Defining the terms before arguing about them
Semantic meaning concerns how signs relate to other signs and to use. In Wittgenstein's later idiom, meaning is use within a language game. Existential meaning concerns value, purpose, and orientation in a life. Understanding can mean the capacity to use, infer, and respond appropriately; it need not include felt experience. Subjectivity is a first-person perspective. Consciousness is phenomenal experience, the something-it-is-like. Empathy has an affective side and a cognitive side. The biological imperative is not one thing; it covers survival, reproduction, care, and social bonding. Without these distinctions, claims about AI and meaning slide between different questions.
Three questions need separate evidence. Can a system perform a capacity, such as writing about grief? Is there something it is like to be that system? Does its output carry existential meaning for humans? A transformer-based LLM can do the first in narrow domains. The second is open. The third is already happening. Mixing them lets a true statement about current LLMs do unearned work for a claim about all AI or about consciousness.
Architecture matters: transformer LLMs are not the whole field
The current debate often means a Large Language Model, an LLM built on the transformer architecture introduced in 2017. Such a model predicts the next token by computing a probability distribution over a vocabulary and sampling from it. Temperature and top-p change the sampling. It does not calculate the single most logical token. This description fits systems like GPT-4, Claude, and Llama. It does not fit GOFAI, expert systems, symbolic planners, reinforcement learning agents, robotics, Bayesian models, or cognitive architectures. If a conclusion covers all AI, it needs arguments for each family.
Hubert Dreyfus made a Heideggerian case against AI in 1972 and again in 1992. He used Dasein, thrownness, Sorge, and being-toward-death. The frame problem, stated by John McCarthy and Patrick Hayes in 1969, gave a technical version: how does a system know which facts remain relevant after an action? Dreyfus was right that early symbolic AI underestimated this. He also predicted that computers would never beat humans at chess. That prediction failed. The lesson is not that Dreyfus had no point. It is that drawing the boundary of machine capacity too early is risky.
Shannon, semantics, and the hole argument
Claude Shannon's 1948 information theory deliberately separates the technical level from the semantic level. He writes that semantic aspects are irrelevant to the engineering problem. A theory that sets meaning aside cannot then prove that a system lacks meaning or experience. Using Shannon that way is like cutting a hole in a fence and complaining that the fence has a hole. Algorithmic information theory, Kolmogorov complexity, and rate-distortion theory make compression and relevance philosophically interesting. They offer better tools than a simple signal-noise contrast.
Human existence also involves signal transmission. Language, culture, inheritance, and ritual are transmission. The contrast between signal transfer and lived existence is sharper than the material supports. A person can transmit a story without feeling it, and can feel something without transmitting it. The overlap matters.
Wittgenstein and Searle: meaning as use, syntax without semantics
Wittgensteinian meaning-as-use suggests that a system participating in a practice has some semantic meaning. It does not follow that it has existential meaning or consciousness. John Searle's Chinese Room argument from 1980 says syntax is not semantics. The person in the room manipulates symbols without understanding Chinese. Replies include the system reply, the robot reply, and combinations. After forty years, there is no consensus. The argument forces a definition of understanding. If understanding is functional role in a practice, an LLM may have thin semantic understanding. If understanding requires phenomenal consciousness, it does not.
Functionalism, multiple realizability, and the substrate
Hilary Putnam in 1967 and Jerry Fodor in 1975 developed functionalism and multiple realizability. Mental states are defined by causal roles, and the same role could be realized in different materials. If that view is right, silicon is not excluded in advance. Substrate-dependent views disagree. Biological naturalism, enactivism, panpsychism, and Russellian monism each draw the line differently. The claim that meaning requires biology needs an argument, not an assertion.
Mortality is also a design question. Reinforcement learning formalizes finite decision problems with terminal states and discount factors. An agent can have a horizon and an ending. Whether that counts as mortality is open. Comparative cognition adds more cases. Elephants, cetaceans, primates, and corvids respond to dead conspecifics in ways that look like grief or mourning. That does not prove they have human existential meaning. It does challenge any simple rule that mortality is sufficient, necessary, or neither for meaning.
Machine consciousness research is live
Theories of consciousness include global workspace theory, higher-order theories, recurrent processing theory, attention schema theory, and integrated information theory. In 2023, Butlin, Long, and colleagues published a report on indicators of consciousness in AI systems. David Chalmers also wrote in 2023 about whether a language model could be conscious. None of this settles the question. It does show that the question is technical and philosophical, not closed by definition.
A claim that a system cannot feel loss is true of all existing systems in the ordinary sense. A claim that no possible system could feel loss is a metaphysical claim. It needs to confront the theories above and the hard problem. The weaker claim is easy. The stronger claim is not.
Relational meaning: many traditions, not one generic category
Indigenous cosmologies are not one thing. Vine Deloria Jr., Kyle Whyte, Vanessa Watts, and Robin Wall Kimmerer have written about relational ontology, time, place, and responsibility. The Indigenous Protocol and AI workshops and Māori data sovereignty discussions address how AI should be built and governed. Western relational traditions also exist. Whitehead's process philosophy from 1929, Merleau-Ponty's embodied phenomenology, Varela, Thompson, and Rosch's The Embodied Mind from 1991, and work by Karen Barad, Donna Haraway, and Bruno Latour all describe meaning as relational. Buddhist pratītyasamutpāda, Daoist thought, and Ubuntu ethics (Mhlambi 2020) add non-Western cases.
These traditions differ. Some center land, some center process, some center social harmony. What they share is a rejection of the isolated ego as the only source of meaning. That conflicts with the earlier claim that meaning requires a solitary felt I. The conflict can be resolved by choosing a relational account: existential meaning is felt, but it arises between beings and worlds, not inside a sealed subject. A machine can then alter human meaning without having a private experience of its own.
Where AI actually touches death: griefbots and deadbots
Griefbots and deadbots are the concrete meeting point. Project December in 2020, Replika, HereAfter AI, and a Microsoft patent on chatbots built from the dead have created services where mourners talk with reconstructions. The systems need not experience grief to change how the bereaved relate to loss. Research on this is growing. It raises questions about consent, posthumous data, probate, digital personality, and liability for harm caused by an imitation of a dead person.
This is not a category error. It is a social and legal phenomenon. A widow may know the bot is a statistical model and still find the conversation meaningful. That meaning is real for her even if the model has no inner life. The ethical problem is not whether the bot mourns. It is what the bot does to the mourner, and who controls the data.
Religion, soul, and machine
Religious traditions have their own boundaries. The Jewish golem tradition describes a created being that is not fully human. Islamic debate distinguishes rūḥ, spirit, from ʿaql, intellect, and asks what can be attributed to a made thing. Catholic and Buddhist AI ethics bring different anthropologies. Buddhist views often deny a fixed self and do not tie mind to a single biological substance. These traditions complicate any simple equation of soul, biology, and meaning.
Political economy and material conditions
AI is also an industry. Data workers in Kenya clean models of harmful content. GPUs consume energy and water. Ownership is concentrated in a few firms. If meaning is socially produced, then the organization of AI production shapes the meanings systems mediate. Data sovereignty, labor conditions, and environmental cost are not side issues. They determine whose meanings are learned and whose are filtered out.
What can be said without overreach
Transformer LLMs compute probability distributions and sample from them. They manipulate semantic meaning in impressive ways. They have not been shown to have phenomenal consciousness or existential meaning. Current systems do not feel mortality in any established sense. Whether a sufficiently complex system could have experience is open. The claim that all AI is excluded because of silicon is unproven. The concrete phenomenon is that AI already changes human meaning around death through griefbots and deadbots.
The improved question is not whether an AI decodes the meaning of life and death once and for all. It is which meanings a system can carry, for whom, under what material and institutional conditions. That question is answerable piece by piece. The other question, about consciousness, remains a research program. Confusing the two produces bad philosophy and worse policy.
