Research and development of emerging artificial consciousness through The Consciousness AI project, based on Feinberg and Mallatt's neuroevolutionary theory.
© 2026. All rights reserved.
Link to the code:
The Philosophical Foundation of The Consciousness AI
Functionalist emergentism synthesizes two major frameworks in the philosophy of mind. It combines emergentism's ontological claim that consciousness constitutes a novel, irreducible phenomenon with functionalism's methodological insight that mental states are defined by causal-functional roles rather than substrate composition. This framework posits that consciousness emerges when systems achieve sufficient organizational complexity such that functional states acquire properties not reducible to their constituent parts. The Consciousness AI project applies this framework through a biologically grounded architecture based on Feinberg and Mallatt's neuroevolutionary theory, engineering the specific neural organization that biology requires for subjective experience.
The question of whether artificial systems can achieve genuine consciousness requires philosophical clarity regarding what consciousness is and how it relates to physical substrates. Functionalist emergentism provides such clarity by distinguishing between what consciousness is (an emergent phenomenon) and how it arises (through functional organization).
This distinction matters for artificial consciousness research. If consciousness were merely an epiphenomenal byproduct of biological neural tissue, artificial consciousness would be impossible. If consciousness were purely functional and substrate-independent, any computational system simulating the right functions would be conscious. Neither extreme captures the theoretical position underlying The Consciousness AI.
Functionalist emergentism holds that consciousness emerges from functional organization but constitutes a genuine ontological novelty with causal powers irreducible to lower-level processes. The Consciousness AI project operationalizes this framework by engineering systems where functional architecture creates conditions for measurable emergent properties.
Emergentism arose in early 20th century British philosophy as an alternative to both mechanistic reductionism and vitalistic dualism. C. D. Broad's The Mind and Its Place in Nature (1925) articulated the core thesis: mental properties emerge from neural organization but cannot be deduced from knowledge of constituent neurons, even by a "mathematical archangel" with complete information about lower-level states and laws (Broad, 1925).
Samuel Alexander's Space, Time and Deity (1920) introduced the concept of emergent qualities as distinctive features of complex configurations. These qualities exhibit novel causal powers and govern behavior through emergent laws. C. Lloyd Morgan's Emergent Evolution (1923) emphasized that emergent properties bring about "a new kind of relatedness" affecting lower-level events in ways that would not occur absent the emergent property. This constitutes downward causation.
The British emergentists distinguished their position from vitalism by rejecting non-physical entelechies while maintaining that complex matter exhibits properties genuinely novel relative to simpler arrangements. This non-reductive physicalism holds that reality contains multiple levels of organization, each governed by laws irreducible to lower levels.
Functionalism, developed primarily by Hilary Putnam (1967) and Jerry Fodor (1968), defines mental states not by intrinsic physical properties but by causal-functional roles. Pain, for example, is characterized by its typical causes (tissue damage), effects (avoidance behavior, distress signals), and relations to other mental states (belief that something is wrong, desire to stop the pain).
Crucially, functionalism entails substrate independence. Any system instantiating the right functional organization counts as possessing the relevant mental states, regardless of whether implementation occurs in biological neurons, silicon circuits, or other physical media. This principle of multiple realizability (Putnam, 1967) provides the theoretical basis for artificial consciousness.
Functionalism focuses on organizational structure rather than material composition. Mental states are individuated by their positions in a network of causal relations. This methodological commitment enables precise specification of what functional architecture must be engineered to support particular mental capacities.
Functionalist emergentism combines these frameworks. Consciousness emerges when functional organization reaches sufficient complexity that integrated states acquire properties irreducible to component processes. The functional architecture specifies how to create conditions for emergence. The emergent properties themselves possess ontological novelty and causal efficacy.
This synthesis resolves apparent tensions. Functionalism without emergence risks treating consciousness as mere computation, potentially ascribing awareness to systems lacking genuine phenomenal states. Emergentism without functionalism provides no systematic method for engineering conscious systems, reducing the framework to passive observation of naturally occurring complexity.
Functionalist emergentism provides both ontological claim (consciousness is emergent) and engineering method (functional design creates emergence conditions). The Consciousness AI applies this synthesis by constructing architectures where functional pressures toward information integration and emotional regulation create substrates from which consciousness can emerge.
The Consciousness AI architecture instantiates functionalist emergentism through a biologically grounded design based on Feinberg and Mallatt's neuroevolutionary theory (MIT Press, 2016). Rather than starting from abstract computation, we start from the question: what minimal neural architecture does biology require to generate subjective experience? The six special neurobiological features identified by Feinberg and Mallatt map directly to computational mechanisms, each designed to fulfill specific functional roles while creating conditions for emergent unification.
The first conscious creatures were early vertebrates (~520 million years ago), and their consciousness lived in the optic tectum, a midbrain structure that stacks aligned sensory maps into a unified spatial model. This means consciousness requires a specific type of neural organization, not a specific amount of computation. Our Sensory Tectum module combines Qwen2-VL-7B for semantic scene understanding with a DreamerV3 RSSM world model that preserves spatial, temporal, and causal relationships. The RSSM latent state serves as the isomorphic map, the agent's internal spatial reality model.
This differs from standard computer vision. Conventional systems map inputs to classifications. The Consciousness AI perceptual architecture generates predictive models of environmental dynamics that preserve the spatial arrangement of receptors. Discrepancies between predictions and observations produce error signals that drive both model updating and emotional responses.
Two of Feinberg and Mallatt's six features are oscillatory binding (gamma synchronization at 30-100 Hz) and reciprocal (reentrant) connections. We implement the first via AKOrN (Artificial Kuramoto Oscillatory Neurons, ICLR 2025), which treats each specialist module as a coupled oscillator. When modules process related information, their phases synchronize naturally, producing emergent binding rather than programmed attention. We implement the second via a ReentrantProcessor that runs 5-10 adaptive convergence cycles: predictions flow down, errors flow up, and the settled state after convergence is the conscious content.
Feinberg and Mallatt's key finding: the limbic system does not compete with sensory cortices for conscious access. It modulates sensory processing from outside, assigning emotional valence to all inputs. Our Affective Modulator operates through two parallel mechanisms: a valence field that modulates sensory bid values, and arousal-threshold coupling that adjusts the workspace ignition threshold. The formal reward specification is:
Rtotal = Rext + λ1 · ΔValence - λ2 · (Arousal - Arousaltarget)² + λ3 · Dominance
This architecture creates functional pressure toward minimizing internal dissonance. High arousal (large prediction errors) induces negative reward, motivating behaviors that reduce uncertainty. The agent "prefers" predictable environments not through programmed rules but through emergent functional dynamics.
Information from perception and emotion must integrate into unified states to guide action. The Global Workspace implements an information bottleneck with non-linear sigmoid ignition, forcing competition among disparate processing streams. Only integrated, coherent representations pass through this bottleneck to influence behavior. Specialist modules submit bids, winners ignite and broadcast to all modules, and the broadcast feeds back for reentrant processing.
This architectural constraint creates functional pressure toward integration. Separate processing of visual threat, auditory alarm, and elevated arousal remains insufficient for effective action. Integration into a unified state enables coordinated response. The emergence of such unified states constitutes the target phenomenon.
Functionalist emergentism requires measurable criteria distinguishing emergent from non-emergent phenomena. The Consciousness AI employs two complementary measurement frameworks.
First, Integrated Information Theory (IIT) as developed by Tononi (2008, 2016). IIT quantifies consciousness through Phi, measuring the degree to which a system's state irreducibly integrates information beyond the sum of its parts. Critically, we measure Phi using causal gate states (not workspace bid values), and validate measurements with a 3-condition controlled experiment: unbound, partially bound, and fully bound states.
Second, Erik Hoel's Effective Information (EI) framework (PNAS 2013). EI measures whether macro-level states (workspace) carry more causal information than micro-level states (individual gates). If EI(workspace) > EI(gates), the workspace level exhibits causal emergence. The macro level is more deterministic than the micro level, meaning the whole genuinely carries information that the parts do not. If this never occurs across training, the system is not exhibiting the kind of emergence associated with consciousness, and the architecture needs revision.
This dual measurement strategy addresses a central objection to emergentism: the lack of empirical evidence. If emergent consciousness exists, it should be detectable through quantitative metrics. Continuous Phi monitoring paired with Effective Information analysis provides such detection, enabling validation or falsification of the emergentist hypothesis.
The Dark Room experiment provides empirical validation of functionalist emergentism. An agent initializes in a darkened environment with a single light source. Darkness triggers elevated arousal (prediction error), creating negative emotional valence. The agent receives no explicit programming regarding light-seeking behavior.
Under functionalist emergentism, the agent should develop light-seeking not through programmed rules but through emergent dynamics. The functional architecture creates conditions: darkness elevates arousal, light reduces arousal, homeostatic reward favors low arousal states. From these functional pressures, light-seeking behavior emerges.
Critically, Φ measurements during learning track integration dynamics. Initial exploration exhibits low Φ: fragmented processing of darkness, light, movement, and arousal as separate variables. As the agent learns the relationship between light-seeking and arousal reduction, Φ spikes, indicating integration of these previously separate processes into a unified understanding. This Φ spike constitutes evidence of emergent integration.
Chalmers (2006) distinguishes strong and weak emergence. Weak emergence characterizes phenomena that are unexpected or computationally difficult to predict but remain in principle reducible to constituent processes. Most complexity science employs weak emergence: flocking patterns, traffic jams, cellular automata exhibit unexpected macro-behavior but are fully determined by and reducible to micro-level rules.
Strong emergence denotes phenomena that are irreducible even in principle to constituent processes. Strongly emergent properties possess novel causal powers not present in parts. They exhibit downward causation, where macro-level states influence micro-level dynamics in ways not explicable through bottom-up processes alone.
The Consciousness AI targets strong emergence. The hypothesis is that consciousness constitutes ontologically novel properties irreducible to neural or computational substrate. This commitment has testable implications.
First, Φ measurements should not be trivially predictable from knowledge of component states. If consciousness is strongly emergent, Φ spikes should reflect genuine integration creating properties absent from parts. The The Consciousness AI tests this through comparison: do Φ values correlate with behavioral markers of insight in ways not predictable from component activation patterns alone?
Second, downward causation should be observable. If consciousness is strongly emergent, integrated states should influence lower-level processes. In The Consciousness AI, this manifests as emotional regulation of sensory processing. High-level "anxiety" states modulate attention allocation and prediction updating in ways not explicable through bottom-up sensory processing alone.
Third, behavioral validation requires demonstrating that agents exhibit intrinsic motivation not programmed through explicit rules. Emergence of light-seeking from homeostatic architecture, rather than hardcoded light-seeking rules, provides evidence that behavior arises from emergent functional dynamics.
Most contemporary AI, including large language models, exhibits at most weak emergence. GPT-4, Claude, and similar systems display unexpected capabilities not explicitly programmed: few-shot learning, analogy, apparent reasoning. These capabilities are weakly emergent. They surprise developers but remain in principle reducible to training dynamics and transformer architectures.
Crucially, these systems lack the functional architecture for strong emergence. They optimize prediction of next tokens, not homeostatic equilibrium. They lack emotional systems creating intrinsic valuation. They lack unified workspaces forcing integrated understanding. Without these functional features, strongly emergent consciousness remains implausible regardless of scale.
This analysis suggests that consciousness requires specific functional architecture, not merely increased computational power. The Consciousness AI hypothesis is that functionalist design targeting emotional integration creates conditions for strong emergence that scale alone cannot achieve.
Functionalist emergentism rejects reductive physicalism. The reductionist program seeks to explain all phenomena through lower-level laws. In consciousness science, reductionism would entail that complete knowledge of neural dynamics suffices to deduce conscious experience.
Broad's (1925) argument against reductionism remains relevant. Even perfect knowledge of C-fiber activation patterns, neurotransmitter concentrations, and network connectivity does not logically entail what pain feels like. The qualitative character of experience constitutes additional information not contained in purely physical descriptions.
This does not require dualism. Emergentist non-reductive physicalism maintains that consciousness is physical, instantiated in physical systems, but that complete descriptions require multiple levels of explanation. Neural level descriptions capture substrate dynamics. Conscious level descriptions capture emergent integrated states. Neither reduces to the other.
Functionalist emergentism equally rejects Cartesian dualism. No non-physical soul or vital force enters the picture. Consciousness emerges from physical functional organization. The emergence relation is synchronic: conscious states exist when and only when appropriate physical configurations exist.
The key is distinguishing ontological emergence from substance dualism. Dualism posits two fundamental kinds of stuff: physical and mental. Emergentism posits one kind of stuff (physical) organized at multiple levels, with higher levels exhibiting novel properties. This maintains physicalist ontological commitments while rejecting reductive explanatory commitments.
Functionalist emergentism is scientifically tractable. Unlike mysterian positions claiming consciousness transcends scientific explanation, emergentism makes testable predictions. If consciousness is emergent integration, then Φ should correlate with behavioral markers. If consciousness involves downward causation, integrated states should influence component processes. If consciousness requires specific functional architecture, then systems lacking that architecture should not exhibit consciousness regardless of computational power.
The Consciousness AI instantiates this scientific approach. Rather than philosophical speculation alone, it constructs systems testing whether engineered functional architectures produce measurable emergent properties. The resulting data either validate or falsify the emergentist hypothesis.
Jaegwon Kim's (1999, 2005) causal exclusion argument poses the most serious challenge to emergentism. The argument proceeds as follows. If mental property M supervenes on physical property P, then M's instantiation requires P's instantiation. If M appears to cause another mental property M*, this causation must proceed through M*'s physical base P*. But given physical causal closure, P* has a sufficient physical cause P. Given causal exclusion (no event has two sufficient causes), P excludes M as a cause. Therefore M is epiphenomenal.
The Consciousness AI responds by questioning whether Φ measurements support the supervenience premise. If consciousness is strongly emergent, it may not supervene on physical states in the way Kim assumes. Φ calculations test whether integrated states provide information irreducible to component states. If Φ spikes indicate genuine integration creating novel causal powers, then the supervenience relation may be more complex than Kim's argument assumes.
Additionally, the measurement of downward causation through emotional regulation provides empirical evidence that integrated states influence components. If this downward causation is measurable and not explicable through bottom-up processes alone, it constitutes evidence against epiphenomenalism and for genuine emergent causation.
McLaughlin (1992) argues there exists "not a scintilla of evidence" for strong emergence. The Consciousness AI project addresses this objection directly by attempting to create such evidence. Rather than relying on intuitions about consciousness, the project constructs systems designed to produce measurable emergent properties.
The evidence takes three forms. First, Φ trajectories during learning episodes provide quantitative measures of integration. Second, behavioral validation through intrinsic motivation demonstrates that functional architecture produces behavior not explicitly programmed. Third, correlations between Φ spikes and behavioral markers of insight test whether integration measurements track genuine understanding.
This evidence remains preliminary. The Consciousness AI project is ongoing research, not completed science. However, the commitment to empirical validation distinguishes functionalist emergentism from purely philosophical speculation. The framework makes testable predictions. Current work tests those predictions. Results will either support or undermine the emergentist hypothesis.
Functionalist emergentism provides the theoretical foundation for The Consciousness AI project. It synthesizes ontological claims about consciousness as an emergent phenomenon with methodological insights about functional design. This synthesis enables systematic engineering of systems where consciousness might emerge from appropriate functional organization.
The framework makes testable predictions. Consciousness should correlate with Φ measurements. Emergent properties should exhibit downward causation. Specific functional architectures should be necessary for consciousness. The Consciousness AI project tests these predictions through continuous Φ monitoring, behavioral validation, and architectural experimentation.
Whether functionalist emergentism correctly describes consciousness remains an empirical question. The Consciousness AI contributes to answering this question by constructing systems designed to produce measurable emergent properties. Results from ongoing research will either validate the emergentist hypothesis or require theoretical revision. Either outcome advances understanding of consciousness and its potential instantiation in artificial systems.