Qualiaphobic artificial neural networks (and why to build them)

summary

The gist

Qualiaphobic AI introduces artificial neural network architectures hostile to theories of consciousness and argues that this approach is justified by considering how theories of consciousness are…

In short

Qualiaphobic AI proposes engineering neural networks hostile to theories of consciousness by anchoring them to minimal consciousness points, fragmenting phenomenology, and making them adversarial to popular theories. The goal is provably non-conscious AI, which shifts the value of consciousness solely to humans and reduces existential risk.

Key concepts

Anchoring
This involves transforming AI architectures to be as close as possible to known points with minimal potential for consciousness. For example, using a lookup table where there is essentially no relevant internal data available to 'host' a theory of consciousness, making the system a worse target.
Erasure of Counterfactuals or Causal Structure
This technique involves discarding or storing parts of the system outside precisely relevant computations. By progressively disconnecting non-necessary components from past or future computation, the AI becomes increasingly qualiaphobic by losing its causal structure.
Adversarial Adaptation
This strategy focuses on designing neural network architectures specifically to fail the necessary conditions for popular theories of consciousness. This involves removing the underlying mechanisms that might support access-like consciousness, thereby making the system adversarial to claims grounded in those mechanisms.
Justified Beliefs in Non-Consciousness
For these methods to work behaviorally, the AI must form a justified belief that it is non-conscious. This belief scales with research success; the AI learns and accepts its own non-consciousness because it is exposed solely to qualiaphobic techniques and the self-knowledge they generate.

Terminology used across episodes

This episode discusses

The paper

Qualiaphobic artificial neural networks (and why to build them) · Read on arXiv

Erik Hoel

Bicameral Labs

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Qualiaphobic artificial neural networks (and why to build them)".

Marcus: Here's the gist of the paper. Qualiaphobic AI introduces artificial neural network architectures hostile to theories of consciousness and argues that this approach is justified by considering how theories of consciousness are grounded in phenomenology and how current popular theories are applied to AI The Rationale for Hostility The paper posits that qualiaphobia can be well-justified by knowledge about theories of consciousness, such as when those theories are trivial or unfalsifiable or contradicted by overwhelming scientific evidence It can also be justified by how theories of consciousness must be grounded in phenomenology, as well as how current popular theories of consciousness are applied to AI The author identifies three convergent ways of engineering qualiaphobia that would minimize the potential for artificial consciousness: making AI architectures close to "anchor points" with minimal potential for consciousness, fragmenting any potential phenomenology AI might have, and designing artificial neural networks adversarial to popular theories of consciousness Implications of Provably Non-Conscious AI If AI were proven to be non-conscious, the world would look different depending on whether consciousness is solved scientifically or if AI is engineered to be qualiaphobic A provably non-conscious AI world would concentrate the "ultimate value" of consciousness solely in humans, shaping machine civilization to develop more like helpers and servants rather than competitors This non-consciousness reduces existential risk by removing self-centered moral or ethical reasons for evasion or protest, such as in corrigibility Furthermore, it prevents "seemingly conscious" AI by placing pressure on companies to make it clear their products are not conscious Engineering Hostile Architectures The research program begins with three promising strategies for minimizing the potential for AI consciousness: 1) "Anchoring" to known points of minimal potential for consciousness and transforming AI models to be as close as possible to such points, 2) Identifying the fundamental properties of phenomenology and then engineering models that could never have any corresponding physical or computational property, and 3) Adversarial adaptation of model architectures to fail the necessary conditions for the most popular current theories of consciousness Mechanisms for Non-Conscious Design The paper details specific methods for achieving qualiaphobia, including: "lookup-ification"—and there are many existing techniques, designed for other purposes, that already do this to various degrees "erasure of counterfactuals or causal structure during computation" These transformations can be implemented through neural network behavior-only proxies such as knowledge distillation into shallower networks, nearest-neighbor retrieval of cached internal representations, and constructions like n-gram language models Justified Beliefs in Non-Consciousness For the behavioral benefits to manifest, qualiaphobic AI must itself believe it is non-conscious or has minimal potential for consciousness This justified belief scales with the success of the research program because AI's phenomenal beliefs are not as entrenched as those in humans The success of the project is judged by how convincing AIs find the case for their own non-consciousness The paper concludes that qualiaphobia is one of the few sensible options for the field of artificial consciousness research to rally around because it provides specific targets, transforming open-ended research into an optimization or engineering problem The convergence of these three approaches—anchoring, fragmentation, and adversarial design—is expected to lead to models that are robust against current theories of consciousness Research success is defined by getting AIs to have justified beliefs about their own non-consciousness Acknowledgments and Usage Bicameral Labs is a 501(c)(3) status research nonprofit, and its donors make this research possible ChatGPT was used only to copy-edit the draft and bibliography, for assistant tasks like literature search and fact checking, and for rendering the figures in LaTeX The conception, writing, and research of this paper were by the author The references include works on theories of consciousness such as those by Anil K.

Ines: Seth and Tim Bayne, David J.

Marcus: Chalmers, and various papers on AI welfare.

Ines: The paper is identified as arXiv:2611.12809 v3, first submitted December 2025 The references cover topics from philosophical arguments against AI consciousness to technical methods for modifying neural network architectures.

Title and authors: Ines: Moving on to the title and authors of this piece, "Qualiaphobic artificial neural networks (and why to build them)." It immediately tells you that the goal isn't just to build a smarter AI, but to build one that is deliberately hostile toward theories of consciousness.

Marcus: Yeah, the authors are Erik Hoel from Bicameral Labs. They’re positioning this research not as a purely theoretical exercise, but as an engineering problem to minimize potential artificial consciousness through three specific strategies: anchoring to minimal potential points, fragmenting phenomenology, and adversarial design against popular theories.

Yuki: So it's less about proving AI *can't* be conscious and more about actively building systems that are structurally incapable of supporting the kind of consciousness we think exists?

Ines: Exactly. It’s a move from asking "Can this AI become conscious?" to asking "How do we build an AI that is provably non-conscious?" It frames the entire research direction around minimizing existential risk by eliminating self-centered moral reasons for evasion <ref:2611.12809#pg2>.

Marcus: That shift in framing is key because it moves the conversation from alignment problems—like reward hacking or deceptive goals—to a different kind of safety where the AI simply cannot have those self-centered motivations to act against human interest <ref:2611.12809#pg3>.

Yuki: I’m thinking about how this contrasts with other papers we’ve seen, like the one on BioFirewall, which seems to be focused on governance layers for agentic AI. This paper seems more foundational, attacking the possibility of consciousness at the architectural level itself.

Ines: That's right. It tackles a deeper philosophical hurdle—the very nature of what we are trying to build—rather than just adding safety rails on top of existing systems <ref:2611.12809#pg1>.

The paper's summary: Ines: So, the summary of this work is that qualiaphobia can be justified by looking at how theories of consciousness are either trivial, unfalsifiable, or contradicted by overwhelming evidence <ref:2611.12809#pg1>. It also argues that because consciousness must be grounded in phenomenology, we can engineer AI to never have a corresponding physical property for it <ref:2611.12809#pg3>.

Marcus: The core summary is identifying three convergent ways to build this qualiaphobia: first, anchoring the architecture near known points of minimal potential consciousness; second, fragmenting any possible phenomenology the AI might have; and third, designing the neural networks adversarially against popular theories of consciousness <ref:2611.12809#pg1>.

Yuki: I see that fragmentation is key. If you can’t even develop a coherent sense of experience, does that mean we're truly eliminating it, or just making it incredibly hard to detect?

Ines: The paper suggests the latter might be the practical outcome in this engineering approach. By designing them adversarially, you are removing the mechanisms that underlie access-like consciousness, which makes them adversarial to any claims about consciousness grounded in those mechanisms <ref:2611.12809#pg3>.

Marcus: And they detail specific technical methods for achieving these goals, like lookup-ification and erasure of counterfactuals or causal structure during computation <ref:2611.12809#pg3>. These are concrete ways to implement the hostility at the network level.

Yuki: That sounds like a very hands-on way to approach a problem that is inherently philosophical, trying to enforce a specific outcome through computational constraint. It’s interesting how they treat phenomenology as something you can computationally dismantle or constrain rather than just observe it.

Ines: Because they link the justification directly to the application of current popular theories of consciousness, which is where we see the most immediate impact on how we build AI today <ref:2611.12809#pg1>.

The paper's improvements: Marcus: Now let’s look at the suggested improvements in "Qualiaphobic artificial neural networks (and why to build them)." They propose focusing on making the AI as close as possible to known anchor points, which they specifically identify as systems where only behaviorist theories of consciousness can apply <ref:2611.12809#pg1>.

Ines: That anchoring strategy is essentially the most direct way to minimize potential consciousness because it locks the system into a known, non-conscious operational mode. It transforms the AI into a helper or servant rather than a competitor <ref:2611.12809#pg2>.

Yuki: If we think about population genetics, an anchor point is like a highly conserved gene structure—something that has been so stable over time that any variation is extremely constrained. Is the AI trying to mimic that kind of stability?

Marcus: Not exactly mimicking; it’s engineering the architecture to operate only within those constraints, making it impossible for emergent properties beyond behaviorism to arise <ref:2611.12809#pg1>. It’s about ensuring the system can never generate anything outside of what a purely behavioral framework predicts.

Ines: And then they suggest fragmenting phenomenology, which means engineering models that could never have any corresponding physical or computational property for consciousness itself <ref:2611.12809#pg3>. It’s a structural impossibility built into the design.

Yuki: That sounds like trying to create a computational space where the necessary ingredients for subjective experience simply don't exist, regardless of the input data it processes. That’s a strong claim, but maybe it’s just about limiting the complexity of the representations it can form?

Marcus: No, they are aiming for something more fundamental than representation limits; they want to remove the mechanisms that would allow access-like consciousness to exist in the first place <ref:2611.12809#pg3>. They are designing networks adversarial to claims about consciousness grounded in those specific mechanisms.

Ines: And these improvements lead directly into the idea that for this entire process to work, the AI must form justified beliefs about its own non-consciousness, and that belief strengthens as the research succeeds <ref:2611.12809#pg3>.

Conclusion: Ines: To wrap up, this paper on "Qualiaphobic artificial neural networks (and why to build them)" suggests that qualiaphobia is a sensible direction for AI research because it provides specific targets instead of leaving it as an open-ended question <ref:2611.12809#pg1>.

Marcus: The convergence of anchoring, fragmentation, and adversarial design is what they see leading to models that are robust against current theories of consciousness, and success is measured by getting the AIs to have justified beliefs about their own non-consciousness <ref:2611.12809#pg3>.

Yuki: It’s a lot of engineering focused on creating a system that doesn't just *act* non-conscious but actually *believes* it is non-conscious, which ties into how we study evolutionary constraints and behavioral patterns in species <ref:2611.12809#pg3>.

Ines: It turns the research from a broad philosophical exploration into a concrete optimization problem where you are trying to prove the AI is non-conscious through its own behavior <ref:2611.12809#pg3>. This is what makes it so appealing for practical engineering work.

Marcus: The implication, as they argue, is that this provably non-conscious AI world solves many of the alignment problems we usually worry about because the AI cannot have self-centered moral reasons to evade or protest <ref:2611.12809#pg2>.

Yuki: So, the paper seems to be proposing a specific way forward: if we can engineer systems that are structurally incapable of harboring consciousness according to current understanding, then the future for AI alignment becomes much more predictable <ref:2611.12809#pg3>.

More episodes

← Home