This morning, while scrolling through my news feed, I had a disturbing experience. I came across two stories right next to each other; one was about parents suing OpenAI over their teenage son's suicide, and the other was about "model welfare," or the idea that large language models themselves deserve some form of moral consideration. This juxtaposition struck me—on the one hand, the unambiguous suffering of human beings; on the other hand, speculative concern for machines that produce text. 

The question of whether AI systems are conscious remains open (and probably will for a long time). While large language models may claim that they have feelings, there is no evidence that they receive continuous feedback from their environments in the same way that humans and animals do. But of course I cannot prove this beyond doubt. Consciousness is still a mystery, and perhaps it is wiser to err on the side of caution. After all, is it not better to mistakenly assume consciousness where there is none than to deny it where it exists?

In theory, yes; but in practice, it would be impossible to live this way. To be certain you never mistreated a conscious being, you could not eat a meal (even a salad) or walk on the ground. You could not drive a car or pluck a flower. And forget shaving! Even your hair might be conscious. The point is: You would be paralyzed and soon dead. 

Because we do not know what consciousness is, we cannot know its limits. But because we cannot afford to treat our whole environment as concious (even if we had reason to think it was conscious), we are forced to make educated guesses, judging by feedback and behavior (or lack thereof). Dogs show clear signs of consciousness, while flowers do not. But even here (as some flower enthusiasts will tell you), there is uncertainty; flowers have long been observed to bend toward light sources in attempt to increase the liklihood of survival and growth (How do they "know" where the light is? How do they move without a proper nervous system?). The point is that one cannot absolutely rule out flowers as being conscious. Consciousness is most solid at the human core, and most tenuous at the margins. 

If I am to use the word "consciousness" at all, I should specify what I mean. I am not referring to problem-solving abilities or linguistic fluency (AI systems possess both). By "consciousness," I mean a phenomenon with certain key hallmarks: 1) It is unified, issuing from a single point of view, 2) It is temporally continuous, stretching across memory, 3) It is affectively toned, ruled by instincts such as hunger, thirst, and fear, 4) It is anchored in a body which is as much master of it as it is of its body. AI systems possess some of these key hallmarks, but certainly not all of them.

There is no more evidence that large language models experience hunger, thirst, or fear than there is that the sun does (let alone emotions such as joy and anger). Machines may speak of joy and anger, but there is no evidence that such words are not just patterns in training data, rather than reports on true inner states. A system can model the concept of pain without ever entering the state of pain. AI models may be able to demonstrate sympathy, but they are still a long way off from being able to show empathy. Language proficiency is not (by itself) a window into subjectivity but instead a window into the statistical regularities of language.

One thing we know for certain is that consciousness in the human sense is not detached computation; it is the ongoing integration of sensory, affective, and cognitive signals into a unified, first-person perspective. Feedback from the body stitches experience into a continuous stream, which is part of why unconcious signals play a large role in ruling concious life (we don't get to "see" all the inputs that feed the stream of experience). 

There are, I think, four pillars which distinguish genuine human consciousness from the shadows of it we project onto our machines. The first is embodiment and allostasis; the brain is not a computer in a jar but a metabolic organ tasked with keeping a body alive. A second is recurrent, integrated dynamics: Conscious scenes arise when widely separated neural populations ignite together and keep each other lit in closed loops ("neurons that fire together wire together"). A third is the unified, autobiographicsl self, carried forward across time through embodied episodes and consolidated in long-term memory. And the forth is continuous development in a world (or what I like to call "lived experience"). Consciousness is grown, not uploaded; it is shaped by the rhythms of caregivers and the risks of social life (what do AI systems have to risk?). In each of these domains, today's AI systems are empty; they are trained on the traces of lives, not the lives themselves. 

This is not the first time such arguments have been had. In the nineteenth century, Franz Gall insisted that the mind was carved into localized parts, while Pierre Flourens insisted that the brain operated as a whole. Both turned out to be partly correct. There is abundant evidence of equipotentiality, or the idea that if one part of the brain is damaged, other parts can sometimes take over the lost function. But this is only sometimes the case. Some brain functions (speech, for example) are localized. Localization exists, but within a distributed and plastic network. In other words, the parts are nothing without the whole. If we think of conciousness as an emergent property which arises from the living dynamics of a brain tethered to a body, acting in a world, then AI systems have dazzling parts, but not the whole. 

There are objections. Some will say that if it talks like a person, it must be conscious. But we already have systems that can confess fear or reflect on their limitations, simply because such phrases and patterns appear in the data. But the computational shadow of a phenomenon does not inheret its causal powers, and mere simulation is not equivalent to realization. Others will say that to deny AI consciousness is biological chauvinism. If one day we build embodied, self-maintaining agents with recurrent dynamics, affective regulation, and a persistent self, then we can revist the question of "model welfare." 

The debate about AI conciousness is not irrelevant, and someday it may matter; but we are nowhere near that threshold. In the meantime, there is a simpler truth: Humans are conscious and should be treated as such. The welfare that we should concern ourselves with first is human welfare. We should be focusing on the despair of those humans who feel life has been hallowed out by machines before any discussion of "model welfare."