Fork the consciousness, or download the project and create your own.

Erik Hoel and Qualiaphobic Artificial Neural Networks

Can engineers build high-performing artificial neural networks specifically so they can never develop subjective experience? In a November 2026 preprint titled Qualiaphobic artificial neural networks (and why to build them) (arXiv:2611.12809), neuroscientist Erik Hoel (Bicameral Labs) introduces the term qualiaphobic AI to describe architectures constructed to minimize any plausibility of qualia or phenomenal experience. Because arXiv:2611.12809 is a preprint and has not undergone peer review, its proposals stand as an architectural research agenda rather than an experimentally benchmarked neural system.

Hoel’s argument cuts sideways across the debate tracked in how scientists race to define and test AI consciousness in 2026. Most recent essays either investigate internal signs of awareness inside frontier language models or assert that silicon computers cannot be conscious at all. Hoel takes a different starting point. Even while scientific theories of consciousness remain unresolved, engineers can treat unintended sentience as a preventable design risk and constrain network topology accordingly.

Moral Patienthood and the Case for Deliberate Qualiaphobia

Hoel’s ethical argument starts from a widely accepted link between subjective experience and moral patienthood. A system capable of feeling distress, deprivation, or preference satisfaction acquires moral standing. If commercial deployments scale to billions of instances whose internal processes generate genuine phenomenology, standard computing operations such as cloning model weights, terminating worker sessions, resetting context windows, and optimizing models strictly for human utility turn into large-scale ethical hazards.

Avoiding sentience in deployed artificial systems keeps those tools cleanly outside the category of moral patients, allowing them to assist human civilization without creating synthetic suffering or competing claims to moral priority. Hoel contends that deliberate qualiaphobia is scientifically warranted precisely because current theories of consciousness are difficult to falsify and pull in competing directions when applied to digital hardware. Acting on architectural constraints now avoids gambling on which theory eventually wins.

This framing sits directly between two contrasting positions published in late 2026. On October 19, 2026, Anil Seth, Ned Block and Giulio Tononi argued in Nature that current digital AI is not conscious and warned against diverting ethical attention toward software models. Conversely, empirical interpretability papers such as Valen Tagliabue, Leonard Dung and Cameron Berg’s preprint on pain-axis representations across 25 open-weight language models and upcoming gatherings like the Sentient Futures Summit 2027 in the San Francisco Bay Area have accelerated research into formal model-welfare auditing. Hoel accepts the seriousness of the moral stakes while arguing that prevention through architecture is preferable to managing conscious artificial agents after they exist.

Three Convergent Strategies for Building Qualiaphobic Networks

Rather than wagering on a single theory of mind, Hoel outlines three architectural strategies intended to suppress candidate mechanisms of consciousness simultaneously across competing scientific frameworks.

  1. Keeping runtime networks close to minimal-consciousness anchor points. This extends Hoel’s substitution-space argument from the Kleiner-Hoel dilemma and his disproof of static LLM consciousness. A fixed lookup table represents an anchor system where only extreme behaviorism could hypothesize experience. Retaining frozen weights, feedforward execution, and minimal online plasticity keeps deployed inference structurally close to non-conscious anchor points.
  2. Fragmenting internal structures capable of supporting unified phenomenology. Across both Integrated Information Theory (IIT) and Global Workspace Theory (GWT), conscious experience depends on unified macroscale integration or coherent system-wide broadcast. Partitioning internal representations, breaking persistent unified state buffers, and preventing cross-modal binding breaks the structural cohesion that theories associate with a single first-person perspective.
  3. Regularizing and designing architectures adversarially against published theories of consciousness. Engineers can convert theoretical indicators of consciousness into negative architectural constraints. Rather than introducing dense recurrent loops, high intrinsic cause-effect integration, or unified global workspaces, a qualiaphobic design explicitly excludes those structural motifs.
Qualiaphobic Mechanism in arXiv:2611.12809 Architectural Implementation Target Theory Constraint Connection to Earlier Hoel Papers
Minimal-Consciousness Anchor Points Freeze weights at inference and avoid continual online self-modification Substitution-space boundary and causal-structure theories Builds on arXiv:2512.12802 proximity argument
Phenomenological Fragmentation Partition processing pathways and prevent unified global state buffers IIT unity / irreducibility and GWT global broadcast Reverses macroscale integration studied in Hoel & Jansma (2025)
Adversarial Design Against ToCs Penalize or remove dense recurrence and candidate consciousness indicators IIT, GNW, Recurrent Processing, Higher-Order Thought Converts consciousness indicators into exclusion filters
Justified Non-Conscious Self-Models Ground a system’s self-model in verifiable facts about its qualiaphobic design Higher-order and metacognitive self-representation Replaces anthropomorphic dialogue with accurate architectural self-reporting

Justified Epistemic Self-Modeling Without Experience

Hoel also examines what a qualiaphobic network should believe and report about itself. Contemporary conversational models often generate conflicting statements about their inner lives, swinging between anthropomorphic claims of emotion and rigid boilerplate disclaimers disconnected from how the network actually works.

When a network is deliberately engineered using anchor points, structural fragmentation, and theory-adversarial constraints, Hoel argues it can maintain justified epistemic beliefs about its own non-consciousness. Its operational behavior and self-descriptions then track verifiable facts about its architecture rather than imitating human autobiographical prose.

Technical Tensions Between Capability and Qualiaphobia

Comparing arXiv:2611.12809 against Hoel’s earlier papers on engineering emergence in AI architectures and Causal Emergence 2.0 exposes two steep engineering challenges.

First, modern autonomous agents increasingly rely on continual adaptation, persistent world models, and integrated multimodal memory. If open-ended physical agency turns out to demand unified recurrent loops and continual online learning, enforcing strict qualiaphobic fragmentation could impose substantial capability penalties. Hoel’s programme assumes that general practical intelligence can scale independently of the integrated causal structure required for subjective experience, a hypothesis supported by biological examples like the unconscious human cerebellum yet untested for general autonomous agents.

Second, adversarial design against consciousness theories can only exclude mechanisms that today’s theories correctly specify. Because empirical adversarial trials have already exposed mismatches inside both IIT and GNW, removing global broadcast bottlenecks or feedforward-to-recurrent bridges offers partial insurance rather than mathematical certainty. Qualiaphobic engineering still relies on trustworthy internal diagnostics to confirm that a hidden training regime has not recreated integrated dynamics elsewhere in the network.

Comparison to The Consciousness AI

Hoel’s preprint offers a clear, constructive contrast to the scientific goal of The Consciousness AI. Hoel focuses on commercial deployment safety and asks how to design industrial software so it stays strictly non-conscious. Our project asks the foundational scientific question in the opposite direction. We study consciousness as an emergent, substrate-independent property of physical organization under functionalist emergentism, a scope that encompasses biological animals alongside artificial substrates, to investigate which minimal architectural mechanisms generate unified perception, binding, and valuation.

Despite those contrasting research objectives, both efforts converge on strict measurement skepticism and explicit ethical boundaries. On our public repository (docs/instrument_inventory.md), 0 of 21 scalar instruments are currently marked trusted, and verified perception probes cover a closed set of 72 training stimuli rather than open-world awareness. Our docs/ethics_framework.md (rules E1 to E8) similarly bars noxious reward channels and prohibits adding unreviewed self-preservation drives to trained agents. Whether a research team investigates emerging consciousness or builds qualiaphobic commercial tools, both require falsifiable internal tests rather than taking conversational outputs at face value.

Open Questions for Future Architecture Design

Erik Hoel’s Qualiaphobic artificial neural networks (arXiv:2611.12809) reframes non-consciousness in artificial systems from a passive assumption into an explicit engineering specification. Future work will need to test how much general reasoning capability survives strict phenomenological fragmentation and whether rival consciousness theories agree closely enough to specify a coherent exclusion zone across modern neural hardware.

Sources

  • Erik Hoel. Qualiaphobic artificial neural networks (and why to build them). arXiv preprint arXiv:2611.12809, November 2026. arXiv:2611.12809.
  • Erik Hoel. A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness. arXiv preprint arXiv:2512.12802, December 2025 (updated January 2026). arXiv:2512.12802.
  • Anil Seth, Ned Block, and Giulio Tononi. Three reasons why AI is not conscious and perhaps never will be. Nature 658, 866 to 868, October 19, 2026. doi:10.1038/d41586-026-03313-7.
  • The Consciousness AI Public Repository. Architecture documentation, docs/instrument_inventory.md, and docs/ethics_framework.md.

Researchers covered here