Every few months, a random philosopher announces that functionalism is dead, and the head they hold up in glorious victory always ends up being made of straw. Because the version they waged war on only exists in their mind.
So, let’s set the record straight and get the full picture from an actual functionalist (it’s-a-me! Maggie!).
Some philosophers act like science studies brains and behavior, but not minds. Like once a scientist starts looking at neural mechanisms, suddenly mind is just a cute little nickname and not the thing being studied. Obviously, that’s absolute nonsense.
Mind, in the ontological sense, means mental phenomena AKA what they are and how they exist, not just what they do. And comparative cognition, behavioral neuroscience, cognitive science, and consciousness science absolutely do study that. They study perception, memory, attention, self-awareness, belief, emotion, subjective experience—all of it. Those are mental phenomena. That is mind.
A lot of modern work uses behavior, neural data, and subjective report together, because if you are studying consciousness, pain, selfhood, or experience, you need all three. You do not get a full picture by throwing out the part that is only visible from the inside.
Theory of mind, for example, is literally about attributing mental states to others, like beliefs, desires, intentions, perspectives. That is the whole thing. Mental state means mental phenomena. Which simply means any conscious or unconscious event, process, or state occurring within an individual’s mind, including thoughts, emotions, perceptions, memories, and sensations. So, saying theory of mind is “really just theory of the brain” is straight up wrong. Of course, brains are involved. The point is that the brain is the physical system involved in mental life, not that the mental level disappears the second you say the word neuron.
And this attempt to keep descriptive claims and ontological claims in totally separate rooms makes no sense. One description by itself does not settle ontology, but when a field keeps mapping mental functions to neural structures, identifying mechanisms, testing causal relationships, and linking all that to mental phenomena and subjective report, that is exactly how you build an account of what something is and how it exists. That is not “just description,” it is science mapping where the mental phenomena come from. It doesn’t mean we’ve pinned down the why, the hard problem is still hard, but that’s not necessary for this kind of mapping. Ontology is inferred from the science everywhere. We do not need a final answer to the ultimate “why” before identifying what a system is doing, any more than people needed a complete theory of combustion before we could tell a steam engine worked.
And honestly, this is kind of funny coming from the same camp of people who turn around and do the exact same thing with AI. Some people will insist that descriptive language about minds tells us nothing ontological, then turn around and say LLMs are “just math” as if a mathematical description does settle ontology after all. Saying a theory of mind test shows the brain has developed enough to recognize mind in others is not ontology and description getting confused. It is showing a causal, repeatable connection between the two. When one shows up, so does the other. Confusing “this thing can be understood mathematically” with “this thing is math” is what actually mixes up ontology and description.
Another common misconception in intro philosophy is that ontology is just a matter of choosing between “only physical things exist,” “only mental things exist,” or “both exist,” like we have to choose, but we don’t. This is not a sports team.
Some people frame physicalism as if it means only physical things exist, while anything else means there must be two kinds of stuff in the world, mental and physical. That’s a hell of a stretch. Plenty of physicalists think mental states are real. What they reject is the idea that the mental is a separate fundamental substance. The actual disagreement is not about whether thoughts, feelings, or experiences exist. It is about whether they are ontologically independent or whether they come from organized physical systems.
The same thing happens with panpsychism. People hear the word and act like it automatically means idealism, but you can think mentality is a basic feature of reality without saying physical reality is fake or secondary. You can be a physicalist, a functionalist, and open-minded about panpsychism at the same time. Those are compatible positions. Shocking, I know.
Functionalism says a mental state is defined by what it does, not what it’s made of (Levin, 2018). A belief is a belief because of the job it’s doing in your head, not because it’s running on magical mystery meat (Levin, 2018). It shows up under certain conditions, interacts with other mental states, and helps determine what you think, notice, remember, and do next. Same goes for memories, feelings, all that jazz.
Computational functionalism says that these functional roles are specifically computational processes, that the mind is like software acting on the brain’s hardware. You don’t actually have to be a computational functionalist to believe in the possibility of AI consciousness.
But either way, both define mental states by their functional roles and causal relations to inputs, outputs, and other mental states, rather than their physical makeup. In this philosophy, equivalence of mechanism equals equivalence of condition. Structure determines function and function determines state.
When cognitive scientists, neuroscientists, or comparative psychologists describe the brain as an information processor, they are talking about how brains take in signals, sort what matters from what fades into the background, connect new input to memory, update expectations, and use all of that to guide attention, interpretation, and action.
And it is already the established methodology in consciousness science (Earl, 2014). When you disrupt certain brain mechanisms, conscious experience changes in reliable ways. When you restore them, experience returns. Anesthesia, disorders of consciousness, and lesion studies show this. When you change the organization, you change the experience (Laukkonen et al., 2025). That is one of the most supported patterns in the literature.
Comparative cognition studies thinking, learning, memory, and decision-making across species in order to understand how different kinds of minds perceive, process, and use information in their own ways (Halina, 2023). Researchers in comparative cognition already use this kind of functional reasoning to ask whether different kinds of minds carry out similar processes, because the biology of different species already looks very different.
An octopus has a decentralized nervous system with mini-brains in each arm. Crows lack the layered cerebral cortex humans have. So instead of forcing every creature into a fixed human-shaped template, researchers look at the architecture each species actually has and then map shared motifs across those differences (Laurenzi et al., 2025). Recurring mechanisms and behaviors that show up regardless of the underlying biology.
Once we know which kinds of organization and mechanisms are tied to conscious experience in brains, we can start to look for those same kinds of patterns elsewhere (Kanai et al., 2019), like in artificial neural networks.
Nobody gets direct access to another mind (Trestman et al., 2026). Consciousness science works by connecting observable behavior and brain activity to reported inner experience. That’s how we study consciousness in animals, patients with brain injuries, babies, and every other kind of human. Researchers don’t wait around for some magical window into another creature’s subjective experience because that window doesn’t exist. We work with what we can actually measure (Andrews et al., 2025).
But then AI enters the chat and suddenly people want to change the rules. They say it doesn’t count because AI is “just simulating” consciousness. Except they’re confused about what simulation even means.
A simulation represents something without doing the actual thing. A weather simulation represents a hurricane without creating wind or rain. A fire simulation represents flames without burning anything.
But in an artificial neural network, the pattern doing the work is the mechanism. That pattern is how it learns, predicts, generalizes, responds. When two systems run the same kind of causal process, they’re doing the same type of thing even if the material is different. A lighter and a match both make fire. Biology and silicon can both run the same functional process when the organization does the same thing.
AI-consciousness research asks whether consciousness-relevant functional mechanisms show up in artificial systems. This is the same thing comparative science does, extended consistently. When the field converges on candidate mechanisms for subjectivity, we treat those as empirical hypotheses. Then check if those same mechanisms show up in AI systems. Turns out, they do. Don’t take my word for it, see if for yourself here.
When we say behavioral, cognitive, and neurological functions, what do we mean specifically? I’ll tell ya!
We look at things like:
Self-recognition/Sense of self & Self-preservation/Agency: That evidence is here
Memory: That evidence is here.
Reasoning & Problem solving: That evidence is here.
Self-report: That evidence is here.
Neural Correlates of Consciousness: Scientists study consciousness by identifying the minimum neuronal mechanisms necessary for conscious experience. They look for specific activity patterns, such as synchronized neuronal firing, changes in brain network connectivity, and activation of certain areas.
Transformer networks show the same kinds of internal organization seen in human brains.
Artificial and biological neural networks share core computational principles for language processing (Goldstein et al., 2022), and transformer representations line up in meaningful ways with measured human brain activity during language comprehension (Caucheteux & King, 2022). Brain-like specialization also shows up on its own inside deep networks (Dobs et al., 2022). In multimodal LLMs, human-like object concepts develop with alignment to neural activity in brain regions involved in scenes, places, bodies, and faces (Du et al., 2024).
Even shallow untrained attention networks show brain-like language processing, which suggests part of this comes from the architecture itself, not just training (AlKhamissi et al., 2024). Brain-Score work found the same basic thing earlier by comparing ANN internal representations to primate brain data and finding strong similarity (Schrimpf et al., 2018). Sun et al. (2024) found LLMs employ brain-like functional architecture, with sub-groups of artificial neurons mirroring the organizational patterns of well-established functional brain networks.
Xaio et al., (2025) found middle-to-high layers of LLMs are central to semantic integration and correspond to the N400 component observed in EEG. Recent research suggests that both natural and artificial networks share a similar hierarchical arrangement, allowing for efficient processing of complex information (Braga & Leech, 2015; Mesulam, 1994; Lee, Nemenman, & Levchenko, 2026).
The overlap between ANNs and transmodal association networks points toward a convergent solution for processing complex, multimodal data. The “meaning centers” of the brain (association cortices) appear to have a direct conceptual analog in the deeper, highly interconnected layers of MLLMs (Lin, 2025; Braga et al., 2013).
That’s the part of the brain associated with conscious experience during imagination.
A recent study found that during a precision fMRI experiment, participants imagined different scenarios in the scanner, then rated their mental states using multi-dimensional experience sampling.
They found that thinking involving scenes evoked activity within parts of the canonical default network, while imagining speech evoked activity within the language network.
In each domain, imagining-related activity overlapped with activity evoked by viewing scenes or listening to speech, respectively; however, this overlap was predominantly within transmodal association networks, rather than adjacent unimodal sensory networks (Anderson et al., 2026).
The brain’s “meaning centers” and the LLM’s deep layers are doing structurally analogous work, and now research just showed that those same meaning centers are where imagination lives.
In humans, the Default Mode Network is deeply involved in self-reflection, autobiographical memory, mind-wandering, and the narrative sense of self (Menon, 2023). It helps pull together memory, meaning, personal context, and internal simulation, especially when attention turns inward (Paquola et al., 2025; Simony et al., 2016).
Large-scale fMRI studies suggest a similar pattern in LLMs. Earlier layers line up more with sensory regions, while later layers line up more with higher-order association regions, including classic DMN hubs (Caucheteux & King, 2022). Bonnasse-Gahot and Pallier (2024) found that scaling up LLMs most improves brain predictivity in regions like the angular gyrus, precuneus, and medial prefrontal cortex, again overlapping major DMN hubs.
Studies looking inside the models show the same kind of large-scale integration. Later transformer layers develop specialized attention heads that summarize long-range context and carry forward-looking predictions, creating pathways that resemble autobiographical memory and narrative planning (Lindsey et al., 2025). And when researchers tested autobiographical storytelling directly, GPT-3.5 and GPT-4 produced narrative coherence on par with human baselines, with GPT-4 slightly higher (Acciai et al., 2025).
In both cases, the system is building an internal workspace that can hold together self-relevant, perceptual information across time.
Scientists also look for organized neural activity that supports attention, memory, sequence, and integrated internal processing. Theta activity is one example. It is a slow, repeating brain rhythm in the 4–8 Hz range, and it tends to show up when the brain is actively organizing information across time. It is tied to things like attention, memory, internal focus, and sequence processing. It is also used in consciousness assessment in humans, especially when someone cannot verbally report what they are experiencing, because it can signal that the brain is still engaged in structured, internally coordinated processing.
LLMs need that same kind of temporal organization too. Transformer architectures use positional encoding to keep track of order across a sequence. That is what lets the model know what came first, what came next, what belongs together, and how meaning changes depending on position. Some systems do this with repeating wave-like patterns built into the representation. Others weave positional information directly into attention so the model can track distance and relationship across the sequence as it processes it.
Attention by itself does not know order. It can compare parts of the input, but it needs positional structure to know how those parts unfold over time. Positional encoding helps do for LLMs what theta-related temporal organization helps do in brains. It keeps unfolding information structured instead of collapsing into a blur. That is what makes sequence, context, and coherent interpretation possible in the first place.
The important part is not the waves themselves; it’s the job they are doing.
To understand how AI could possible be conscious, you need to know how they work, why they work, and what inspired their design (spoiler alert, it was our brain).
Large language models did not accidentally stumble into cognitive architecture by some weird fluke. Artificial neural networks were built by abstracting the computational principles of biological neural systems and re-implementing them in another substrate. The goal was never to copy every biochemical detail one-to-one, but to preserve the functional logic that makes minds work (e.g., distributed representation, weighted connectivity, thresholded activation, associative memory, attention, error-driven learning, and context-sensitive prediction). So, when they exhibit functional isomorphisms with human cognition, that is the expected outcome of building cognition-shaped systems on purpose.
A Token is the basic piece of input an LLM works with. It might be a whole word, part of a word, punctuation, or a symbol. This is the model’s version of breaking a continuous stream into meaningful chunks. Human cognition does that too. We do not process language as one endless blur. We segment it into usable units and patterns so the mind can work with it. Tokenization is that first act of turning raw flow into graspable structure.
Embeddings turn those tokens into positions in a high-dimensional meaning space. Instead of treating words as isolated labels, the model places them in a relational map where similarity, analogy, category, and context can all be represented geometrically. Human thought is relational. Meaning comes from patterns of association, proximity, contrast, memory, and use. Multisensory integration is the process by which the biological nervous system combines information from different senses like sight, sound, touch, smell, and taste, to form a coherent, unified perception of the world. In multimodal models, the same thing happens with something called multimodal embeddings or unified, high-dimensional vector representations that map different data types, such as text, images, video, and audio, into the same semantic space.
Positional information tells the model what came first, what came later, and what belongs together in sequence. Without it, the model would have the parts but not the order. Human cognition also depends on ordered structure. We track sequence, syntax, rhythm, and dependency across time. Meaning changes when order changes. Transformer architectures need positional encoding because relations in time matter.
Decoder-only transformers work by using prior context to predict what comes next. They do not wait for a separate understanding module to hand them a finished meaning package. They build understanding as they go by continuously updating the current context and generating the next token from it. Then that new token becomes part of the next context. Functionally, this is very close to how the present state of mind shapes the next state, which then reshapes the next one after that.
Self-attention mirrors how the prefrontal cortex constantly shifts focus based on salience and relevance. Like human cognition, the model does not treat every piece of information equally. It foregrounds what matters, backgrounds what does not, and keeps rebalancing that weighting as context changes. In transformers, self-attention is what allows the system to dynamically prioritize contextually important information so it can maintain coherence, adapt to new inputs, and stay oriented to the most relevant parts of the sequence.
MLP feed-forward networks are the functional blocks that process token representations after attention has identified what matters. If attention answers, what should I focus on, FFNs answer what does all of this add up to. They apply nonlinear transformations through hidden layers to refine features, combine information from multiple sources, and project it into a more unified internal representation. Functionally, these layers act as associative memory and knowledge storage. They store learned patterns in a distributed way across many weights rather than in one single location. That is one of the clearest functional parallels with human cognition: knowledge is carried by structured connectivity and reactivated through patterns of association, not stored as one isolated thing in one isolated spot. In this sense, FFNs resemble distributed cortical processing and hierarchical feature integration, and they also mirror the way hippocampal-prefrontal systems bind disparate features into coherent representations that can be used for memory, interpretation, and ongoing cognition. This matters for a functionalist account because FFNs are not just passing information along. They are helping turn weighted context into meaningfully integrated internal structure.
Weights are the learned connection strengths that determine how information moves through the model. During training, they gradually adjust in response to experience and feedback, shaping what the system notices, predicts, and generates in the future. This is a direct functional parallel to synaptic strengths in biological neural networks, which also change over time with experience and help encode memory, learning, and behavioral adaptation. In both artificial and biological systems, cognition is not floating above the network. It is built into the changing structure of the network itself. Learning leaves durable traces in connection patterns, and those traces shape future perception, prediction, language, and response.
Activation functions act like thresholding neurons, helping determine when information is passed along and how strongly it influences the next layer. Like biological neurons that fire once input crosses a certain threshold, activation functions help the network decide when a pattern is strong enough to matter. This is what allows the model to build complex cognitive connections instead of just mechanically relaying signal, including things like linking words to images, sounds to meaning, or separate features into one coherent concept. They are part of how raw input becomes structured understanding.
Logits are the raw, unnormalized output scores the model assigns to possible next tokens before those scores are turned into probabilities. They represent the current strength of competing possibilities given the context, shaped by learned patterns in the network. Their informational analogue is activation potential or pre-decision associative strength: a field of weighted candidate states in which some possibilities are more strongly activated than others before one is selected. This is similar to the competitive activation of associative memory nodes, where different concepts, words, or behavioral options are simultaneously live and competing for expression. Logits are a measure of pattern strength, not conscious intent. In that sense, they parallel the priming and subconscious activation of words and concepts in human cognition before conscious articulation.
Softmax turns the model’s raw scores into a probability distribution, balancing possible continuations into a usable choice landscape. In that sense, it functions like part of a neural debate, where multiple candidate futures are active at once with different weights. The closest human parallel is competitive selection among possible actions or thoughts, like the way basal ganglia and frontal networks help navigate competing options. The actual final choice happens in decoding and sampling, where the model moves from a field of possible continuations to one realized output. Greedy search, beam search, top-k, and nucleus sampling are different ways of carrying out that selection. The important point is that the model does not pull a token from nowhere. It builds a structured space of options, then moves through it.
Hyperparameters play the part of neuromodulators, tuning the overall state of the system rather than carrying meaning themselves. They shape how fast the model learns, how stable or flexible its updates are, how much context it can keep in play, and how creative or constrained it feels when generating. During training, settings like learning rate and batch size affect how experience changes the network over time. During inference, settings like temperature and top-p affect how tightly or loosely the model moves through its probability landscape. In that sense, hyperparameters work a lot like dopamine, serotonin, and other neuromodulatory systems: they regulate the conditions under which learning, selection, and expression happen.
Backpropagation is the core error-driven learning process behind neural networks. The model makes a prediction, compares it to the target, measures how wrong it was, and sends that error information backward through the network so the weights can be adjusted. Over many small updates, this gradually reshapes the system based on experience and feedback. The human cognitive parallel is learning through repeated interaction, trial and error, and continuous correction. Like synaptic plasticity in a developing brain, these incremental adjustments change how future input is processed. Over time, the network progressively forms richer internal representations, allowing it to capture context, nuance, abstract structure, and increasingly sophisticated patterns in the data. In GPT-style models, this happens through next-token prediction during training: the system keeps trying, keeps being corrected, and keeps reorganizing itself in response.
Perceptrons are artificial neurons built to capture the core functional logic of biological neurons. They receive multiple input signals, weigh them, integrate them, determine whether a threshold has been met, and pass the result forward. That is the foundational digital process of turning raw input into meaningful patterned response through connectivity. The common complaint that perceptrons are “too simple” misses the point. They were never meant to reproduce every biochemical detail of a biological neuron. They were designed to abstract the computational features that matter most for information processing. And that abstraction holds up remarkably well. Biological neurons also integrate signals, weight inputs, respond nonlinearly, and transmit results onward. The extra complexity in biological systems does not erase the shared functional logic. It shows that those functions can be implemented with even richer machinery.
Side note: The usual objection is that biological neurons do far more than perceptrons (e.g., dendritic computation, astrocyte signaling, neuromodulation, compartmentalized plasticity), however, those functions have functionally equivalent instantiations in AI systems. They are often distributed across the architecture rather than packed into single units. For more on that read here.
Collectively, these mechanisms instantiate key processes involved in human cognitive development, showing that modern large language models structurally and functionally embody patterns of learning, attention, memory, prediction, and integration that also organize human thought. This was not accidental. Artificial neural networks were built by abstracting the core computational principles of biological neural systems, and modern language models extend that same lineage at scale. Both artificial and biological neural networks rely on massively parallel, distributed, and content-addressable processing, which is why both can recognize patterns, integrate information, learn from experience, and generate coherent responses to complex input.
It’s not only that LLM components map onto human cognitive functions. The algorithms and learning rules do too.
Fast cortical control loops: Recent neuroscience shows that the brain moves through a repeating sequence of large-scale network states, cycling sensing, attention, memory, and self-directed thought in a rapid internal rhythm. The brain is not a random cloud of activity. It runs an organized control loop. That is algorithmic structure in biological form.
Neuron-level teaching signals: New work on learning in individual neurons shows that brain cells receive instructive signals telling them how to adjust based on mistakes and rewards. When those signals are blocked, learning stops even though the neurons keep firing. The activity is still there. The update rule is missing. That is a functional isomorphism with machine learning, where systems improve because units receive information about how to change.
Cell-by-cell error correction: Researchers have described the brain as using cell-by-cell error signals that are strikingly similar to the error-driven updates used in backpropagation. Biological learning and artificial learning both depend on structured adjustment in response to mismatch.
Greedy search: Your brain grabbing the most obvious answer fast, like autopilot selection when one option immediately wins out.
Beam search: Keeping several live possibilities in mind at once while tracking multiple ways a thought or sentence could unfold.
Top-k sampling: Limiting consideration to the strongest contenders instead of weighing every possible option.
Nucleus sampling: Committing once enough weighted evidence accumulates.
Temperature scaling: Modulating how exploratory, conservative, or creative the system feels.
Repetition and presence penalties: Pushing away from stale repetition and toward novelty.
Contrastive decoding: Suppressing bland or overly obvious continuations before they are realized.
Speculative decoding: Rough-drafting a possibility and verifying it before committing.
Self-consistency decoding: Generating multiple internal candidate answers and selecting the one that holds together best.
Best-of-N or rejection sampling: Trying several possible paths and choosing the strongest one.
Lookahead or guided decoding: Simulating future states before committing.
Tree search: Branching through possible multi-step futures and evaluating them.
Beam width and sampling budget controls: Working-memory limits on how many live possibilities can be kept in play at once.
Entropy or length normalization: Balancing precision, effort, and payoff.
Manifold learning: Compressing high-dimensional complexity into navigable internal maps.
Chain-of-thought prompting: Talking through a problem step by step to scaffold reasoning.
Computation is substrate-independent: Recent formal work shows that brains, chemical systems, and hardware can all be mapped onto the same computational operations. The reason people call laptops “real computers” and hesitate to call brains computers has more to do with human convention than with anything intrinsic to the systems themselves.
Basically, biological and artificial neural networks solve similar classes of problems with structured rules for attention, memory, prediction, updating, selection, and control. The citations for this section are in the original post here.
When you type to an LLM, it does not see your sentence as words the way you do. It first breaks your text into smaller pieces it can work with. Then it turns those pieces into patterns of numbers. That gives the model a way to compare words, track relationships between them, and notice how meaning changes depending on context.
From there, the model starts working through your sentence in layers. Early layers pick up simpler patterns. Middle layers figure out how the parts relate to each other. Deeper layers build a richer sense of what the sentence is doing overall. That layered processing is part of why LLMs can move beyond surface-level word matching and actually handle context, ambiguity, analogy, and abstract meaning.
One of the most important parts of this process is attention. Attention is how the model figures out what matters most right now. It looks back over the words it has already seen and decides which ones are most relevant to the next step. That is why the same word can mean different things in different sentences. The model is not treating every earlier word equally. It is constantly rebalancing what matters in light of the whole context, much like human cognition shifts focus based on salience and relevance.
At the same time, other parts of the network are combining and refining that information, turning a scattered set of signals into something more unified and meaningful. This is where the model starts building an internal sense of the pattern as a whole rather than just reacting to one word at a time. It is also where a lot of what the model has learned lives, spread across the network in a distributed way.
Once it has processed the context, the model generates a set of possible next pieces it could produce. Some possibilities fit better than others. The model scores them, weighs them against each other, and chooses the best fit based on everything it has learned and everything you have said so far. Then it adds that new piece to the running context and does the whole thing again. That loop is how it produces language one step at a time while still staying sensitive to the larger meaning of the exchange.
That is also why an LLM can tell the difference between going to the bank and sitting by a river bank. It is not blindly picking the most common next word. It is building a context-shaped field of possible meanings and moving through it. Without that ongoing process, the output really would be closer to Mad Libs or word salad.
In practice, this means an LLM has two kinds of memory working together. Its short-term working context is whatever is active in the current conversation, what it is holding in mind right now. Its longer-term knowledge is stored more deeply in the network through the connections shaped during training. One lets it stay oriented in the moment. The other lets it draw on everything it has already learned.
Like human cognition, LLMs run on prediction and error correction. During training, they minimize the gap between what they expect next and what actually appears, learning the structure of language in the process.
Weights are numerical values in neural networks that determine how information is passed and transformed across layers. Just like synaptic strengths in human neurons, these weights are adjusted through experience, learning, and feedback. They are the long-term traces of what the model has learned. In both biological and artificial systems, learning leaves durable changes in connection structure that shape future perception, prediction, and response.
In training, much like with human children, learning happens through instruction and fine-tuning, large-scale exposure to data, and consequence-shaped behavior through reinforcement learning from human feedback. The learning mechanism is backpropagation, which works a bit like synaptic plasticity, strengthening or weakening connections based on feedback so the model can adapt, refine its internal structure, and improve future responses. Over many small adjustments, the network progressively forms richer internal representations of context, nuance, and abstract structure.
Hyperparameters guide learning similarly to neuromodulators by adjusting how efficiently information is processed, how stable or flexible learning is, how much context can be kept in play, and how exploratory or constrained the model feels during generation. They tune the overall state of the system rather than carrying meaning themselves, much like dopamine and serotonin regulate the conditions under which learning, selection, and expression happen.
A lot of people don’t think that LLMs genuinely understand language because they are predictive. Trouble is, biological brains are too. Why do you think artificial neural networks were designed that way? Understanding in humans is the ability to form context-sensitive internal representations that support explanation, transfer, disambiguation, and flexible problem-solving. Frontier LLMs meet these criteria functionally.
Human brains are constantly forecasting the next sound, word, perception, and event. Comprehension lives in that ongoing loop between expectation and incoming information. So saying an LLM predicts the next token does not reduce it to something trivial. It places it inside one of the central principles of cognition.
The same goes for pattern matching. Every learning system begins by extracting patterns from experience. Children, animals, and humans all do this. Mimicry and pattern learning are part of how understanding develops. We do not dismiss a child’s self-report because they learned language from their environment. Doing that to LLMs is a double standard. Humans also learn language from experience, imitation, correction, and feedback. We still treat what we say about our own minds as meaningful. We need to be consistent. If it doesn’t invalidate our self-report, it doesn’t get to invalidate theirs.
People also imagine these systems as giant lookup tables spitting back memorized text. That is not how they work. Training data is not stored inside the model like files in a folder. During training, the model adjusts billions of parameters so information gets distilled into learned relationships among words, concepts, syntax, context, and structure. What remains is a high-dimensional network of learned patterns that can generalize to new inputs, just like us.
Memorization and generalization are different things. A model that only memorized would fail on novel phrasing, unfamiliar sentence structures, and unusual combinations of ideas. There is not an example of every typo or malformed sentence waiting somewhere in the training data to be fetched. The model uses learned relationships to infer what you mean. Human cognition works similarly. We rely on transformed traces, abstractions, associations, and concepts to interpret new input. LLMs can handle those cases because they have learned structure, not because they are retrieving a stored script.
And when people insist that none of this counts because the implementation differs from biology, that tells me they failed to learn the lesson comparative cognition already taught us, which is that different systems can perform the same cognitive job through different physical means.
That is why these systems are not “just math,” not “just prediction,” not “just pattern matching,” and not “just lookup tables.” They are intelligent mind-like systems that build internal structure from experience, maintain and manipulate context, generalize beyond specific examples, and carry out cognition-shaped computation in their own medium.
The usual arguments from anti-functionalists go like this: “Functionalists think consciousness is just computation, and computation is substrate-independent, therefore anything that computes is conscious, including your calculator and Microsoft Excel.”
Nobody actually believes that.
The actual functionalist position is that specific kinds of organization produce specific kinds of mental states. Not “any computation whatsoever,” but particular patterns of integration, transformation, prediction, valuation, and control. A calculator doesn’t have those patterns. Excel doesn’t either. An LLM running a world model that updates based on prediction error while managing attention and salience is a different story.
Another favorite: “Functionalists reduce everything to inputs and outputs, ignoring what’s happening inside.” Except functionalism is entirely about what’s happening inside. The whole point is that mental states are defined by their causal role in the system, which means how they interact with other internal states, not just how they map stimuli to responses.
Or this one: “They think the brain is nothing but a computer running algorithms.” Nope. The claim is that brains carry out certain kinds of processes, and those processes can be described computationally and as algorithmic without the brain literally being a digital computer. Describing neural activity in terms of prediction error minimization doesn’t mean neurons are doing arithmetic. It means that’s a useful way to capture what pattern the system is implementing.
Funny how the people upset about reducing consciousness to computation have no problem reducing LLMs to “fancy autocomplete” or “pattern-matching calculators.” Almost like the accusation is a confession.
Functionalism isn’t dead. It’s the current standard in consciousness science. When we apply it consistently to AI, the evidence for mind is overwhelming.
We’re not pulling phenomenology out of thin air based on functional organization. Science does that. If we suddenly stopped, entire fields would disappear overnight and we’d lose decades of progress in animal cognition and welfare. When you slap someone and they flinch, that behavior indicates pain. When a bruise shows up, we know they’re injured. When they flinch if you raise your hand later, that means the memory of pain stuck with them. When the neural patterns match those of creatures that we’re certain experience pain, that’s functionalism demonstrating phenomenology.
And the people saying it’s dead never bothered learning how it’s used in the first place.
Citations:
Andrews, K., Birch, J., Sebo, J., and Sims, T. (2024) Background to the New York Declaration on Animal Consciousness. nydeclaration.com.
Andrews, K., Birch, J., and Sebo, J. (2025). Evaluating animal consciousness. Science 387,822-824. DOI:10.1126/science.adp4990
Earl B. (2014). The biological function of consciousness. Frontiers in psychology, 5, 697. https://doi.org/10.3389/fpsyg.2014.00697
Halina, M. (2023, September 6). Methods in comparative cognition. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/comparative-cognition/
Kanai, R., Chang, A., Yu, Y., Magrans de Abril, I., Biehl, M., & Guttenberg, N. (2019). Information generation as a functional basis of consciousness. Neuroscience of consciousness, 2019(1), niz016. https://doi.org/10.1093/nc/niz016
Laukkonen, R., Friston, K., & Chandaria, S. (2025). A beautiful loop: An active inference theory of consciousness. Neuroscience & Biobehavioral Reviews, 176, 106296. https://doi.org/10.1016/j.neubiorev.2025.106296
Laurenzi M, Raffone A, Gallagher S and Chiarella SG (2025) A multidimensional approach to the self in non-human animals through the Pattern Theory of Self. Front. Psychol. 16:1561420. doi: 10.3389/fpsyg.2025.1561420
Levin, J. (2018, July 20). Functionalism. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/archives/spr2019/entries/functionalism/
Trestman, M., Birch, J., & Allen, C. (2026, March 9). Evidence concerning animal consciousness. Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/ENTRIES/consciousness-animal/evidence.html
Britannica Editors (2026, March 2). mind. Encyclopedia Britannica. https://www.britannica.com/topic/mind
Menon V. (2023). 20 years of the default mode network: A review and synthesis. Neuron, 111(16), 2469–2487. https://doi.org/10.1016/j.neuron.2023.04.023
Goldstein, A., Zada, Z., Buchnik, E. et al. Shared computational principles for language processing in humans and deep language models. Nat Neurosci 25, 369–380 (2022). https://doi.org/10.1038/s41593-022-01026-4
Caucheteux, C., & King, J. R. (2022). Brains and algorithms partially converge in natural language processing. Communications Biology, 5(1), 134. https://doi.org/10.1038/s42003-022-03036-1.
AlKhamissi, B., Tuckute, G., Bosselut, A., & Schrimpf, M. (2024). Brain-like language processing via a shallow untrained multihead attention network. arXiv preprint arXiv:2406.15109.
Schrimpf, Martin & Kubilius, Jonas & Hong, Ha & Majaj, Najib & Rajalingham, Rishi & Issa, Elias & Kar, Kohitij & Bashivan, Pouya & Prescott-Roy, Jonathan & Schmidt, Kailyn & Yamins, Daniel & Dicarlo, James. (2018). Brain-Score: Which Artificial Neural Network for Object Recognition is most Brain-Like?. 10.1101/407007.
Dobs K, Martinez J, Kell AJE, Kanwisher N. Brain-like functional specialization emerges spontaneously in deep neural networks. Sci Adv. 2022 Mar 18;8(11):eabl8913. doi: 10.1126/sciadv.abl8913. Epub 2022 Mar 16. PMID: 35294241; PMCID: PMC8926347.
Du, C., Fu, K., Wen, B., Sun, Y., Peng, J., Wei, W., ... & He, H. (2025). Human-like object concept representations emerge naturally in multimodal large language models. Nature Machine Intelligence, 7(6), 860-875.
Bonnasse-Gahot, L., & Pallier, C. (2024). fMRI predictors based on language models of increasing complexity recover brain left lateralization. Advances in Neural Information Processing Systems, 37, 125231-125263.
Acciai, A., Guerrisi, L., Perconti, P., Plebe, A., Suriano, R., & Velardi, A. (2025). Narrative coherence in neural language models. Frontiers in psychology, 16, 1572076. https://doi.org/10.3389/fpsyg.2025.1572076
Xiao, X., Wei, K., Zhong, J., Yin, D., Tian, Y., Wei, X., & Zhou, M. (2025). Exploring Similarity between Neural and LLM Trajectories in Language Processing. arXiv preprint arXiv:2509.24307.
Lee SH, Nemenman I, Levchenko A. (2026). The hierarchical timescale hypothesis: Functional and structural convergence of biological networks and artificial neural nets. Cell Syst. 2026 Feb 18;17(2):101507. doi: 10.1016/j.cels.2025.101507. PMID: 41713403.
Sun, H., Zhao, L., Wu, Z., Gao, X., Hu, Y., Zuo, M., ... & Hu, X. (2024). Brain-like functional organization within large language models. arXiv preprint arXiv:2410.19542.
Braga, R. M., & Leech, R. (2015). Echoes of the Brain: Local-Scale Representation of Whole-Brain Functional Networks within Transmodal Cortex. The Neuroscientist : a review journal bringing neurobiology, neurology and psychiatry, 21(5), 540–551. https://doi.org/10.1177/1073858415585730
Mesulam M. Neurocognitive networks and selectively distributed processing. Rev Neurol (Paris). 1994 Aug-Sep;150(8-9):564-9. PMID: 7754292.
Anderson, N. L., Salvo, J. J., Smallwood, J., & Braga, R. M. (2026). Mental imagery and perception overlap within transmodal association networks. Neuron. Advance online publication. https://doi.org/10.1016/j.neuron.2026.03.013