GPT

Depression, Anxiety, and the FEP Lens

Under the FEP, depression and anxiety can be framed as maladaptive states where the system gets stuck in patterns that minimize surprise in ways that are pathological.

• Depression as Pathological Certainty
  Depression often manifests as a rigid, low-energy state, where the brain generates predictions that the world is futile, effort is meaningless, and change is unlikely. From an FEP perspective, this could be seen as the brain settling into a low-surprise attractor state—a cognitive and affective landscape where expectations of failure and suffering become self-fulfilling. Rather than engaging with the world dynamically, the brain “learns” to expect negative outcomes and ceases to update its priors effectively. The system essentially overfits to a pessimistic model of the world.

• Anxiety as Hyperactive Surprise Avoidance
  If depression is an excessive commitment to low-energy, predictable states, anxiety is its inverse: a state of heightened sensitivity to prediction errors, where the system is constantly scanning for threats and generating worst-case scenarios. Instead of allowing prediction errors to be updated properly, anxiety loops keep the system in a state of hypervigilance, where the cost of being wrong is perceived as catastrophically high. It’s not that the system is bad at prediction—it’s that it predicts too much, too often, without resolving uncertainty in productive ways.

Simulated Neural Networks and Reinforcement Learning

Recent work in reinforcement-learning-based artificial neural networks (especially those incorporating predictive coding principles) provides a compelling model for how depressive and anxious states might emerge in biological systems. Some key insights:

1. Overfitting and Learned Helplessness:
   If a reinforcement-learning model is trained on highly negative or unpredictable environments, it can learn to expect bad outcomes regardless of its actions, leading to policy freezing—the equivalent of behavioral inhibition in depression. The system essentially decides that action is not worth the energy cost.

2. Hyperpriors and Anxiety:
   In neural networks trained to minimize uncertainty at all costs, a form of hypervigilance emerges, where the model constantly anticipates threats. This mirrors anxiety, where the system biases itself toward erring on the side of caution—even at the cost of excessive stress and reduced exploratory behavior.

3. Dopaminergic Weighting and Motivation:
   Simulated models that incorporate dopaminergic-like reinforcement signals show that if prediction errors aren’t properly rewarded or punished, learning stagnates. In depression, this can be likened to anhedonia—the failure to update reward expectations, making everything feel flat and unmotivating.

What This Means for Understanding Mental States

FEP-based approaches suggest that depression and anxiety aren’t just chemical imbalances or cognitive distortions but deeply entrenched ways of modeling the world. The reinforcement-learning analogy highlights how these states emerge when the system overcommits to certain priors (in depression’s case, low-value expectations; in anxiety’s case, excessive threat anticipation) and fails to recalibrate properly.

This also raises an interesting ethical and existential question: If our mental suffering is the result of overly rigid or maladaptive priors, what does that say about the narratives we construct about ourselves? Do we suffer because we are caught in an epistemic trap—a loop of self-confirming beliefs that feel inescapable not because they are true, but because they minimize surprise? And if so, how do we update a model that no longer believes updating is possible?