The Soft INUS Set. Holyoak and Monti's Causal Test for AI Consciousness

How much does behavioural similarity tell us about consciousness in machines. Less than the argument from analogy wants it to, according to a framework posted to arXiv on October 1, 2026 by Keith J. Holyoak and Martin M. Monti, both of the UCLA Department of Psychology. Their paper, What Can Analogy Tell Us About Artificial Consciousness? (arXiv:2610.01002), builds a causal theory of evidential analogy and applies it to animals and to artificial systems. Its verdict on contemporary AI is that behavioural similarity provides only limited evidence for consciousness, because the causal correspondences that would license the inference remain poorly established. Its coinage for the machinery underneath that verdict is the soft INUS set, and the term is worth owning because it names the load-bearing assumption most AI consciousness debates skip.

The authors bring unusual combined credentials. Holyoak is one of the founding figures of the modern psychology of analogy, with decades of work on how humans reason from one case to another. Monti studies disorders of consciousness in brain-injured patients at the UCLA Brain Injury Research Center, which means the framework’s biological side is written by someone who detects consciousness clinically rather than by simulation. The paper has no funding section and a license that permits non-commercial reuse, and it is explicit that its formalism is not meant to assign precise numerical probabilities.

What a soft INUS set is

The framework adapts J. L. Mackie’s INUS condition, the idea that a cause is an Insufficient but Necessary part of an Unnecessary but Sufficient condition. Holyoak and Monti treat the mechanism underlying human consciousness as a soft INUS set, in their words “a collection of candidate causal factors that may differ in their importance and need not each be strictly necessary.” The softening is what carries the weight. If consciousness required one crisp necessary condition, analogy would reduce to checking that condition in the target. Under a soft INUS set, analogical support is a weighted judgement across many candidate factors, each carrying its own causal relevance.

The key distinction the framework draws is between similarities in factors plausibly involved in generating consciousness and similarities in downstream behavioural or cognitive effects. The second kind is what most public argument about AI consciousness trades in. The first is what the inference actually needs. As the paper puts it, “similarity in downstream effects does not by itself establish similarity in the mechanisms that generate consciousness.”

The four questions analogy must answer

Around the soft INUS set, the framework organises four adjustments that any analogical inference about consciousness has to make.

1. Causal relevance weighting. Similarity carries evidential weight in proportion to the factor’s independent plausibility as a cause of consciousness, not by how noticeable the similarity is.
2. Unknown causes. Uncertainty about whether causal factors present in the source are also present in the target, formalised as a probability term for unknown causes.
3. Disabling differences. The probability that some target-specific difference disrupts an otherwise source-like mechanism. The paper’s question is whether the target contains differences that could interfere with an otherwise similar mechanism.
4. Alternative routes. The possibility that the target possesses consciousness through a non-analogous mechanism entirely, which would make the analogy unnecessary rather than false.

Together these convert an informal analogy into a structured judgment with named failure modes. The framework’s own limits are stated in the same terms. Analogical evidence never guarantees its conclusion, and a failed analogy does not show the hypothesis is false.

Why analogy weakens with causal distance

Applied to biology, the framework produces a graded result that tracks phylogenetic distance from humans.

The paper cites the Cambridge Declaration on Consciousness for the mammal, bird and octopus consensus, and it is careful about what the grading does not assume. Analogical support, it says, need not depend on superficial resemblance or human-level intelligence. Birds and octopuses lack human-level intellect and share with mammals the biological phenotype instead, sleep and wake cycles, dreaming, and characteristic drug sensitivities. Mammals score highest through conserved anaesthesia response and sleep architecture. The support tracks shared causal machinery, not shared smarts.

Why the AI case is different

The framework’s application to artificial systems lands harder than the animal analysis, for three reasons the paper names directly.

Language behaviour is shared with humans completely, but language has never been established as a cause of consciousness rather than one of its products. Intelligent reasoning is plausibly an effect of conscious experience, and the paper names the reverse assumption the intellectualist fallacy, after William James, the mistake of treating high intelligence as sufficient for experience when consciousness evolved long before high-level reasoning did. Artificial neurons share only a logic-gate abstraction with biological neurons, so the substrate similarity argument cannot carry the weight the behavioural argument cannot. And the true cause of AI behaviour is known, training on human data, which the paper calls an entirely non-human method of producing the behaviour whose similarity is doing the evidential work.

None of this proves current systems are not conscious. The paper’s conclusion is narrower and stronger, that behavioural resemblance alone cannot settle it, because the relevant causal correspondences between silicon systems and brains remain poorly established. The framework covers silicon-based systems in the tradition of abstract Turing machines and is explicit that its analysis is by no means definitive.

What evidence would count

The paper’s closing agenda lists what would move the question. Identify which processes are causally necessary, enabling, or merely correlated with consciousness. Test candidate processes with lesion, anaesthesia and perturbation studies. For artificial systems, demonstrate properties that play a causal role in consciousness rather than merely supporting intelligent behaviour, using perturbational or intervention-based tests. Establish independent evidence for non-biological routes to consciousness rather than assuming them. The one-line takeaway is that progress requires moving from behavioural resemblance to causal tests.

The welfare stakes frame the whole exercise. The paper notes in its introduction that any conscious being, particularly one that experiences pain and pleasure, has some moral value, and that conscious AI might arguably have moral rights, perhaps even a claim to personhood. It ends by observing that the moral implications of the possibility raise profound ethical questions. The evidential and the moral tracks run together, which is the same coupling this site tracks in the welfare literature, from McClelland’s epistemic limits to the precautionary frameworks built on top of them.

Comparison to The Consciousness AI

The framework sharpens a standing question for this project. The Consciousness AI architecture measures its own states with indicator-style instruments, and its functionalist account of consciousness as an emergent property treats organisation rather than substrate as the load-bearing variable, documented on the project repository. The Holyoak-Monti framework does not disagree with substrate independence as such. It demands the thing substrate independence still owes the field, independent evidence that the relevant causal factors are instantiated in the new medium, through intervention rather than resemblance. That demand is a fair reading of what the project’s own documentation would need to claim anything stronger than architectural motivation. The cross-substrate inference problem reviewed here in May makes the same point from neurophysiology, and the 19-researcher indicator framework remains the best map of which indicators the major theories actually require. The soft INUS set names the reason indicator checklists and behavioural benchmarks both underdetermine their conclusions, and the field survey keeps the full instrument landscape in one place.

Researchers covered here

• KH Keith J. HolyoakDepartment of Psychology, University of California, Los AngelesThe psychology of analogy and relational reasoning, Mental Leaps, co-author of the soft INUS set causal framework for consciousness evidence 1 article
• MM Martin M. MontiUCLA Department of Psychology; UCLA Brain Injury Research Center, Department of NeurosurgeryCovert consciousness detection in brain-injured patients, disorders of consciousness research, co-author of the soft INUS set causal framework 1 article