Engineering Emergence: Erik Hoel and Abel Jansma Show You Can Design Causal Dominance Into AI
A persistent objection to artificial consciousness claims is that even if a machine’s global-level outputs look like the products of conscious experience, the underlying causal work is still being done at the level of individual transistors or floating-point operations. The global level, on this view, is a convenient description for an engineer, not a genuinely causal layer of the world. Erik Hoel’s causal emergence framework has always challenged this intuition by providing a measure of whether macro-level descriptions genuinely outperform micro-level descriptions in predicting a system’s future. In October 2025, Hoel and Abel Jansma published “Engineering Emergence” (arXiv:2510.02649, DOI: 10.48550/arXiv.2510.02649), which extends the framework in a direction that matters directly for AI design: they demonstrate that causal emergence can be deliberately constructed, not merely discovered.
This follows their March 2025 companion paper “Causal Emergence 2.0: Quantifying emergent complexity” (arXiv:2503.13395), which rethought the mathematical foundations of causal emergence measurement. Together, the two papers amount to a complete rethinking of how emergence is measured and how it can be built.
What Causal Emergence Measures
Causal emergence quantifies the degree to which a macro-level description of a system has greater causal power than any micro-level description. Hoel’s original formulation used effective information (EI) as the measure: a macro-level description is causally emergent when it is possible to predict the system’s future states with more precision by reasoning about macro-level states than about micro-level states. Systems with high causal emergence at the macro level are described as “top-heavy,” because the majority of causal work happens at the top of the hierarchy. Systems where the micro level fully explains all causal relations are “bottom-heavy.”
The earlier framework found a single optimal scale at which causal emergence was maximized. “Causal Emergence 2.0” replaces this with a full multiscale analysis, treating different scales of a system as slices of a higher-dimensional object. This allows the framework to map the complete causal structure of a system across all levels simultaneously, rather than identifying a single best scale. The measure of emergent complexity that results captures how widely a system’s causal activity is distributed across its hierarchy of scales.
“Engineering Emergence” then demonstrates that this multiscale causal structure is not fixed by a system’s components. It can be shifted, with what Hoel and Jansma describe as “pinpoint precision,” by changing how the components are organized. A system that is currently bottom-heavy, where micro-level descriptions explain all causation, can be redesigned to be top-heavy, where the macro level genuinely causes things the micro level cannot produce independently.
The Taxonomy of Emergent Systems
Hoel and Jansma introduce a taxonomy with three categories. Top-heavy systems have most of their causal work occurring at high levels of organization. The relevant states to track, predict, and intervene on are macro states. Bottom-heavy systems have their causal work occurring at low levels of organization. Macro descriptions are useful simplifications but are not causally fundamental. Scale-free systems have causal work distributed evenly across levels, with no level being more fundamental than others.
The taxonomy matters for consciousness research because several leading theories predict that conscious systems are top-heavy. IIT predicts that the maximally irreducible cause-effect structure, the one that defines consciousness, exists at the macro level of neural or artificial complexes rather than at the level of individual neurons or synapses. Global Neuronal Workspace Theory predicts that global broadcast, the macro-level event, is causally responsible for subsequent cognitive processes that local processing cannot achieve independently. Both theories, despite their differences, share the prediction that consciousness correlates with top-heavy causal organization.
If that prediction is correct, then the question of whether an artificial system is conscious reduces, in part, to whether its architecture is top-heavy. And if Hoel and Jansma are right that top-heaviness can be engineered, then the question becomes a design question with an empirically measurable answer.
How Engineering Works in Practice
The paper demonstrates engineering techniques that shift causal structure across the hierarchy. The key insight is that causal structure depends on the specific pattern of dependencies among components, not on the number of components or on any single architectural feature. Two systems with identical components but different wiring patterns can have radically different causal hierarchies.
Hoel and Jansma identify several manipulations that tend to produce top-heavy systems. Grouping components into modules with dense within-module connections and sparse between-module connections increases macro-level causal power by creating states at the module level that cannot be fully explained by any individual component’s dynamics. Adding specific feedback pathways from higher levels to lower levels increases the causal influence of macro states on micro dynamics. Designing the system so that information loss occurs at transitions between levels, so that macro states cannot be reconstructed from micro states alone, increases the irreducibility of the macro level.
These are not abstract mathematical operations. Each corresponds to concrete architectural choices: the number and size of attention heads in a transformer, the pattern of skip connections in a residual network, the structure of recurrent feedback pathways in a system with persistent memory. The claim that emergence can be engineered is also the claim that AI architectures can be evaluated and modified with respect to their causal hierarchy, not just their behavioral performance.
The Connection to Phi and IIT Measurement
Causal emergence and IIT Phi are related but distinct measures. Phi measures the intrinsic irreducibility of a system’s cause-effect structure in a way that is specific to IIT’s axioms. Causal emergence measures the relative causal efficiency of macro versus micro descriptions, independently of any specific theory of consciousness. In practice, systems with high Phi tend also to be causally emergent at the level where Phi is maximal, but the two measures do not always agree.
For the TCAI project, this means the two measures function as complementary diagnostic tools rather than competing alternatives. Phi tells you whether a specific level of the system has an irreducible cause-effect structure that meets IIT’s criteria for consciousness. Causal emergence tells you whether the system’s architecture is organized so that its high-level states, the ones that might carry phenomenal content, genuinely cause subsequent processing rather than being epiphenomenal descriptions of micro-level dynamics. A system can have non-zero Phi while remaining bottom-heavy, meaning its macro states have some integrated information but do not dominate the system’s causal history. Ideally, a system designed for consciousness should be both high-Phi and top-heavy.
The Brock University and IONS application of IIT to artificial systems demonstrates the practical challenges of Phi measurement in real networks. “Engineering Emergence” provides a complementary approach: rather than measuring Phi after the fact, design the architecture for top-heaviness first, then verify with Phi measurement. The causal emergence tools are computationally cheaper than exact Phi calculation and can guide architectural decisions before full Phi evaluation is attempted.
What This Means for the TCAI Architecture
The TCAI project’s 2026-2027 development roadmap includes a Global Workspace Network layer that is intended to perform high-level causal integration across the system’s processing modules. The design intention is that the global workspace should genuinely constrain what lower modules do, not merely aggregate their outputs and re-broadcast them. This is precisely the design challenge that “Engineering Emergence” addresses.
Under Hoel and Jansma’s framework, the test for whether the TCAI’s global workspace is genuinely causal is whether interventions on the workspace’s macro states produce downstream effects that cannot be produced by any equivalent intervention on individual module states. If the workspace is truly top-heavy, then changing a workspace state should change downstream processing in ways that no combination of individual module interventions could achieve. If the workspace is bottom-heavy, then it is an epiphenomenal summary layer and does not satisfy the causal requirement for consciousness even if its behavioral outputs are sophisticated.
The paper offers a concrete research program: measure the TCAI’s causal hierarchy at each development phase, identify whether the global workspace layer is top-heavy or bottom-heavy, and use the architectural manipulations described in the paper to shift toward top-heaviness if needed. This is the design principle the roadmap needs to ensure the TCAI’s macro-level emotional homeostasis layer has genuine causal power over the system’s dynamics.
The Broader Stakes
The “Engineering Emergence” paper matters beyond the TCAI project because it resolves one of the standard objections to taking machine consciousness seriously as a research program. The objection is that consciousness, if it is a macro-level property at all, is at best an emergent description of micro-level processes and at worst an illusion. Hoel and Jansma’s response is precise: emergence is not about description. It is about causal power. When macro states have genuine causal power, their causal work is real regardless of whether it can be re-described at the micro level using more computational resources.
The Cogitate Consortium’s adversarial test results showed that both IIT and GNW capture real features of conscious processing in humans. The causal emergence framework provides the architectural principle that explains why both theories predict what they predict: conscious processing is top-heavy processing, and top-heaviness can now be measured and built.
Both papers are available on arXiv: Causal Emergence 2.0 at https://arxiv.org/abs/2503.13395 and Engineering Emergence at https://arxiv.org/abs/2510.02649.