Yael Niv is a professor of neuroscience and psychology at Princeton University. Her lab studies the computational processes underlying reinforcement learning, focusing on how attention, memory and learning interact to construct task representations that allow efficient learning through optimal generalization. She is co-founder and co-director of the Rutgers-Princeton Center for Computational Cognitive Neuropsychiatry, where she is applying ideas from reinforcement learning to understanding and treating mental health. She is also the director of a Latent Cause Inference Conte Center joint between Princeton and Rutgers. Her proudest career accomplishment is winning a graduate mentoring award. In her nonexistent spare time, she is a mom to two awesome boys, and an activist within and outside academia.

The vast amounts of human-generated data available from the internet, government agencies, and other sources present exciting possibilities for understanding how people decide in the real world. In this talk I discuss two different lines of work which highlight how the analysis of large naturalistic datasets can be used to circumvent constraints of laboratory experimentation. For example, a recent body of work suggests that fluctuations in mood states are driven by unpredictable outcomes in daily life, but also appear to drive consequential behaviors such as risk-taking. By analyzing day-to-day mood language extracted from several million social media posts, we find that real-world ‘prediction errors’—local outcomes that deviate positively from expectations—predict day-to-day mood shifts observable at the level of a city, which in turn predicted increased per-person lottery gambling rates, elucidating the interplay between prediction errors, moods, and risk attitudes. In a newer line of work examining a massive dataset of UK grocery store purchases, we examine real-world evidence for decoy effects—whereby people’s choices between two valuable “target” options are swayed by a third, objectively inferior, “decoy” option— in consumer choice. When comparing products with attributes that effectively trade off (e.g., quality and price), the presence of inferior, decoy options systematically biased the likelihood of choosing a specific target option. This study provides a proof of principle demonstrating that these sorts of context effects are detectable in richer, complex real-world consumer choice settings with multiple decoy options, underscoring how real-world, naturalistic datasets afford unique understanding of consequential choice behaviors.

Achieving human-AI alignment in complex multi-agent games is crucial for creating trustworthy AI agents that enhance gameplay. We propose a method to evaluate this alignment using an interpretable task-sets framework, focusing on high-level behavioral tasks instead of low-level policies. Our approach has three components. First, we analyze extensive human gameplay data from Xbox's Bleeding Edge (100K+ games), uncovering behavioral patterns in a complex task space. This task space serves as a basis set for a behavior manifold capturing interpretable axes: fight-flight, explore-exploit, and solo-multi-agent. Second, we train an AI agent to play Bleeding Edge using a Generative Pretrained Causal Transformer and measure its behavior. Third, we project human and AI gameplay to the proposed behavior manifold to compare and contrast. This allows us to interpret differences in policy as higher-level behavioral concepts, e.g., we find that while human players exhibit variability in fight-flight and explore-exploit behavior, AI players tend towards uniformity. Furthermore, AI agents predominantly engage in solo play, while humans often engage in cooperative and competitive multi-agent patterns. These stark differences underscore the need for interpretable evaluation, design, and integration of AI in human-aligned applications. Our study advances the alignment discussion in AI and especially generative AI research, offering a measurable framework for interpretable human-agent alignment in multiplayer gaming.

Prediction errors (PEs)—the difference between expected and actual outcomes—are thought to be critical drivers of learning and emotion, shaping momentary affective reactions, future expectations, and longer-term mood states. While theories of affect propose that emotions arise from outcome valence and PEs, most evidence comes from low-stakes laboratory studies, which may lack the ecological validity needed to capture the richness and complexity of real-world emotional experiences. To address this, we developed an event-triggered procedure to measure emotion and expectation updating in university students receiving goal-relevant exam grades—a high-stakes, personally meaningful event that elicits robust and sustained emotional responses. Using real-world contexts, we demonstrate that PEs are strong predictors of emotion and significantly influence long-term mood, with their effects persisting for up to two weeks. These findings underscore the importance of deviations from expectations in shaping affect in daily life and highlight the critical role of ecologically valid paradigms in understanding emotion.

Real-world, high-stakes events also provide a powerful framework for building computational models to understand how psychopathology affects emotional experience. In contrast to null results from some laboratory studies, we find that in real-world settings—where outcomes are deeply consequential and emotionally engaging—people with depression display a specific reduction in affective reactivity to positive PEs. Notably, there are no differences in responses to negative PEs or outcome values, implicating impaired positive PE processing as a key mechanism in emotional dysregulation. Together, these findings demonstrate that PEs are central to the dynamics of real-world emotion, linking momentary reactions to longer-term affective states. They also highlight the necessity of studying emotion in ecologically valid contexts to uncover meaningful insights into both typical and disordered emotional processing.

Social media has become a central arena for human interaction, with billions of daily users worldwide. The intense popularity of social media is often attributed to a psychological need for social rewards (“likes”), portraying the online world as a Skinner Box for modern humans. Despite such portrayals, empirical evidence for social media engagement as reward-based behavior have been limited. In this talk, I will focus on a study where we applied a computational approach to test whether reinforcement learning mechanisms shape social media behavior in naturalistic settings. Analyzing over one million posts from more than 4,000 individuals across multiple platforms, we used reinforcement learning models to show that human behavior on social media conforms qualitatively and quantitatively to reward learning principles. Specifically, we found that users adaptively space their posts to maximize their rate of social rewards, balancing effort costs and opportunity costs. Furthermore, we identified meaningful individual differences in social reinforcement learning profiles, and through an online experiment, we provided causal evidence that social rewards directly influence behavior in line with our computational model. This work was among the first to establish large-scale, real-world evidence for reinforcement learning dynamics in social media engagement. I will also discuss some challenges of using naturalistic data for RL modeling and possible future directions for studying real-world sequential decision-making in dynamic social environments.