Freud and the algorithm: neuropsychoanalysis as a framework to understand artificial general intelligence

The paper advances the thesis that neuropsychoanalysis and affective neuroscience offer a superior paradigm for AGI by integrating mind and brain within a dual-aspect monist framework (after Solms). It argues current AGI research is overly cortex- and cognition-centric, overlooking subcortical affective systems that constitute the basis of intelligence and consciousness. The purpose is to outline design principles for an AGI grounded in Panksepp’s seven basic affective systems and to show how these can be formalized computationally—moving beyond affective computing’s predominantly cognitivist stance. The study positions AGI as possible if built from affect upward, providing a preliminary, programmatic exploration of this approach.

The paper situates its argument against traditional cognitivist AI and affective computing approaches (e.g., Picard), emphasizing that emotions are intrinsic subcortical functions rather than products of higher cognition. It reviews neuropsychoanalysis (Solms, Kaplan-Solms, Turnbull) and dual-aspect monism, linking Freudian metapsychology with contemporary neuroscience. It contrasts Damasio’s homeostasis-linked account of emotions with Panksepp’s affective neuroscience, which posits seven primary emotional systems (SEEKING, RAGE, FEAR, LUST, CARE, PANIC, PLAY) as evolutionarily conserved, subcortical circuits foundational to cognition. The paper references broader debates and critiques of neuropsychoanalysis (e.g., Hobson; Blass & Carmeli) and reviews AGI definitions (Minsky; Russell & Norvig; Shanahan) and philosophical critiques (Dreyfus), setting the stage for a causal modeling alternative (Pearl) to Bayesian associationist frameworks (Friston’s active inference).

This is a theoretical and conceptual methodology. The author: (1) maps Freudian metapsychology (id–ego–superego, Pcpt-cs) to neural substrates per neuropsychoanalysis; (2) adopts Panksepp’s seven primary affective systems as the generative basis for intelligence; (3) analyzes and critiques the Solms–Friston active inference/Markov blanket/free energy framework as insufficient for modeling emotional affects (though suitable for homeostatic and sensory affects); (4) proposes integrating Judea Pearl’s causal analysis—causal diagrams and do-calculus—to model emotional systems as causal networks. The paper identifies a recurring four-phase structure across affective systems: emotion generation (biological reaction), evaluation (valence/feeling), anticipation/prediction (emotional memory and projections), and action. It formalizes an AGI architecture in which data (D) feed hypothesis formation (H), and an inference engine (evaluation, E) uses causal diagrams (CD) and do-calculus (DC) to generate predictions (P), actions (A), and counterfactuals (C). A worked example (SEEKING in a survival scenario) illustrates identifying confounders and computing interventions via do-calculus. The methodology extends to proposing embodiment avenues (biohybrid xenobots) and a functional account of sleep in AI (noise-induced stabilization in SNNs) aligned with free-energy optimization during off-line phases.

- Emotions as primary, subcortical processes should foundationally structure AGI; cognition emerges from affects (Panksepp), challenging cortex-first AGI design. - The Solms–Friston free-energy/active inference framework aptly models homeostatic and sensory affects but is too tied to pleasure–unpleasure and probabilistic association to capture emotional affects’ causal complexity. - Pearl’s causal diagrams and do-calculus better formalize emotional systems as causal networks, enabling interventionist and counterfactual reasoning (ascending Pearl’s ladder of causation). - A common four-phase causal structure (emotion, evaluation, anticipation, action) spans primary affective systems (e.g., SEEKING, RAGE, FEAR, LUST), supporting a unified design pattern for AGI modules. - Counterfactual reasoning (do-calculus) provides a bridge from affective processes to higher cognition and self-awareness, aligning with neuropsychoanalytic views of consciousness emerging from affects. - Sleep-like off-line phases stabilize learning in neuromorphic systems (supporting an analogy to slow-wave/REM processes), suggesting AGI requires primary-process functions (sleep, instincts) to support secondary processes (language, logic). - Embodiment via programmable biological matter (xenobots) offers a pathway to implementing homeostatic and sensory affects, addressing embodiment critiques. - A subcortical AGI aligned with human basic affective systems could improve AI control by sharing core human values and learning schemes, reducing reliance on external rule constraints.

By grounding AGI design in neuropsychoanalytic affective architectures, the paper argues it addresses the central research question: how to conceive AGI that genuinely mirrors human general intelligence. Modeling primary emotional systems as causal networks enables AGI to reason about interventions and counterfactuals, supporting adaptive, value-laden behavior that more closely aligns with human needs. The integration of do-calculus augments active inference, overcoming limits of purely probabilistic association and enabling planning and imagination—key components of general intelligence. Embedding sleep-like stabilization and considering biological embodiment strengthens the system’s robustness and ecological validity. Collectively, these elements suggest a route to AGI that is both neurally and psychoanalytically plausible and potentially more controllable by virtue of shared affective foundations.

The paper concludes that AGI is feasible if constructed from subcortical affective systems upward. It proposes an AGI architecture that instantiates Panksepp’s seven primary emotions as interacting causal networks, uses Pearl’s do-calculus for interventionist and counterfactual reasoning, and complements Solms–Friston active inference for homeostatic regulation. It argues such a subcortical AGI could address AI control by sharing human foundational affective values and learning schemes. Future work includes extensive empirical and comparative research to specify accurate causal diagrams for each affective system, developing processors that generate and update these diagrams autonomously, exploring biological embodiment pathways, and formalizing sleep/regression mechanisms within AGI. The paper suggests an “AI psychoanalysis” may become part of educating and aligning emotionally developing AGI systems.

The work is explicitly preliminary and theoretical, offering broad outlines rather than empirical validation. It depends on contested frameworks (neuropsychoanalysis; Panksepp’s basic emotions) and critiques established models (active inference) without experimental tests. Implementing accurate causal diagrams for subcortical systems requires massive interdisciplinary research on animals and humans. The proposal for biological embodiment (xenobots) is speculative and early-stage. The analogies between sleep in neuromorphic systems and biological sleep remain suggestive. Overall generalizability and practical feasibility are yet to be demonstrated.