Project: Recommender Systems and Cultural Evolution (Ongoing)
Can algorithmic recommendations improve on social learning in complex cultural spaces?
The efficient accumulation of culture depends on well-adapted social learning strategies that enable individuals to benefit from and build on the knowledge of others (Laland, 2004, Social learning strategies; Thompson et al., 2022, Complex cognitive algorithms preserved by selective social learning in experimental populations). Traditionally, social learning has been considered in the context of peer-to-peer communication, where individuals learn or copy cultural elements through their direct social relationships. This has changed dramatically in the digital age, where vast online information ecosystems and reservoirs of cultural content have emerged. In such environments, individuals acquire culture not (only) through their immediate social contacts; they also have easy access to information that comes from outside their social environment (Acerbi, 2019, Cultural evolution in the digital age). Due to their immense size and reach, most of these online digital environments could not function without the use of algorithms to filter and redirect flows of information. Therefore, individuals usually access information through recommender systems that help users navigate online environments and facilitate the discovery of relevant information that is often highly personalized, with notable examples ranging from friend recommendations on social media sites to e-commerce platforms. Algorithmic information filtering mechanisms in recommender systems are usually based either on the particular nature of the content or on other properties of cultural elements (or cultural models), such as their frequency/novelty, prestige, or similarity.
Here, we ask how recommender systems might be able to improve on cultural accumulation, compared to a basic variant of social learning in which individuals self-select what information they copy. We hypothesize that recommender systems might be particularly beneficial in cultural spaces with a relatively higher complexity because these should be harder to navigate. We first construct an agent-based model addressing this question. Agents in the model attempt to maximize their payoff by combining different cultural items, which sometimes can result in the discovery of a new, better-performing item. Importantly, our cultural space is constructed hierarchically in such a way that better-performing items branch out into separate independent trajectories (Figure 1), representing specialization or personalization (in the case of recommendations). Complexity is varied with the number of these specialized branches, and agents always only learn either socially or receive algorithmic recommendations within a single run.
Figure 1. Visualization of the cultural space for a “simple” environment. Agents/participants can reach better-performing items by recombining known items and moving upwards in a specific branch. A more complex cultural space would exhibit more branches and starting items.
We find that generally recommenders outperform social learning in this task. Furthermore, higher complexity is detrimental for agents relying on social learning, while recommendation reaches optimal performance for an intermediate complexity of a given space. In a second step, we want to test the robustness of these model results in an online experimental version of the task that involves human participants. Focusing on the central results for complexity, we will construct a 2x2 manipulation of recommenders/social learning and low/high complexity, predicting a pattern similar to the one found in the model with respect to participants’ performance.
Sample Project: Human–Machine Cultural Evolution (Ongoing)
The project investigates the role of selective social learning on culturally maintaining algorithmic solutions that conflict with human biases.
How can human cultural evolution benefit from making use of artificial intelligence? Previous research has shown that AI can help humans on tasks where myopic behavior is maladaptive (Lieder et al., 2019, Cognitive prostheses for goal achievement; Brinkmann et al., 2022). In our previous study (Brinkmann et al., 2022), we showed that unintuitive but better-performing solutions created by AI can also be imitated by humans. This represented the first controlled experiment showing social learning by humans from AI, yet the effect was short-lived: Subsequent human learners did not adopt the superior AI strategy. Recently, an experimental study by Thompson et al. (2022, Complex cognitive algorithms preserved by selective social learning in experimental populations) demonstrated that complex cognitive strategies can successfully be transmitted and spread culturally if human participants have the opportunity for selective social learning, i.e., they can choose to learn from successful individuals. Can an unusual but superior AI solution be adopted successfully by humans, if they can use selective social learning?
In a follow-up experiment to our original task, we will present participants with similar “reward networks” that task them with finding an optimal path between nodes connected by routes with varying payoffs. Importantly, we know that humans are biased against superior long-term solutions and prefer myopic solutions instead, while AI has no such bias and will find the optimal solution. In contrast to Brinkmann et al. (2022), we will replace the transmission chain design of the experiment with chains of groups, such that later generations of participants will have the option to select from a group of previous players. Our manipulation concerns the presence of (superior) AI solutions in that group of previous players in one condition.
We suggest that the option for selective social learning will enable participants who are offered the unintuitive AI solutions to outperform typical human strategies even in later generations, once they are solely transmitting from human to human. This would constitute experimental evidence that AI has the potential to transform human culture persistently by introducing novel and high-performing solutions, which can be maintained in subsequent human cultural transmission.
Figure 2. Left: The reward network task as presented to participants. Each possible move on the network is associated with a positive or negative reward. Some of the networks we designed have an optimal strategy that conflicts with human biases, i.e. they require accepting large losses initially to eventually obtain a large reward. Right: The upcoming experiment is structured with multiple participants constituting a generation. Participants have the option of selecting two participants of the previous generation to learn from, which represents selective social learning. We compare two different conditions: a human-only condition, and a condition in which algorithmic agents seed unbiased superior solutions.
Key Reference