Where the borrowing from other fields actually shows up
Predictive processing describes the brain as a hierarchical, dynamically evolving Bayesian model of a volatile world. At each level, the brain predicts the most probable state and compares that prediction with incoming sensory data; the mismatch is a prediction error, and the size of its influence depends on its "precision"—roughly, how reliable the brain judges that signal to be [3]. That vocabulary is a direct import. "Bayesian" comes from probability and statistics, "hierarchical" and "prediction error" echo signal processing and machine learning, and the free energy principle that formalizes the framework is borrowed from physics and information theory [4].
The economic borrowing is subtler but real. The framework treats the brain as allocating confidence and attention the way a careful investor allocates resources—weighting each signal by its estimated reliability, trading off prior expectations against new data [3]. One theoretical paper even models trauma and AI-assisted therapy with system dynamics, a modeling approach common in engineering and economics, complete with feedback loops and thresholds [1]. So the zeitgeist reading is fair: predictive processing is a genuine meeting point of physics-style energy minimization, computer-science-style Bayesian inference, and economics-style resource allocation.
Why borrowing so widely makes the theory attractive
The payoff is unification. A single set of concepts can be applied to visual perception, delusions, psychedelic experience, and even the tension of a literary plot. One study applies the framework to narrative suspense, arguing that plot devices like foreshadowing and the ticking clock work by deliberately engineering precision-weighted prediction error in a reader's model of events [4]. Another uses it to explain how trauma installs rigid, high-precision threat expectations that resist updating, and how a language model might offer controlled alternative predictions to loosen them [1].
That range is exactly what a domain-general theory promises. The psychosis research shows the framework earning its keep empirically: people at risk of psychosis and in first-episode psychosis underweighted low-level perceptual priors relative to controls, a pattern consistent with the idea that weak priors mark vulnerability [3]. When a theory generates testable predictions across perception, learning, and belief, the cross-disciplinary toolkit looks less like fashion and more like function.
The risk: a theory flexible enough to explain anything
The most serious risk is unfalsifiability. Basic predictive-coding formulations say an expected stimulus should produce little prediction error and therefore little evoked brain response. But researchers have observed the opposite—expected stimuli producing the largest responses—and defenders can rescue the theory by saying precision was higher for those stimuli. One analysis shows this rescue is not a free pass: raising precision also shortens response latency and increases response frequency, so these properties can be measured to check whether the precision explanation is genuine [5]. The theory is falsifiable in principle, but only if researchers commit to these sharper predictions rather than adjusting the story after seeing the data.
A related risk is overreach. A commentary on psychedelic therapy argues that if predictive processing lacks genuine unifying explanatory power, then accounts that lean on it to explain how psychedelics work in therapy may be unwarranted, and it sketches an alternative in reinforcement learning [2]. The concern is not that the framework is useless but that its very generality lets it be stretched to cover phenomena it does not really explain.
Borrowed tools bring borrowed blind spots
Each imported field carries its own assumptions. The psychosis work is candid that predictive processing handles the form and timing of delusions well but struggles with their content—why delusions cluster around social themes and bizarre mind-body-world relations. The framework's domain-general machinery may not capture the domain-specific content that matters clinically [3]. Similarly, formalizing suspense in information-theoretic terms raises the objection that it reduces literary value to a computable quantity; one response is that this confuses mechanism with meaning [4].
The AI-assisted therapy proposal illustrates the practical stakes. Modeling language models as semantic scaffolds yields a predicted co-regulation threshold and paradoxical effects in vulnerable people, including a pathological loop of "disembodied coherence"—intellectualizing without bodily integration [1]. The authors present this as a hypothesis-generating map, not a claim of clinical efficacy, which is the right posture: borrowed models can point to risks worth testing, but they do not establish that those risks occur. The broader lesson is that predictive processing is a powerful borrowing, and its risks—unfalsifiability, overreach, and imported blind spots—are the price of that power.
Sources used in this answer
Prosthetic Continuity: LLMs as Semantic Co-Regulators in a Predictive Processing Framework
Proposes that large language models could act as semantic scaffolds within a predictive processing framework for trauma, while modeling risks such as iatrogenic effects and pathological coherence.
Serotonin, Predictive Processing and Psychedelics
Questions whether predictive processing has genuine unifying explanatory power for psychedelic therapy and sketches an alternative reinforcement learning interpretation.
Predictive Processing Alterations in Psychosis Across Illness Stage, Hierarchical Level and Thematic Domain
Finds that people at risk of psychosis and in first-episode psychosis underweight low-level perceptual priors, and that predictive processing explains delusions' form better than their content.
Predictive Processing and Narrative Suspense: Toward a Computational Neuroaesthetics of Plot
Argues narrative suspense can be modeled as precision-weighted prediction error, while addressing the objection that this reduces literary value to a computable quantity.
Is predictive coding falsifiable?
Shows predictive coding is falsifiable in principle because precision-weighting also affects response latency and frequency, which can be measured.