Growing up in the 1980s in Santa Cruz, California, where redwood-covered mountains descend to the rocky edge of the Pacific, might sound idyllic. But in the dark wake of the drug-fueled ’70s, the beach town could also be frightening. There was a bully at my high school who once chased me down the street threatening to hurt me. Unsurprisingly, catching sight of him in the hallways or at the skate park filled me with dread. Just walking past his house would trigger a wave of anxiety. Yet if I saw him in class, with teachers present, I felt more at ease.
How did my brain know to fear him only in specific circumstances? More broadly, how did I infer emotional significance from the world around me? The fact that I or anyone can make these judgments suggests that emotion arises from an internal model in the brain that supports inference, abstraction and flexible, context-dependent evaluations of threat or safety. These model-based emotion systems helped me infer danger from otherwise innocuous features of the environment, such as the bully’s house, or to downgrade my alarm, as I did when an adult was present.
Understanding the neural basis of emotion is a central question in neuroscience, with profound implications for the treatment of anxiety, trauma and mood disorders. Yet the field remains divided over what emotions are and how they should be defined, limiting progress. On one side are neurobiologists focused on the neural underpinnings of simple learned and innate defensive behaviors. On the other are psychological theorists who view emotions as subjective experiences arising from complex conceptual brain models of the world that are unique to humans.
This divide fuels persistent arguments over whether emotion should be defined primarily as a conscious state or not. Though subjective feelings are undeniably important, limiting our definitions to conscious phenomena prevents us from studying the underlying mechanisms in nonhuman species. To move forward, we need to identify the conserved neural processes that support higher-order, internal-model-based emotional experiences across species, regardless of whether they rise to consciousness.
H
Starting in the 2010s, though, cracks in this approach became increasingly apparent. Some worried that using the term “fear” implied that rodents experienced human-like subjective states, when the experiments truly probed only simple associative learning and defensive reactions, such as freezing. Moreover, despite remarkable mechanistic advances, clinical translation remained limited. Many disabling features of human anxiety and trauma disorders reflect dysfunction in inferential, model-based processes, such as persistent, overgeneralized fear and failures to adapt to different contexts, and not simply disrupted stimulus-outcome learning. When fireworks trigger fear in someone with post-traumatic stress disorder, the issue is not that they incorrectly learned to fear fireworks; it’s that their internal model can’t discern the safe context of a local fair.
Most importantly, a gulf persisted between how human and rodent researchers viewed emotion. Influential psychological theories increasingly framed emotions as individualized, subjective experiences arising from complex, human-specific conceptual brain models that integrate sensory, interoceptive, cognitive and cultural information to generate emotional meaning. These accounts capture phenomenology but have been difficult to map onto specific neural mechanisms. Moreover, because of their focus on conscious human emotional experience, these theories cannot be studied in animals.
So how can we move forward as a field? A broader perspective focused on model-based emotion systems offers a middle ground. These systems provide an intermediate representational layer—one that is more abstract than simple associations yet not dependent on conscious experience or complex human specific capabilities. They may not generate subjective feelings directly, but they enable an organism to draw conclusions from incomplete information—to infer danger from a bully’s house, for example, or to modulate emotional responses depending on context.
Crucially, cross-species examination of model-based aspects of emotion has already begun. Experimental studies in rodents and primates have identified brain regions involved in emotional inference and generalization. Work from my laboratory, for example, leveraged a sensory preconditioning task as a tool to probe emotional inference, showing that rodents can learn to feel threatened by a stimulus that was never directly associated with an aversive experience, much like my own learned fear of my bully’s house.
Critically, these behaviors were supported by a model-based representation encoded in the medial prefrontal cortex that links both direct and indirect predictors of threat with the aversive experience. Notably, this model-based system coordinates simple associative networks in the amygdala that are traditionally thought to govern fear conditioning. Research on interoception highlights the role of the brainstem and association cortices in integrating bodily state with external information, a central feature of emotional models in human theories.
Other studies demonstrate that emotional states can be sustained over time, supporting an internal context that helps animals evaluate the environment over longer timescales. From the theoretical side, conceptual models have proposed similar intermediary cognitive-like representations in emotion systems. Computational frameworks provide an organizing principle for this work, modeling emotional processes as forms of hidden-state inference, in which an organism’s response to threat or reward is contingent upon its understanding of the context; my fear of the bully, for example, depends on whether I believe I’m in a safe or dangerous environment.
A
Similarly, paradigms that examine how defensive responses or decision-making change across aversive or safety states offer tractable models of context-dependent emotional processing. At the neural level, evidence for model-based representations can come from asking whether neural activity reflects inferred latent states or associative structure, as captured by quantitative computational models. Importantly, these approaches lend themselves naturally to cross-species designs, enabling direct comparisons of higher-order emotional processing.
There is strong precedent for distinguishing between conscious experience and model-based strategies in other areas of neuroscience. Vision research has revealed how hierarchical circuits move from detecting simple sensory features to encoding abstract representations of objects and scenes. Decision-making research has mapped conserved model-based control circuits in the basal ganglia and cortex without requiring any assumption about conscious deliberation.
Emotion research can follow similar principles. We don’t need to resolve philosophical debates about consciousness to make progress. Instead, we must commit to identifying the relevant computations for model-based emotional processing and their underlying neural mechanisms across species—using careful task design and powerful genetic, optogenetic, imaging and computational tools.
Emotion, like vision, may be organized hierarchically, with increasingly abstract functions emerging from interactions among the brainstem and the hypothalamic, amygdala and cortical circuits. Using cross-species tasks that isolate specific higher-order emotional computations will be key to uncovering shared circuitry and mechanisms across species. In parallel, researchers can examine how uniquely human forms of emotion and consciousness arise from cortical expansion and how these elaborations interact with conserved circuits to expand the human emotional repertoire.
This framework also offers new translational opportunities. Emotional models are not fixed; they are gradually sculpted across development through learning, stress and social experience. Understanding how early-life adversity distorts emotional inference, and how neuromodulation and interoception interact with internal models, may illuminate pathways to better treat anxiety, trauma and mood disorders. Such an approach could yield new treatments that target the mechanisms responsible for model-based emotional processing instead of simple associative learning mechanisms.
