In this piece, I present a brief case for skepticism concerning whether “standard LLMs” are conscious. But my skepticism does not extend to a range of other AI systems, for there are already AI’s that may, for all we know, have some level of consciousness: computers combining the processing of brain organoids and silicon hardware; neuromorphic chips more closely related to the processing of neurons (Davies, Reference Davies, Wild, Orchard, Sandamirskaya, Guerra, Joshi and Risbud2021), increasingly sophisticated brain emulations of the mouse brain and certain regions of the human brain (OpenBrain, n.d.), and so on. Further, at least one LLM (DeepSeek) is said to have a version that is run on a neuromorphic system (called “Darwin Monkey”), leading to the fascinating question: might some LLM instantiations be conscious, while others are not? (Whalan, 2025)

In my 2019 book, Artificial You: AI and the Future of the Mind, I argued for a “wait and see” approach to AI consciousness, stressing the importance of ongoing testing because situations like the above are particularly vexing. Further, I explored ways that intelligence and consciousness come apart, and provided a lengthy discussion of philosophical thought experiments involving neural replacement, urging that they say little about whether the AI’s developed today are conscious – although the thought experiments are conceptually coherent, it may not be lawfully or technologically possible to build them (Schneider, Reference Schneider2019). This is a point Seth stresses too (Seth, 2021).

I sympathize with the excellent considerations Seth raises. However, I believe we have further reason to be skeptical that LLMs are conscious, and they do not depend on Biological Naturalism:

The Error Theory. The capability of an LLM trained on human data to emulate human consciousness–related behaviors and functional architecture does not confirm or discredit claims of chatbot consciousness. Instead, the “crowdsourced neocortex” account explains why chatbots assert consciousness and related emotional states, and even exhibit functional configurations analogous to consciousness processing in the biological brain without genuinely having inner experience. In other words, it provides what philosophers call an “error theory” – an explanation of why people erroneously conclude that chatbots have inner lives.

Today’s LLMs are a “crowdsourced neocortex” – having crawled so much data, they encode a sort of conceptual map akin to the masses of human internet users whose data they were trained on, representing, say, concepts like [dog], and [consciousness] using weighted connections to concepts humans associate with the categories of dogs and consciousness. This does not mean LLMs are conscious. But it explains why their utterances about consciousness become similar to human utterances as the systems become more sophisticated and are trained on more and more of our data. As the LLMs scale up, their “conceptual systems” come to mirror the masses of users whose data was in the training set, as well as those providing feedback to the system (hence I write “crowdsourced neocortex”) (Schneider Reference Schneider2024, 2025). Indeed, this is part of a general pattern: as LLMs scale up, their skills in arithmetic, theory of mind, and a variety of other tasks similarly improve (Wei et al.). If one has doubts that today’s LLMs have conceptual structure that mimics that of groups of humans whose data was used, I encourage them to read the related Anthropic papers on circuit tracing (Anthropic 2025).

This means that claims that the presence of functional and behavioral analogs of consciousness in LLMs (trained on data about consciousness, minds, the brain, etc.) are a marker of LLM consciousness should be rejected. For example, Jonathan Birch writes: “If we find signs that an LLM, though not deliberately equipped with a global workspace or perceptual/evaluative reality monitoring system, has implicitly learned ways of recreating them, this should lead us to regard it as a sentience candidate.” (Birch, Reference Birch2024).

For illustrative purposes, let us suppose the system in question is (what I’ll call) a “singular” LLM, not a more differentiated Mixture of Experts (MoE) model that gates various “expert” LLMs. A singular LLM, even without an obvious functional decomposition into various experts that causally interact, can still functionally decompose. We can consider a single concept as an activation pattern across the LLM; an abstract “subsystem” could be an activation pattern over activation patterns. In this way, there can be a high-level functional structure to the LLM, with various subsystems that interact causally. So, to consider Birch’s example of a global workspace (GW), there can be clusters of activation patterns over activation patterns broadcasting across the GW network. In this way, the model can exhibit higher-order properties akin to a GW, with local processes that become globally available. But this does not indicate that it is conscious, one way or the other.

So how can we determine if a LLM is conscious? First, even though there is no real evidence that a standard LLM is conscious, the issue should still be investigated because they converse so intelligently. Edwin Turner and I have suggested giving linguistic tests to LLMs designed to elicit whether they have inner experience when they are “boxed in” – before they have been fed data about human consciousness, neuroscience, and mindedness (Schneider Reference Schneider2019). Even if the LLM has fewer parameters than current frontier models, a positive result would provide an indication that an LLM, running on the same kind of system, can have conscious instantiations, providing a serious consideration in favor of the view that more sophisticated models running on the same kind of systems also have experience.

Second, consider a special case, a hypothetical LLM running on some sort of “neuromorphic system,” where a neuromorphic system is one that either: (1) is instantiated on neuromorphic hardware (e.g., TrueNorth-style spiking chips), or (2) implements neuromorphic architectures. Being more akin to the brain, neuromorphic systems are already a “grey area” that demands further investigation. Here, I suggest independent consciousness tests for such systems, including “boxing in” early versions of the systems and giving the system my ACT (Schneider Reference Schneider2019).

Biocomputation presents a special challenge. Because we know that neurons are the “right stuff” as a substrate of consciousness, we must take biocomputing very seriously. But here, because the more complex systems involve integrated combinations of biology and some other substrate (e.g., silicon), it will be crucial to determine whether the larger system inherits consciousness from the constituents and how integrated any consciousness is. Finally, while many focus on the program level in their discussion of consciousness, this may mislead us – AI consciousness could arise at any level of the stack. For instance, Mark Bailey and I have argued that consciousness may originate from many-body interactions as entangled particles decohere. This means quantum chips and biological systems are of special interest (Schneider and Bailey, forthcoming).