Suppose I push a needle into the paw of a stuffed bear. You, watching, may wince. But nothing feels bad to the bear itself, because it isn’t a conscious subject. Indeed, nothing can be bad for the bear. From our perspective, it might be bad if the bear were destroyed – perhaps because it wouldn’t then comfort anyone. But nothing is either good or bad for the bear itself, independent of what anyone might think about it – it isn’t what we’ll call a valuing subject. The bear is not alone: nor will a simulated person who winces when pinched feel pain. Their apparent good depends on our (external) interpretation of the simulation because what a system computes depends on how the system is interpreted (cf. Rescorla, Reference Rescorla2013).
In this commentary, we pursue Seth’s conjecture that only living things can be conscious (“biological naturalism”). Seth motivates his view with several theories (the free energy principle (FEP), predictive processing (PP), and active inference) and attention to the autopoietic and action-oriented aspects of living things. We build on this by offering a specific proposal about why life may be necessary for consciousness.
On our view, conscious subjects are always valuing subjects, and valuing subjects are generally alive (although depending on the definition of “life,” there might be exceptions). This is important because FEP, PP, and even some versions of the global workspace theory of consciousness (GWT) specify only computational/functional structure, entirely applicable to systems that seem unlikely to be either conscious or alive (e.g., video compression algorithms or email servers). By requiring that the states broadcast to the global workspace (or the errors minimized in PP) be evaluative states that play a distinctive role in making things go well with respect to the valuing subjects’ own good, we rule out non-living systems that lack their own good.
Why are living things valuing subjects? Two defining features of life are suggestive; both involve being subject to an ongoing optimization process that robustly corrects deviations away from a “goal-state.” First, living things sustain themselves metabolically through a set of interwoven homeostatic (or allostatic) processes, with regulatory mechanisms operating to maintain physiologically favorable conditions (Bich & Bechtel, Reference Bechtel and Bich2025). Second, they are subject to Darwinian processes that push populations towards higher average fitness. Moreover, whether the systems persist or reproduce themselves depends on the success of the optimization. So perhaps valuing subjects must be subject to existential optimization – optimization with skin in the game.
We don’t claim that all valuing subjects (or all living things) are conscious. Perhaps plants are not. For consciousness, we further need the existential optimization to be internalized and the evaluative states employed therein to be integrated.
Once you’re sensitive to different ways the world can be, and can evaluate whether they are good for you, you must still respond appropriately. Initially, you might only care about a few things: light or dark, or the concentration of one kind of molecule. An isolated response to some input may not be conscious. But once you’re tracking several states at once, you will need to be able to integrate many different factors – the water at the oasis smells good, but you see a predator there. Your own internal state matters: how thirsty are you anyway? To determine a coherent next move (perhaps even as a bacterium with no specialized decision-making systems), you must integrate cues from multiple sources and weigh their relative importance (Klein & Barron, Reference Klein and Barron2025).
The allostatic mechanisms of metabolic individuals are examples of just such internalized and integrated optimization processes: specifying a target range, tracking whether things are going well or poorly, and responding appropriately to that evaluation in the service of survival. Predictive error or free energy minimization, (only) when carried out by valuing subjects as part of assessing and acting for their own good, are further examples.
By contrast, a population under natural selection doesn’t typically have these features, despite also being subject to existential optimization. Evolution is ultimately responsible for the emergence of such internal optimization mechanisms, but is not itself one of them; in natural selection, the optimization process is external to the thing being optimized.
In animals like ourselves, perceptual and interoceptive states are inherently associated with a valuation through the concurrent activity of modulatory neurotransmitter systems (cholinergic and dopaminergic) with widespread and diffuse projections, responsible for controlling attention and tracking reward. Thus, information about the valuation of current inputs is in principle widely available – and in just the way that contents of a “global workspace” responsible for consciousness are supposed to be.
This brings us to a nice convergence: between a conceptual analysis of what consciousness requires, functionally speaking, and suggestive empirical facts about how evaluation systems operate in creatures like ourselves.
The “conceptual analysis” we have in mind hinges on the idea that conscious states must be available to downstream systems in order to drive “System 2” processes implicated in intentional action, memory, learning, explicit reasoning, and so on. Any states unavailable to drive either behavior (such as verbal report) or downstream cognition would be states that the subject itself would have no meaningful access to – so it would be perverse to call them “conscious.” Functionalist versions of GWT encapsulate this conceptual analysis. Perhaps surprisingly, in our way of thinking, the functional integration of any evaluative system like ours may turn out to be substrate-dependent, and so, a functionalist GWT need not imply any commitment to substrate-neutral multiple realizability (Cao, Reference Cao2022). Rather, the complex intertwinings of biochemical interactions may even be necessary to achieve the highly integrated evaluative system we think consciousness requires.
Our proposal differs from the standard GWT in its emphasis on the significance of what is broadcast. Mere information isn’t enough. What is broadcast must be evaluatively relevant for a valuing subject. An evaluation system can then serve the role that Seth highlights as crucial for consciousness: “multimodal survival-relevant integration” across diverse inputs for controlling whole-system behaviors.
Valuation is the necessary connection between life and consciousness, but the considerations developed here don’t yet show that an artificial intelligence couldn’t be conscious – only that an artificial intelligence, if conscious, must also be a valuing subject.
You have
Access
Suppose I push a needle into the paw of a stuffed bear. You, watching, may wince. But nothing feels bad to the bear itself, because it isn’t a conscious subject. Indeed, nothing can be bad for the bear. From our perspective, it might be bad if the bear were destroyed – perhaps because it wouldn’t then comfort anyone. But nothing is either good or bad for the bear itself, independent of what anyone might think about it – it isn’t what we’ll call a valuing subject. The bear is not alone: nor will a simulated person who winces when pinched feel pain. Their apparent good depends on our (external) interpretation of the simulation because what a system computes depends on how the system is interpreted (cf. Rescorla, Reference Rescorla2013).
In this commentary, we pursue Seth’s conjecture that only living things can be conscious (“biological naturalism”). Seth motivates his view with several theories (the free energy principle (FEP), predictive processing (PP), and active inference) and attention to the autopoietic and action-oriented aspects of living things. We build on this by offering a specific proposal about why life may be necessary for consciousness.
On our view, conscious subjects are always valuing subjects, and valuing subjects are generally alive (although depending on the definition of “life,” there might be exceptions). This is important because FEP, PP, and even some versions of the global workspace theory of consciousness (GWT) specify only computational/functional structure, entirely applicable to systems that seem unlikely to be either conscious or alive (e.g., video compression algorithms or email servers). By requiring that the states broadcast to the global workspace (or the errors minimized in PP) be evaluative states that play a distinctive role in making things go well with respect to the valuing subjects’ own good, we rule out non-living systems that lack their own good.
Why are living things valuing subjects? Two defining features of life are suggestive; both involve being subject to an ongoing optimization process that robustly corrects deviations away from a “goal-state.” First, living things sustain themselves metabolically through a set of interwoven homeostatic (or allostatic) processes, with regulatory mechanisms operating to maintain physiologically favorable conditions (Bich & Bechtel, Reference Bechtel and Bich2025). Second, they are subject to Darwinian processes that push populations towards higher average fitness. Moreover, whether the systems persist or reproduce themselves depends on the success of the optimization. So perhaps valuing subjects must be subject to existential optimization – optimization with skin in the game.
We don’t claim that all valuing subjects (or all living things) are conscious. Perhaps plants are not. For consciousness, we further need the existential optimization to be internalized and the evaluative states employed therein to be integrated.
Once you’re sensitive to different ways the world can be, and can evaluate whether they are good for you, you must still respond appropriately. Initially, you might only care about a few things: light or dark, or the concentration of one kind of molecule. An isolated response to some input may not be conscious. But once you’re tracking several states at once, you will need to be able to integrate many different factors – the water at the oasis smells good, but you see a predator there. Your own internal state matters: how thirsty are you anyway? To determine a coherent next move (perhaps even as a bacterium with no specialized decision-making systems), you must integrate cues from multiple sources and weigh their relative importance (Klein & Barron, Reference Klein and Barron2025).
The allostatic mechanisms of metabolic individuals are examples of just such internalized and integrated optimization processes: specifying a target range, tracking whether things are going well or poorly, and responding appropriately to that evaluation in the service of survival. Predictive error or free energy minimization, (only) when carried out by valuing subjects as part of assessing and acting for their own good, are further examples.
By contrast, a population under natural selection doesn’t typically have these features, despite also being subject to existential optimization. Evolution is ultimately responsible for the emergence of such internal optimization mechanisms, but is not itself one of them; in natural selection, the optimization process is external to the thing being optimized.
In animals like ourselves, perceptual and interoceptive states are inherently associated with a valuation through the concurrent activity of modulatory neurotransmitter systems (cholinergic and dopaminergic) with widespread and diffuse projections, responsible for controlling attention and tracking reward. Thus, information about the valuation of current inputs is in principle widely available – and in just the way that contents of a “global workspace” responsible for consciousness are supposed to be.
This brings us to a nice convergence: between a conceptual analysis of what consciousness requires, functionally speaking, and suggestive empirical facts about how evaluation systems operate in creatures like ourselves.
The “conceptual analysis” we have in mind hinges on the idea that conscious states must be available to downstream systems in order to drive “System 2” processes implicated in intentional action, memory, learning, explicit reasoning, and so on. Any states unavailable to drive either behavior (such as verbal report) or downstream cognition would be states that the subject itself would have no meaningful access to – so it would be perverse to call them “conscious.” Functionalist versions of GWT encapsulate this conceptual analysis. Perhaps surprisingly, in our way of thinking, the functional integration of any evaluative system like ours may turn out to be substrate-dependent, and so, a functionalist GWT need not imply any commitment to substrate-neutral multiple realizability (Cao, Reference Cao2022). Rather, the complex intertwinings of biochemical interactions may even be necessary to achieve the highly integrated evaluative system we think consciousness requires.
Our proposal differs from the standard GWT in its emphasis on the significance of what is broadcast. Mere information isn’t enough. What is broadcast must be evaluatively relevant for a valuing subject. An evaluation system can then serve the role that Seth highlights as crucial for consciousness: “multimodal survival-relevant integration” across diverse inputs for controlling whole-system behaviors.
Valuation is the necessary connection between life and consciousness, but the considerations developed here don’t yet show that an artificial intelligence couldn’t be conscious – only that an artificial intelligence, if conscious, must also be a valuing subject.
Acknowledgements
None.
Financial support
J.M. acknowledges funding from the Stanford Interdisciplinary Graduate Fellowship and the Center for Affective Science Graduate Fellowship.
Competing interests
The authors declare no competing interests.