A computational reward learning account of social media engagement

The widespread and often obsessive use of social media platforms like Instagram, Facebook, and Twitter, involving billions of users spending hours daily, has prompted comparisons to addiction. This engagement is frequently attributed to the pursuit of positive social feedback, such as 'likes,' framing social media as a modern-day Skinner Box. However, empirical evidence directly supporting this reward-based learning account has been limited. This research utilizes a computational approach to analyze large-scale social media data to test if reward learning mechanisms contribute to social media behavior. The study aims to provide novel insights into this prevalent form of human interaction and simultaneously test a learning theory model on an unprecedented scale. Social media 'likes' are assumed to function as social rewards, engaging similar motivational mechanisms as other rewards like food or money. Neuroimaging studies show that social rewards activate brain regions similar to those activated by non-social rewards, suggesting that social media use might involve reward maximization. While previous research indicates that 'likes' correlate with satisfaction and increased social media activity, direct evidence for a reward learning account of social media behavior remains lacking. Existing studies using reinforcement learning (RL) often focus on optimizing software rather than understanding psychological mechanisms. Furthermore, quantitative studies yield mixed results. This study directly addresses this gap by applying a computational approach to rigorously test the reward learning hypothesis on real-world social media behavior.

Prior research has explored the motivations behind social media usage, ranging from simple self-expression to complex relational dynamics. The prevalent view of social media as a reward-driven environment has some support from neuroimaging studies showing that social rewards (e.g., 'likes') activate brain regions associated with reward processing. This overlap with the neural processing of non-social rewards suggests the possibility of reward maximization in social media engagement. Several studies have found correlations between the number of 'likes' and self-reported satisfaction, happiness, and increased subsequent activity. Moreover, the subjective value of likes is also influenced by social comparison, similar to non-social rewards. These findings support the idea that social media engagement involves reward mechanisms, but existing studies primarily rely on self-report methods lacking direct evidence of reward-based learning. Previous RL applications in social media have largely focused on optimizing software, neglecting the underlying psychological mechanisms. While some studies have attempted a quantitative approach to human behavior, the results are inconsistent, highlighting the need for a more comprehensive and rigorous investigation using large-scale datasets and advanced computational models.

This study employed a computational approach using four independent social media datasets, comprising over one million posts from over 4000 individuals across multiple platforms. Study 1 utilized a large dataset of Instagram posts, while Study 2 expanded the analysis to three different topic-focused social media sites to mitigate potential biases associated with economic incentives. A key prediction of reward learning theory is the relationship between response latency (time between posts) and the average rate of rewards. The study used two model-independent approaches to assess reward sensitivity. First, it evaluated whether a hyperbolic function (quantitative law of effect) better explained the relationship between likes and response rates compared to a linear function. Second, it used Granger causality analysis to assess whether the history of likes predicted response latencies. A generative model based on reinforcement learning theory was developed to formally test the reward learning hypothesis. This model assumes agents balance effort costs and opportunity costs to maximize net reward rate, resulting in shorter response latencies with higher reward rates. Model parameters were estimated for each individual, and the model's explanatory power was compared to a null model lacking reward learning. The study also included an online experiment (Study 3) with 176 participants to experimentally manipulate reward rates and observe their impact on response latencies. The experiment involved participants posting memes and receiving varying numbers of 'likes' in a controlled environment. Mixed-effects models were used to analyze the data, accounting for individual differences and potential confounding factors.

The study's key findings consistently demonstrate that social media behavior aligns with reward learning principles. Across all four social media datasets, both model-independent analyses (hyperbolic function and Granger causality) and the reinforcement learning (RL) model indicated a significant association between social rewards and posting behavior. The quantitative law of effect showed that response rates better followed a hyperbolic function of reward rates than a linear one. Furthermore, the Granger causality tests confirmed that past social rewards predicted future posting latencies. The RL model outperformed the null model (without reward learning) in explaining the timing of posts for a substantial proportion of users in all datasets. A higher average reward rate (R) was consistently associated with shorter posting latencies across all platforms and topics, aligning with the RL model's predictions. The magnitude of this effect was stronger for individuals whose behavior was better explained by the RL model. Model simulations accurately reproduced the observed relationship between reward rate and posting latency. Study 2 investigated the potential influence of social comparison on reward learning and found preliminary evidence suggesting its contribution. Unsupervised cluster analysis revealed four distinct computational phenotypes (individual difference profiles) in social reward learning parameters. Finally, the online experiment (Study 3) provided causal evidence that manipulated higher social reward rates led to significantly shorter posting latencies.

The findings provide strong evidence that social media engagement is driven, at least in part, by reward learning mechanisms. The consistency across various platforms and the support from an experimental manipulation solidify the conclusions. The results validate the popular conceptualization of social media as a Skinner Box, emphasizing the role of social rewards in shaping online behavior. These findings have implications for understanding a wide array of online behaviors, including those in dating apps, and contribute to our understanding of the psychological basis of phenomena such as online moral outrage and polarization. The study highlights the relevance of basic, cross-species reinforcement learning principles to understanding complex human social interactions. The RL model, originally developed for non-human animals, successfully explained behavior on time scales far exceeding those typically observed in laboratory settings. The link established between response latencies and average reward rates opens avenues for future research examining the underlying neural mechanisms, potentially involving dopamine pathways. Individual differences in reward learning parameters suggest potential for personalized interventions based on computational phenotypes. While the study focused on the effect of social rewards, future research should incorporate the influence of negative feedback and other motivations such as self-expression and relational development.

This study provides compelling evidence that basic reward learning mechanisms substantially contribute to human behavior on social media. The integration of large-scale data analysis with reinforcement learning models offers a powerful approach to understanding this prevalent form of human interaction. Future research should explore demographic moderators, the influence of negative feedback, the role of action selection in reward maximization, and the detailed mechanisms underlying social comparison in shaping social media engagement. The findings offer a valuable framework for developing theoretically grounded interventions and design solutions targeting diverse aspects of social media use.

The study's reliance on anonymous social media data prevented the researchers from including demographic variables in their analysis, limiting the understanding of potential demographic effects on social media reward learning. While the study focused primarily on 'likes' as a form of social reward, other forms of feedback like comments and shares were not directly considered, potentially influencing the results. The correlational nature of the findings from Studies 1 and 2 warrants caution. Although the experimental manipulation provided causal evidence, the experimental setting might not perfectly replicate the complexity of real-world social media interactions. The participants in the online experiment might have been influenced by their awareness of participating in a study, which could have altered their usual behavior.