Active Inference Account of Substance Use Disorder

The free energy principle (FEP) and active inference framework offer a revolutionary lens for understanding substance use disorder (SUD), reframing addiction not merely as a disease of the brain, but as a complex disorder of inference, decision-making, and environmental interaction. This computational approach conceptualizes addiction as emerging from the brain’s attempts to minimize free energy—a measure of surprise or uncertainty—through maladaptive predictive models that become pathologically entrenched.

Traditional models of addiction have focused primarily on dopaminergic dysfunction and reward learning abnormalities, leading to the widespread adoption of the “disease model.” While this framework explains the compulsive nature of drug-seeking behavior through dysfunctional reward systems, it provides an incomplete picture of the underlying mechanisms. The predictive processing theory within active inference offers a more nuanced understanding: rather than being passive recipients of reward signals, agents actively predict rewarding outcomes and select actions that minimize long-term prediction error.

In this framework, dopamine doesn’t simply signal reward prediction error but instead encodes the precision—or confidence—of affordances, which represent opportunities for action available in the environment. This distinction is crucial for understanding addiction. When individuals use substances, they experience what feels like successful error reduction, but this represents false feedback that creates an illusion of minimizing uncertainty while actually increasing it.

Precision & Uncertainty in Addiction

Recent computational models have identified two seemingly contradictory mechanisms underlying addictive behavior: hyper-precision and hypo-precision. The hyper-precision model suggests that addiction involves excessive confidence in drug-related predictions, creating rigid behavioral patterns focused narrowly on substance use. Individuals become locked into what can be described as fixed-point attractors—predictable behavioral cycles that persist despite their detrimental outcomes.

Conversely, the hypo-precision model proposes that addiction results from reduced model-based control, where individuals show decreased ability to plan and consider long-term consequences. This manifests as a shift from goal-directed behavior to habitual, automatic responses—a transition from thoughtful decision-making to compulsive action. Environmental factors such as stress, poverty, and social instability can exacerbate this by reducing the precision of model-based predictions, pushing individuals toward more immediate, model-free control strategies. Along the same lines, the active inference perspective emphasizes that addiction cannot be understood solely through individual brain dysfunction but must be viewed as a breakdown in the agent-environment system. Environmental unpredictability—whether through inconsistent social support, economic instability, or uncertain drug supply—compromises the brain’s ability to form accurate predictive models. This uncertainty paradoxically doesn’t deter drug use but instead exacerbates harmful patterns by making individuals more reliant on the immediate, albeit false, certainty that substances appear to provide. Research demonstrates that factors like reduced social support and poverty significantly impact addiction and relapse rates, not merely as social determinants but as computational challenges that impair model-based decision-making. When individuals encounter drug-related cues under conditions of high environmental uncertainty, their working memory-intensive processes for reasoning through action-outcome associations become compromised, leading to steep discounting of future outcomes and maintenance of addictive behaviors.

Generative Models And their Clinical Implications

Perhaps most profoundly, the FEP framework reveals how addiction becomes integrated into an individual’s very sense of self. Agents maintain their identity through biological autonomy, which involves fulfilling predictions about their environment through internal generative models—hierarchical prediction systems that guide interaction with the world. In addiction, substance use becomes so central to these generative models that changing behavior requires fundamental alterations to one’s sense of identity. This integration helps explain why addiction is particularly resistant to change and why successful treatment often involves not just addressing substance use but reconstructing fundamental aspects of identity and life structure. The predictability and reliability of substance effects can provide stability in otherwise chaotic circumstances, making the drug-related lifestyle feel necessary for maintaining coherent selfhood.

The Bayesian inference framework embedded within the FEP provides crucial insights into why addiction persists despite obvious negative consequences. Rather than viewing addictive behavior as fundamentally irrational, this approach suggests that such behavior may be optimal given the individual’s generative model of the world. The apparent irrationality emerges not from broken inference mechanisms but from maladaptive generative models that have become entrenched through repeated experience. This perspective shifts focus from presumed optimal behavior to understanding the specific parameters and characteristics of individual generative models that lead to pathological outcomes. For instance, an individual’s generative model might accurately predict that drug use will provide immediate relief from withdrawal symptoms, even while failing to adequately weight longer-term negative consequences.

The FEP approach to addiction suggests several novel therapeutic directions. Rather than focusing solely on blocking reward pathways or teaching willpower, treatments might target the precision of different types of predictions. This could involve therapies aimed at reducing the precision assigned to drug-related affordances while simultaneously strengthening model-based control systems and environmental predictability. Interventions might also focus on reconstructing generative models through carefully structured experiences that challenge pathological predictions and support the development of more adaptive behavioral repertoires. This could include creating therapeutic environments that provide the reliability and predictability that substances falsely promised, while gradually expanding the range of behaviors through which individuals can achieve successful error reduction. The framework also suggests the importance of addressing environmental factors that contribute to uncertainty and stress, which compromise model-based decision-making. Policy implications include ensuring access to stable housing, healthcare, and social support—not merely as social goods but as computational necessities for healthy inference and decision-making.

Conclusions

The free energy principle offers a sophisticated and integrative framework for understanding addiction that moves beyond simple disease models toward a comprehensive account of how individuals interact with their environment through prediction and action. This approach recognizes addiction as neither purely biological nor purely social but as an emergent property of complex systems trying to maintain coherence in uncertain environments. While this theoretical framework is promising, significant challenges remain in translating these insights into practical clinical applications. The mathematical complexity of these models and the deeply embedded nature of addictive patterns in identity and social context make intervention challenging. Nevertheless, this approach offers hope for more effective treatments that address the fundamental computational and inferential processes underlying addictive behavior, potentially leading to more personalized and effective therapeutic approaches. The key insight from this computational perspective is that addiction represents not a failure of willpower or moral character, but a predictable outcome of particular generative models operating in specific environmental contexts. Understanding these mechanisms offers pathways toward more compassionate, effective, and scientifically grounded approaches to treatment and prevention.

Aspect	Traditional View	FEP/Active Inference View
Core Problem	Reward system dysfunction	Inference and prediction errors
Dopamine Role	Reward prediction error	Precision of affordances/confidence
Addiction Mechanism	Hijacked reward circuits	Maladaptive generative models
Environmental Factors	External stressors	Computational challenges to inference
Treatment Focus	Block rewards, teach control	Reconstruct predictive models, reduce precision
Identity	Separate from addiction	Integrated into self-model

References

• Holmes J. Friston’s free energy principle: new life for psychoanalysis? BJPsych Bull. 2022;46(3):134-139

• Smith R, et al. Imprecise Action Selection in Substance Use Disorder. Front Hum Neurosci. 2020;14:366

• Miller M, et al. Embodying addiction: A predictive processing account. Brain Cogn. 2020;138:105495

• Zhang Z, et al. An Overview of the Free Energy Principle and Related Research. Neural Comput. 2024;36(5):713-757

• Schwartenbeck P, et al. Optimal inference with suboptimal models: Addiction and active Bayesian inference. Med Hypotheses. 2015;84(2):109-117

• Haarsma J, et al. Precision weighting of cortical unsigned prediction error signals benefits learning and is impaired in psychosis. Mol Psychiatry. 2021;26(9):5320-5333

• Li L, et al. Formalizing Lacanian psychoanalysis through the free energy principle. Front Psychol. 2025;16:1574650

• Smith R. Active Inference and its Application to Empirical Data. Yale Medicine Psychiatry. 2025

• Blum K, et al. Introducing Precision Addiction Management of Reward Deficiency Syndrome. Front Psychiatry. 2018;9:548

• Albarracin M. Addiction as a Disruption of the Markov Blanket. Glob J Addict Rehabil Med. 2024;7(4):555717

• Takahashi H, et al. Dopamine D1 Receptors and Nonlinear Probability Weighting in Risky Choice. J Neurosci. 2010;30(49):16567-16572

• Friston K, et al. Computational psychiatry: from synapses to sentience. Mol Psychiatry. 2023;28(1):256-268

• Chakroun K, et al. Dopamine regulates decision thresholds in human subthalamic nucleus. Nat Commun. 2023;14(1):5506

• Taylor S, et al. Active learning impairments in substance use disorders. Drug Alcohol Depend. 2023;252:110932

• Groman S, et al. Model-free and model-based influences in addiction-related behaviors. Biol Psychiatry. 2019;85(11):936-945

• Doyle W, et al. Environmental Uncertainty and Substance Use Disorders. Perspect Behav Sci. 2023;46(1):191-213

• Robinson A, et al. Model-based and model-free mechanisms in methamphetamine use disorder. Addiction. 2024;119(3):456-466

• Chen H, et al. Model-Based and Model-Free Control Predicts Alcohol Use Trajectories. Biol Psychiatry. 2021;89(10):980-990

• FitzGerald T, et al. Dopamine, reward learning, and active inference. Front Neurosci. 2015;9:136

• Friston K, et al. Dopamine, Affordance and Active Inference. PLoS Comput Biol. 2012;8(1):e1002327

• The anatomy of choice: active inference and agency
• Embodying addiction: A predictive processing account
• Optimal inference with suboptimal models: Addiction and active Bayesian inference
• Pathologies of precision: A Bayesian account of goals, habits, and episodic foresight in addiction
• Imprecise Action Selection in Substance Use Disorder: Evidence for Active Learning Impairments When Solving the Explore-Exploit Dilemma
• Active inference and learning