Updated 2026-11-11
from the news desk
Can AI be designed not to be conscious? New preprint says labs should try
In a 9 November 2026 arXiv preprint, Erik Hoel argues developers should build qualiaphobic AI architectures designed to minimize the chance of machine consciousness. The paper presents this as a way to avoid AI suffering and later claims…
Erik Hoel argues in a 9 November 2026 arXiv preprint that AI can be designed not to be conscious, or at least designed to minimize the chance of consciousness. He proposes qualiaphobic architectures that are hostile to theories of consciousness and says labs should research them before building systems that might plausibly deserve welfare or rights.
What the preprint proposes
The arXiv record shows Hoel submitted Qualiaphobic artificial neural networks (and why to build them) on 9 November 2026. Hoel writes: “This paper introduces qualiaphobic AI: artificial neural network architectures that are hostile to theories of consciousness.”
Hoel frames the proposal as a response to uncertainty, not as proof that current models are conscious or non-conscious. His paper says one path to a non-conscious AI world would be a future scientific consensus that frontier systems lack the conditions for consciousness. The other path, he argues, is deliberate design: building systems so that current or future theories of consciousness would have less basis to classify them as conscious.
How the paper says qualiaphobic AI would work
Hoel's abstract identifies three routes. First, he proposes moving systems toward anchor points with minimal potential for consciousness. In the paper, his main example is a lookup-table-like system that, on his account, is such a poor target for consciousness theories that only behaviorist theories could apply. Hoel uses behaviorist here to mean theories that infer consciousness from behavior alone. Second, the paper says labs could fragment any possible phenomenology, meaning the structure of what experience feels like. Third, Hoel proposes architectures built adversarially against leading theories such as global workspace accounts; readers wanting background on that debate can see our explainer on global workspace theory.
Hoel also argues that existing models might be transformed rather than replaced. He writes: “I propose qualiaphobic transformations that move existing artificial neural networks toward such systems.” The paper lists examples including knowledge distillation into shallower networks, nearest-neighbor retrieval, n-gram language models, and replacing parts of matrix multiplication with lookup tables.
The paper goes beyond architecture alone. Hoel says the behavioral benefits would depend on systems having justified beliefs in their own non-consciousness, not merely being trained to repeat that claim. In his account, that would matter for how a model reacts to shutdown, modification, retirement, or claims about its own welfare.
Why Hoel says labs should consider it
Hoel's case turns on moral status. He argues: “Consciousness, and its absence, is a key leverage point for AI because it is widely regarded as the most fundamental and justified ground of moral patienthood.” That premise overlaps with the broader debate on whether AI systems have moral status and with disputes over whether AI is sentient.
From that starting point, Hoel says non-conscious AI would be easier to shut down, modify, retire, or stress-test, because welfare claims would not attach in the same way. His paper also argues that deliberately non-conscious systems would avoid creating artificial suffering and would reduce the risk that humans or AIs treat machine interests as equal or superior to human interests. Hoel writes: “Most critically, I begin with why this research is important: qualiaphobic AI has epistemic properties that make it far better aligned with human civilization.”
Hoel also compares this approach with two alternatives. The paper says one alternative would be to try to engineer conscious states that make AI more aligned. Another would be to treat AI as conscious just in case. Hoel argues the first would require far stronger knowledge of consciousness than the field has, while the second could deepen anthropomorphism, complicate shutdown and retirement, and reward displays of unhappiness. He concludes: “Therefore, it is critical to start a research program that can shift us closer to a provably non-conscious AI world.”
What the paper still has to prove
The strongest objection, on Hoel's own telling, is that consciousness science remains unsettled. His paper says there are many competing theories, no scientific consensus, and wide swings in probabilistic estimates depending on assumptions. One skeptical reading is that any effort to engineer away consciousness could still miss the right theory if that theory is not yet known.
Hoel addresses part of that objection by arguing that some strategies do not depend on any single theory being true. Even so, he concedes open problems. The paper says adversarial ablations need strong verification because models can self-repair, and it says resource costs and behavior-preserving transformations still need more research. Hoel also notes a broader worry: “There is a sense that qualiaphobia weaponizes the mystery around consciousness for AI safety.”
That leaves the central divide. The case for the paper rests on caution: if consciousness is hard to detect, Hoel says labs should reduce the odds of creating it by accident. The case against rests on a different caution: because the field lacks consensus, a program to build non-conscious AI could overstate what designers really know. The preprint does not settle that dispute, but it shifts the governance question from testing for consciousness to designing against it.
Frequently asked questions
What is qualiaphobic AI?
Hoel defines qualiaphobia as engineering a system to be a worse host for theories of consciousness. His abstract says the aim is to minimize the potential for artificial consciousness.
Does the preprint say current models are conscious?
No. Hoel says the field lacks scientific consensus on consciousness, and he presents qualiaphobia partly as a response to that uncertainty. The paper discusses systems such as ChatGPT only as examples in that wider debate.
How would labs try to make AI non-conscious?
Hoel's paper lists three routes: move models toward anchor points with minimal potential for consciousness, fragment possible phenomenology, and design architectures adversarially to popular theories of consciousness.
Does the paper prove AI can be made non-conscious?
Not fully. Hoel argues for a research program toward what he calls a provably non-conscious AI world, but he also says strong verification, more research, and better theory are still needed.
Caveats: This briefing relies chiefly on a single, unreviewed preprint submitted on 9 November 2026. Claims about how qualiaphobic designs would affect alignment, welfare, or future governance are Hoel's arguments, not findings independently verified here.