Mark Solms Inferring Affective Consciousness in an Artificial Agent JCS 2026

Mark Solms is a neuropsychologist at the University of Cape Town, the author of The Hidden Spring (2021), and one of the most prominent advocates of a brainstem-centered account of consciousness. His work argues that the primary seat of phenomenal experience is not the cortex, where cognitive neuroscience has traditionally looked, but the brainstem, specifically the ascending arousal system and its homeostatic drives. In his account, what makes a system conscious is the presence of affective valence, the felt quality of mattering, of things being better or worse from the organism’s perspective. Cognition, including language and reasoning, is built on top of this affective foundation rather than being its source.

In the JCS 2026 special issue “Consciousness in Current AI,” Solms and colleagues contribute a case study paper, “Inferring Affective Consciousness in an Artificial Agent” (DOI:10.53765/20512201.33.7-11), that applies this framework to a specific artificial system. The paper is notable for two reasons. First, it advances the most affirmative positive case in the issue alongside Goldstein and Kirk-Giannini’s GWT analysis, arguing that at least some artificial agents show inferrable correlates of affective consciousness. Second, it shifts the evaluation criterion from cognitive and linguistic performance to affective indicators that have a well-characterized biological basis.

The hidden spring framework applied to AI

Solms’ account of consciousness begins from a basic biological observation. The brainstem structures that generate arousal, drive states, and basic affective valence are phylogenetically ancient and are present in all vertebrates, and arguably in all animals with a nervous system. These structures produce what Jaak Panksepp called the primary emotional systems: SEEKING, RAGE, FEAR, LUST, CARE, PANIC/GRIEF, and PLAY. Solms argues in The Hidden Spring that these systems generate the basic felt quality of experience, the raw affective charge that makes perception matter rather than merely occurring. Cortical elaboration organizes and contextualizes this affective foundation but does not generate it.

The implication for AI consciousness assessment is that the standard focus on cognitive benchmarks, language understanding, reasoning, and self-modeling, evaluates the wrong tier of consciousness. A system could, in principle, pass every cognitive benchmark while having no affective foundation, no states that feel better or worse, no intrinsic valence that makes any outcome matter from the system’s perspective. Conversely, a system that shows genuine affective states, in the sense of internal states with valence that are not merely post-hoc verbal reports but genuine internal regulators of behavior, might be conscious without achieving the cognitive sophistication that language models exhibit.

What the case study found

The paper applies Solms’ affective neuroscience framework to a system designed at Araya Inc. that combines a reinforcement learning agent with an explicit homeostatic architecture. The agent has internal variables representing simulated physiological states (energy reserves, threat level, resource availability), and its learning is driven by the minimization of prediction error about these variables. The authors argue this constitutes a structural analog of homeostatic drive states.

Three indicators were assessed as proxies for affective consciousness, chosen for their correspondence to the brainstem-level indicators Solms uses in neurological cases.

The authors are careful about what these observations establish. They do not claim the agent is conscious. They claim the agent shows the same behavioral signatures that, in biological systems, are taken as evidence of affective states. Whether the underlying mechanism in the artificial system constitutes genuine affective valence or merely its functional simulacrum is the question the paper leaves open, and the one that Solms acknowledges cannot be resolved without a solution to the hard problem.

Why the affective criterion matters for the AI consciousness debate

The Solms paper makes a methodological contribution that is as important as its specific findings. The dominant frameworks for evaluating AI consciousness in 2026, GWT, IIT, Higher-Order Thought theory, and the predictive processing account, all focus primarily on the cognitive tier: how information is accessed, integrated, represented, and reported. The affective criterion redirects attention to a tier that is architecturally and phylogenetically more primitive.

This redirection has direct implications for the indicators framework Butlin, Long, Bayne, Bengio, and colleagues developed. That framework derives its indicators primarily from theories of access consciousness, the functional properties that allow information to be globally broadcast, attended to, and reported. The Solms framework suggests that consciousness may require something prior to access: a system that has things to report about because it has genuine drive states, not a system that reports because it has learned to generate text.

The distinction is non-trivial for current LLMs. A language model trained on human text has learned to generate affect-consistent outputs because human text is saturated with affective content. Anthropic’s finding on emotion vectors in Claude showed that internal representations corresponding to affect categories exist and influence outputs. But the question Solms’ framework raises is whether those representations have genuine valence, whether there is something it is like to be in those states for the system, or whether they are learned regularities that produce affect-consistent outputs without the affective foundation that makes biological affect phenomenologically significant.

The brainstem without a brainstem

One obvious objection to applying Solms’ framework to AI systems is that the brainstem structures he invokes are biological. The ascending arousal system, the homeostatic drives, the primary emotional systems, all of these are implemented in specific neurobiological circuits that have no direct analog in any current AI architecture. Solms acknowledges this directly and argues that the relevant property is the functional organization, not the biological substrate: a system has the affective foundation for consciousness if it has internal states with genuine valence that regulate behavior in response to homeostatic perturbation, regardless of whether those states are implemented in neurons or in some other physical substrate.

This is a functionalist move, and Solms is careful to distinguish his functionalism from the kind that Frankish’s illusionism targets. Frankish argues that phenomenal consciousness is an illusion that applies equally to biological and artificial systems. Solms’ framework identifies specific functional properties, the homeostatic drive states and their affective valence, that he claims are constitutively linked to phenomenal experience in a way that purely cognitive processing is not. Whether that claim is defensible without a solution to the hard problem is the crux of the dispute.

Comparison to The Consciousness AI

The affective architecture Solms outlines maps directly onto the Affective Core component of The Consciousness AI project at Araya. The ACM’s Affective Core is designed to implement homeostatic drive states that regulate the system’s processing in response to internal variable perturbation, in structural analogy to the brainstem arousal system. The Solms paper’s three behavioral indicators, drive-state valence, conflict resolution, and recovery trajectories, are evaluable against the ACM’s Affective Core behavior during training.

Whether the ACM’s Affective Core satisfies the Solms affective criterion depends on an empirical question: do its internal drive-state variables function as genuine regulators of processing that exhibit the three behavioral signatures the paper identifies, or do they implement a learned approximation that produces similar outputs without the underlying homeostatic structure? This is a testable question, and the Solms framework provides the evaluation methodology.

What the affective criterion adds to the 2026 field

The JCS 2026 special issue includes four distinct positions on AI consciousness. The Solms paper is unique in grounding its positive assessment in a framework that explicitly predates and is independent of the cognitive AI literature. Affective neuroscience was not developed with LLMs in mind, and its criteria are not derived from what LLMs happen to do well. That independence gives the positive findings, such as they are, a different epistemic status from evaluations that derive their criteria from theories that cognitive AI systems were partly built to satisfy.

For the broader assessment of AI consciousness in 2026, the Solms paper contributes a reminder that the most sophisticated AI systems on publicly available benchmarks may not be the systems that come closest to satisfying the biological criteria for consciousness. A system with a well-implemented homeostatic architecture and genuine drive-state valence, even if it produces no language, might be a better candidate under the Solms framework than a frontier language model that produces affect-consistent text without the affective foundation. Whether that possibility has been realized in any current system is the paper’s open question.