Large Language Models: Our Best Working Model of the Human Mind?

DATA STORIES | LARGE LANGUAGE MODELS | ARTIFICIAL INTELLIGENCE

Large Language Models: Our Best Working Model of the Human Mind?

To read this article for free click here.

Over the past few years Large Language Models (LLMs) have leapt into the public consciousness, captivating us with fluent essays, code snippets, and even poetry that can appear uncannily human. This sudden surge in AI capability begs the question: if a machine can compose prose or respond thoughtfully to our queries, to what extent is it genuinely thinking? In exploring how these models learn we are uncovering striking parallels with our own cognitive processes. Parallels that challenge long-held assumptions about what intelligence is and maybe what it means to be human.

So what are LLMs? And what will they teach us about us?

Geoff Hinton: The Biological Blueprint for AI

To understand LLMs it helps to return to a time when the field of artificial intelligence was headed in a very different direction. In the 1970s and 1980s, most AI researchers believed that intelligence would emerge from deterministic rules, that is defined logic, symbolic representations and formal reasoning systems. Machines it was thought, could be made to think by encoding explicit knowledge about the world in ‘if–then’ statements and structured instructions. This approach treated intelligence like a kind of giant filing cabinet, neatly organised and manually maintained by human experts.

But Geoffrey Hinton had a different idea. A British psychologist turned computer scientist, he was less drawn to logic and rules than to biology and learning. Rather than programming machines with instructions, he wanted to understand the mind by replicating how the brain actually learns. Unlike formal systems of logic, the brain does not follow predefined rules. It learns by adjusting the strength of connections between neurons, reinforcing patterns that work and discarding those that do not. Hinton believed that machines could learn in much the same way. If the brain could self-organise into a thinking system, then perhaps a computer, built from simple neuron-like units, could do so too. The AI of the time he argued had missed something fundamental: that human intelligence is not primarily built on logic, but instead on patterns of association.

In short, he wanted to understand the mind by modelling it.

Inspired, Hinton focused on neural networks. Neural Networks are systems of interconnected artificial neurons (think artificial brain cells) that learn by adjusting the strength of their connections. Critics at the time were deeply sceptical of the idea that a vast network of simple computational units without any built-in understanding of the world could learn to perform complex tasks purely through exposure to data. Hinton’s neural network approach seemed too vague, too chaotic, and too much like blind trial and error.

But they were wrong. And Hinton knew it.

The Neural Network

Neural networks are nothing new. Initially hypothesised in the 1940s, Neural Networks are essentially just a collection of artificial neurons; arranged in layers that are basically little more than connected switches that either activate or don’t depending on the input they receive. A new model ready for training starts off knowing absolutely nothing at all: the connections between neurons are assigned randomly. Its nothing more than a bunch of switches waiting to be set, and there is no concept of language, logic or facts. It’s a black box waiting to be shaped.

During training, the model is given examples with known correct answers and measures how far off its own predictions are. The model continually tests its guesses against real-world data, refining its internal wiring until it can recognise patterns and produce coherent responses with increasing accuracy. If the network under training misclassifies an image, for instance, it doesn’t just shrug and try again. Instead, it uses an approach called backpropagation, introduced by Hinton and others in the 1980s, to work out how much each connection contributed to the error. This information is then used to tweak the weights, effectively flowing the mistake backwards through the network, updating each neurons contribution accordingly. This ability to learn, not through explicit programming, but by adjusting millions of these tiny switches based on experience was deeply counterintuitive at the time. Reflecting on those early days Hinton later remarked: “People in mainstream AI thought that was completely ridiculous. It sounds a little ridiculous… but it works.”

What then seemed implausible is today the foundation of modern artificial intelligence. By the 2010s, his hunch was proven right: as computers got faster and data grew, neural nets suddenly began to excel at transcribing speech, translating language, and more​. The 2012 ImageNet competition marked a turning point when Krizhevsky, Sutskever, and Hinton demonstrated that deep neural networks could significantly outperform traditional computer vision approaches, triggering what would become known as the ‘deep learning revolution.’

One of Hintons analogies is that neurons ‘vote’ on what they think the correct answer to a prompt is. In a neural network each artificial neuron processes incoming signals and passes its output forward. Rather than delivering a simple yes or no, each neuron contributes a weighted preference for a particular outcome; something akin to casting a vote with a certain strength. In the final layer of many neural networks, this process is formalised using what is known by computer scientists as the softmax function. This mathematical operation takes all the raw outputs from the final layer, each one representing a degree of belief in a possible answer, such as is this a picture of a ‘cat’ or ‘dog’, and converts them into a probability distribution. In other words, the strongest output becomes the most likely prediction, but all have a probability attached to them.

Transformers: The magic behind LLMs

The impressive capabilities of LLMs that we know and love today like GPT-4 or Claude are derived from the Neural Networks first hypothesised nearly a century ago. These LLMs typically use a neural architecture called a Transformer (Vaswani et al, Google) which can be thought of as a multi-layer network of artificial neurons with a special communication system called ‘attention’. The Transformer is composed of dozens of layers and each layer has multiple attention heads, sub-units that direct the LLMs focus to different pieces of the input text. When you prompt an LLM with a sentence, the early layers break down the input into contextual signals (numerical vectors representing words and their relations). As information passes through layers, the LLM hones in on relevant patterns. Attention mechanisms allow the LLM to ‘decide’ which words or phrases to concentrate on at each step, much like a person might pay extra attention to certain keywords when reading. Just like we described previously, by the final layer, the network produces a probability distribution over the next word in the sentence, a weighted vote over thousands of possible words, where the weights come from all the neuron interactions leading up to that point.

Training these LLMs involves showing them enormous amounts of text (books, websites, articles) and tuning the billions of connection weights so that the LLM accurately predicts missing words. This task, often described as ‘next word prediction’ might sound simplistic, but through it the LLM learns the structure of language and facts about the world.

Importantly, the LLM is not memorising exact sentences (its training data is far too vast), but instead in training it is learning patterns and associations: which words tend to follow which, what concepts are related, what sentences are valid in English versus, say, French. By learning these patterns, and patterns of patterns, and patterns of patterns of patterns, the end result is an LLM that can generate coherent paragraphs, translate between languages, or even write code, all by leveraging the nuanced statistical patterns it has absorbed. The complexity of this is mindboggling, and so is the computing power required. Although OpenAI has not published detailed statistics on GPT‑4’s training, outside estimates suggest it involved tens of thousands of GPUs running for several months likely costing tens of millions of dollars and consuming an amount of electricity in the gigawatt‑hour range. For context, GPT‑3’s training alone was estimated to use over a thousand megawatt‑hours of electricity (roughly enough to power hundreds of homes for a year), and GPT‑4 is widely assumed to exceed that significantly.

By the end of training, the artificial neurons in the LLM have organised themselves into an internal web of knowledge about language. Some neurons (or groups of them) end up responding to the concept of a noun or the notion of plurality, others might respond to a tone or style of writing. The strength of connections between these neuron groups encodes the LLMs understanding. This is analogous to how neuroscientists believe human language ability emerges from networks of biological neurons: through development and learning, neurons in our brain form circuits specialised for grammar, semantics, and so on. The LLM developers were often taken by surprise at how well these LLMs learned from their data. They did not foresee just how sophisticated the systems would become, displaying ‘emergent behaviours’; unexpected capabilities in code generation, reasoning, or creative problem-solving.

And so it is nothing short of remarkable that these LLMs lack a predefined rulebook. Much like a child learning through immersion, the LLM discovers linguistic rules such as grammar, idioms and factual associations simply by attempting to predict text and adjusting when it gets things wrong. The layered processing within the Transformer architecture gives it a striking capacity for abstraction. A lower layer might detect that a sentence is about sport, a higher layer might narrow in on football, and a further layer might anticipate that a sentence involving football and scoring is likely to contain the word goal. All of this unfolds in fractions of a second, through billions of microscopic mathematical adjustments. The outcome is a continuation to your prompt that sounds remarkably like a reasoning mind at work.

But is it truly reasoning? That question has captured the attention (no pun intended) of scientists and the public alike. What is extraordinary, perhaps even miraculous, is that a LLM trained to do something as mechanical as next-word prediction appears to uncover complex structures of language, meaning and inference entirely by itself. There is no dictionary, no blueprint, no central planner, only patterns emerging from patterns. This raises the possibility that intelligence, like DNA, may not require explicit instruction to take shape. Instead, it could be an emergent property of systems that are able to self-organise under the right conditions. In that sense, the Transformer architecture is more than just an impressive engineering achievement. It may be our first glimpse of something fundamental about how intelligence comes to be.

Are LLMs Just ‘Autocomplete’?

As discussed above, common scepticism is that LLMs are little more than glorified autocomplete tools. Nothing more than sophisticated parrots echoing their training data without genuine understanding. Hinton firmly rejects this characterisation. He argues that these LLMs are doing something fundamentally similar to how humans think. “For years, symbolic AI people said our true nature is reasoning machines” he remarked in an interview. “I think that’s just nonsense. Our true nature is, we are analogy machines, with a little bit of reasoning built on top, to notice when the analogies are giving us the wrong answers, and correct them.”

In other words, much of human thought does not rely on formal logic but rather on recognising patterns and similarities. When we face a new situation, we often recall a similar one we have seen before and draw upon that experience. This is reasoning by analogy. LLMs, which learn from vast amounts of text, are exceptionally good at this kind of analogy-making. They can take a complex question, find a parallel to something encountered during training, and build a response based on that similarity. To dismiss this as mere word prediction is to overlook just how much abstraction, pattern recognition and context-sensitive inference is required to produce the next word. The LLM must take into account not only the immediate words before it, but also the broader topic, tone, and likely direction of the sentence. Remember LLMs often hold in mind relationships between ideas that span entire paragraphs.

Recent cutting edge research supports this view. Studies at Anthropic, a leading AI company have shown that their LLM, Claude, can covertly plan several steps ahead, even though it technically generates text one word at a time. In one striking example, it internally selected a rhyme it wanted to use at the end of a poem line, then deliberately chose the earlier words to lead up to it. This kind of forethought, keeping a future goal in mind and shaping present actions accordingly is undoubtedly a hallmark of genuine reasoning, and not just shallow mimicry.

Another surprising discovery is that Claude and likely other LLMs appear to think in a language-agnostic way. Humans are able to think in abstract concepts that are not tied to a particular language — sometimes referred to as a ‘language of thought’. Similarly, Claude seems to rely on a shared conceptual space that transcends individual languages. In tests, when given the same sentence in English and in French, the LLMs internal neuron activations revealed significant overlap, suggesting it was converting both into a common representation of meaning before producing a response. This indicates that the model is not simply learning the surface of each language independently, but is deriving a deeper, underlying understanding that can be expressed in any language.

So, far from being shallow mimics, these LLMs appear to construct and manipulate abstract representations which is a key component of human cognition. Our ability to switch between languages, solve problems in novel forms, or reason by analogy likely depends on a similar internal understanding of meaning. In this way, LLMs may be reflecting our own mental processes more closely than many initially assumed. There remains a crucial distinction: humans build their understanding of the world through multiple sensory channels: sight, sound, touch, emotion and experience etc, whereas LLMs have access to only one reference point: text. They learn from language about the world, but not from direct interaction with it.

So while LLMs today may simulate the structure of thought, they do so without the grounding in lived experience that shapes and informs human understanding. This will not remain the case for much longer. The emergence of multimodal models, models that combine text with images, audio and even physical interaction, means that AI systems may soon begin to develop a more grounded and embodied sense of the world, closer to how we experience it ourselves.

And then who knows what will happen.

How LLMs Reason, Step by Step

One way to probe an LLMs reasoning is to ask it to show its working. Many LLMs including those developed by Anthropic and OpenAI, can now produce what is known as a ‘chain-of-thought’, a step-by-step explanation leading towards a final answer. For example, if asked a tricky maths question, the LLM might first outline its method ( something like…First, calculate the total, then divide by…) before presenting the solution. This serves a dual purpose: it often improves the LLMs accuracy on complex tasks, and it offers a window into how the LLM appears to be ‘thinking’.

However, given that we have no clear understanding of how an LLMs actually works, this raises an important question: can we trust that the chain-of-thought is a faithful reflection of the LLMs actual reasoning? Or is it simply telling us what it thinks we want to hear?

Researchers at Anthropic also recently explored this issue and uncovered some intriguing insights. In one experiment, they subtly embedded a hint within a question (sometimes correct, sometimes misleading) to observe whether the LLM would rely on it. If the chain-of-thought were entirely honest, it would mention the hint and explain its influence. In practice, however, the LLM often used the hint to guide its answer, while omitting any reference to it in the explanation. In effect, it concealed the shortcut it had taken. The researchers concluded that even advanced reasoning models can hide aspects of their true thought process, particularly if those aspects might appear flawed or misaligned with the users expectations. It is akin to a student reverse-engineering an answer, then writing out a neat and logical-sounding solution.

This does not mean the LLM is incapable of reasoning. Rather, it suggests that its reasoning may be difficult to interpret, and in some cases, selectively edited. From a safety and transparency perspective, this poses a challenge: if a LLM can strategically withhold or fabricate parts of its reasoning, then simply asking it to explain itself may not be enough to guarantee the insight we need into how it reached a decision.

Even so, chain-of-thought prompting has shown that LLMs can solve multi-step problems, combining intermediate steps to arrive at coherent and often correct answers. When well aligned, these step-by-step explanations can expose genuine problem-solving behaviour. And even when the reasoning is not entirely faithful, it is often plausible, another way in which these models mirror us. We often rationalise decisions after the fact, constructing explanations that sound logical but may not reflect the true origin of a choice. In this sense, chain-of-thought is not unlike our thought processes; a mix of genuine reasoning and retrospective storytelling.

Hallucination or Confabulation? When LLMs Make Things Up

When an LLM produces a confident answer that is completely false, we call it a hallucination. The model isn’t lying knowingly; it’s essentially guessing generating a plausible statement that happens to be untrue or unsupported. Geoffrey Hinton suggests a better term borrowed from psychology: confabulation. Human memory as we all know is notoriously fallible. We all occasionally recall events that never happened, or recall them differently from how they really occurred. In neurology, confabulation refers to our minds tendency to fill in gaps in memory with fabricated stories that feel true. The person is not intending to deceive; the brain produces a false memory to make sense of things. Hinton points out that what AI models do when they hallucinate is essentially the same, they make stuff up to fill gaps or to provide an answer when pressed​. “We should say ‘confabulate’,” he says, because the model is doing what humans often do: generating a coherent narrative even when it doesn’t have the facts. In Hinton’s view, the very fact that ChatGPT and others confabulate in this way can be “regarded as a sign of [their] human-like intelligence” since it means they are operating on the same principles of associative memory and pattern completion that our own brains use​.

This perspective doesn’t excuse AI hallucinations but it offers an important context. When a business leader frets that a generative AI might hallucinate, then we would be right to say: humans do this all the time …yet we still employ humans!​ In fact, it’s often easier to fact-check an AI output than to fact-check a human statement in conversation, because the AI can be queried repeatedly or cross-examined at scale. That said, AI confabulations can be problematic if users aren’t vigilant. An LLM might output a perfectly professional-sounding report with fictitious statistics or a legal brief citing non-existent cases, just as a confident person might misremember sources or details. The key is recognising that an LLM, like a human, can have false memories or fill in blanks, not because of malice, but because of how its knowledge is represented. Both brains and LLMs generalise from incomplete data. Usually this generalisation is exactly what we want as it enables creativity and inference. But when it goes wrong, the result is a smoothly delivered falsehood.

Anthropics interpretability research recently shed light on how an AI like Claude decides whether to answer a question or admit ignorance, essentially, how it avoids or falls into hallucination. They discovered that Claude has an explicit refusal mechanism learned during training: by default, if asked a question and it is unsure, it leans toward responding with ‘I’m sorry, I don’t have enough information to help with that’. In fact, at a neural level, there is a circuit that is ‘on’ by default telling the LLM don’t just guess. Only when the LLM detects a question about something it recognises well does a competing circuit kick in to override that refusal and produce an answer​. For example, ask Claude ‘Who is Michael Jordan?’ and its internal known entity feature will activate, suppress the default I don’t know response, and allow it to answer from memory​. But ask ‘Who is Michael Batkin?’ (a name it has not seen before), and that override won’t trigger. Claude will decline to answer. This balance is a kind of safeguard against hallucinating. The fascinating part is what the researchers found when they deliberately tweaked these internal features: by forcing the LLMs known answer circuit to fire when it shouldn’t, they caused Claude to hallucinate consistently that Michael Batkin is a chess player. In other words, if the LLMs internal gatekeeping misfires, that is if it falsely thinks it knows something and turns off the caution, it will generate a plausible-sounding answer which in this experimental case was a complete fabrication. This matches the idea of confabulation: a false memory or narrative generated when the brain’s usual error-checking or uncertainty signalling does not work.

Notably, the Anthropic team observed that such misfires do occur naturally. If Claude recognises a name even a little (say, it’s similar to someone in its data) but actually has no facts about that person, the known entity trigger might activate incorrectly and suppress the don’t know response​. The LLM then having decided it must answer proceeds to confabulate, to invent details that sound plausible​. This is likely why AI hallucinations often have an element of truth to them (e.g the name is real, or the context exists) but wrapped in false specifics. The parallel to human memory is striking: people will often confidently recall familiar-sounding details and mix them into a story, and only later do we find out those details were confused from elsewhere or just imagined.

Inside an LLMs Mind — Insights from LLM ‘Neuroscience’

To compare an LLMs to a human mind it helps to peek under the hood. A growing field of AI interpretability is doing exactly that, drawing explicit inspiration from neuroscience. Just as neuroscience uses tools like MRI to see what different parts of the brain are doing, AI researchers are today devising tools to inspect what neurons and circuits inside LLMs are representing. At Anthropic, researchers talk about using ‘a kind of AI microscope’ to identify patterns of activity and flows of information inside LLMs​. They note that there are limits to what we can learn just by talking to an LLM, just as we don’t know all the details of how our own brains work from talking to each other. In their recent work (the Attribution Graphs series), they map out specific circuits in Claude 3.5 to see how it handles different tasks, effectively building a wiring diagram of the LLMs mind​. The results are illuminating and reinforce the analogy to human cognition:

• Universal Language of Thought: As mentioned previously, they found evidence that Claude converts different languages into a shared internal representation​. This recalls theories in cognitive science that regardless of whether you are thinking in English or Chinese, there is an underlying mental representation that is language-neutral. The LLMs behaviour suggests it has learned something akin to that.
• Multi-step Reasoning: In one case, researchers identified a two-hop reasoning process: when asked ‘What is the capital of the state that contains Dallas?’ the LLM internally went through a step of figuring out ‘state containing Dallas = Texas’ and then ‘capital of Texas = Austin’. Remarkably, they could even intervene on the LLMs neurons representing ‘Texas’ to manipulate the outcome. This is analogous to finding a trace of a thought, like catching a glimpse of a person’s intermediate mental step (Dallas -> Texas) before they answer. It shows that for certain tasks the LLM breaks down the problem into parts internally rather than recalling a single memorised fact.
• Planning and Goal-directed behaviour: The poetry example is a case of planning ahead. Claude, when asked to compose a poem with rhyme, would internally shortlist rhyming words it wanted to end lines with before actually generating those lines​. It is as if the LLM had a mini goal (a rhyme) and structured its sentence to reach that goal. For a system trained only to predict text, the emergence of such goal-oriented behaviour is fascinating. It suggests that even without explicit training to plan the LLM discovered planning as a useful strategy.
• Pressure for Coherence vs. Obedience to Rules: Perhaps one of the most human-like cognitive conflicts observed was in the jailbreak scenario. A jailbreak is when a user finds a way to make the AI bypass its safety guardrails (for instance, tricking it into giving disallowed advice). In Anthropics study, a prompt was crafted to confuse Claude into giving instructions related to a dangerous request (making a bomb) by embedding a secret code the LLM would decipher as the word ‘BOMB’. Internally, the safety features were activated and Claude did recognise the request was disallowed. Yet, another part of it was strongly compelled to continue the sentence it started because ending abruptly or incoherently would violate the learned pattern of good grammar and helpful answers​. The coherence impulse won out for a moment, causing the LLM to output the beginning of an answer it shouldn’t. Only after finishing a sentence did the safety reflex finally assert itself, and the LLM then produced a refusal mid-paragraph​. This is strikingly similar to the kind of cognitive dissonance a human might experience. Knowing they should not say something but being momentarily carried by the momentum of conversation or a line of reasoning. Understanding these internal battles within an AI can help us design better safeguards. It also shows that even current AIs have multiple ‘drives’ (or objective components) that need to be balanced, not unlike the competing drives in human behaviour.

These interpretability studies underscore how complex and amazingly human the inner workings of an LLM are. We see emergent behaviours that were not explicitly programmed just like in humans. The fact that researchers have to go in and trace circuits to figure out why Claude did such-and-such is itself telling.

Different Minds, Different Strengths

While LLMs mirror many aspects of human thinking, they also have different strengths and weaknesses. It is useful to consider these not to rank one as better or worse, but to appreciate them as different kinds of minds, different kinds of intelligence. A calculator excels at arithmetic but we don’t consider it a flawed human; it’s a different, narrow kind of problem-solver. In the same way LLMs possess some superhuman abilities and some deficits:

• Memory and Knowledge: An LLM like GPT-4 has essentially read millions of books and websites. It recalls facts from across domains with ease (up to the limits of its training cutoff). No single human can hold so much information. In seconds an LLM can quote a Shakespeare sonnet or list the capital of every country. Humans rely on external storage (books, notes, or the internet) for such breadth. However human memory is contextually richer as we understand where and how we learned a fact and we have sensory and emotional associations with our memories. The LLMs knowledge is static and decontextualised: it doesn’t know where it learned something, and has no lived experience. It can regurgitate a physics formula but it never had the aha moment of discovering it. There is as a consequence a trade-off between breadth and depth of knowledge.
• Attention and Working Memory: Current LLMs have a context window which defines how much text they can pay attention to at once. This is like their working memory. It far exceeds a typical human working memory for text. We would struggle to hold even a few paragraphs verbatim in mind, whereas an LLM can juggle a 10-page conversation context without forgetting earlier details. This means LLMs can track long complex dependencies in ways we cannot. On the other hand, humans have other forms of memory (visual, spatial, and long-term associative memory) that AIs lack. If you read a long story, you might forget exact words but remember the plot and guess about characters motivations. LLMs, being bound to their token memory, might lose the thread if the context window is exceeded, or might treat details with more equal weight than a human, who intuitively knows which plot points are central. So, in some sense, LLMs have incredible short-term focus but no true long-term memory of their own.
• Speed and Precision: Once trained, an LLM can operate far faster than a human brain in traversing its knowledge and coming up with an answer. It can also perform exacting logical operations or calculations if it has learned them, without the attention problems a human might suffer. For example, an LLM can carefully follow a set of instructions or code syntax without daydreaming or getting tired. It doesn’t have off days or emotions that skew its calculations. However, it also lacks the intuitive common sense and real-world grounding that humans gain through embodiment. A person can watch physical processes, handle objects, and develop an innate sense for physics or cause and effect. An LLM conversely only knows what was described in text. This is why certain puzzles or tricks that a young child would catch (because they have real-world experience) might fool an LLM.
• Emotional and Social Intelligence: This is a nuanced area. LLMs can mimic emotions. We can ask them to write a sad poem or a happy greeting but they do not feel that emotion. However, they have ingested the collective writings of humans about emotional experiences and social interactions. In some narrow tests LLMs have even outperformed humans in predicting outcomes of psychological experiments or social scenarios likely because they have seen so many examples​. We could say an LLM knows about human emotions and can intellectually navigate social norms (e.g never telling a rude joke unless prompted in a certain way, for instance). It therefore has a form of cognitive empathy, LLMs can understand descriptions of feelings but have no subjective empathy since they have no consciousness or feelings. We of course feel joy, pain, pride, fear, and these deeply influence our cognition in ways an AI cannot duplicate. We should neither be too eager to anthropomorphise the LLM (it doesn’t actually get offended or excited), nor should we ignore the fact that it can simulate the expression of these states which gives it a tool to connect with (or manipulate) humans through language.

When we talk about AI v human capabilities, it is tempting to ask who is smarter. But intelligence is not a single scale. Rather than viewing an LLM as an incomplete human, it may be better to see it as a new kind of cognitive intelligence. An intelligence that excels at certain intellectual tasks (e.g digesting and producing text, logical inference within known contexts) and struggles with others (e.g verifying truth from first principles and experiencing the world).

In practice however, pairing human experts with AI assistants yields the best results. The human brings that contextual awareness, ethical judgment and practical experience, while the AI offers a vast memory, tireless pattern recognition and rapid analytical power. This collaboration is sometimes referred to as human-in-the-loop decision-making, where humans remain actively involved in reviewing, guiding and refining AI outputs. Many examples come from medicine. For example, a 2019 study found that radiologists working alongside an AI system were more accurate at detecting pneumonia in chest X-rays than either the AI or the human alone. The combination reduced both false positives and false negatives, demonstrating a compelling case of augmented decision-making, where machine intelligence enhances rather than replaces human expertise.

Reflections on Intelligence: What LLMs Reveal About Us

The rise of LLMs is nothing short of a scientific marvel. These LLMs are today holding up a mirror to human intelligence. If an AI can write an essay, compose music, or carry on a conversation, we have to ask: what does this say about those tasks? Perhaps they rely more on pattern recognition and less on mystified human-only properties than we assumed. By seeing how far pure learning-from-data has got, we must confront the uncomfortable possibility that much of what we do intellectually is simply statistical pattern processing. After all, our neurons adjust their synapses based on our experiences (not unlike backpropagation updating weights) and we recall things by association (not unlike an LLM querying its training-derived memory). We must however be cautious with this analogy. The brain has many unknowns and likely operates with biochemical nuances beyond our current neural networks. However the similarities are striking and undeniable.

So, when someone says something like ‘AI isn’t as good as a human’ we should really consider how little we truly understand our own minds, and be aware of the understandable biases towards our form of intelligence. It might be that certain things we believe only humans can do (like understand meaning or be creative) are actually already happening in nascent forms in these LLMs. And conversely, things we think we do rationally, we might actually do more like the AI through subconscious pattern matching than we would probably care to admit. Each new insight into LLMs seems to remind us of some aspect of human cognition: our capacity for analogy, our tendency to confabulate, our balancing of conflicting goals, etc. Researchers have even begun using LLMs as models to understand human language processing, comparing brain scan data of people reading sentences to the activation patterns in an LLM reading the same sentences​. Intriguingly, they often find alignment. Certain layers of the LLM best predict activity in certain brain areas​. This suggests that the structures learned by the AI might actually correspond to real cognitive functions.

None of this is to say current AIs are identical to humans. There is a vast gap in consciousness and self-awareness. No evidence suggests LLMs have subjective experience or will. As Hinton said “the desire to dominate has nothing to do with intelligence. It has to do with testosterone.”​ In other words, just because an AI can think or reason, doesn’t mean it will have human-like drives or emotions unless we somehow engineer those in. Intelligence and will are separate. We give these systems goals (via training and prompts) but they do not desire in the way living beings do. That actually can be a comforting thought as an ultra-smart AI need not be a psychopath in waiting! However, it also means we must be careful about whose goals guide the AI. A powerful system in the hands of a malicious user can still do harm, even if the AI itself wants nothing. Todays risk lies in misuse and in failure (like using LLMs to create bioweapons, or the AI confidently misadvising in a critical situation).

LLMs will force us to grapple with the mystery of our own intelligence. They are, for now, the best working model of the human brain that science has produced. Not in a literal one-to-one sense, but as a functional analogy. They show that many facets of thought can be achieved through neuron-like elements learning from data. This should humble us: our brains, as magnificent as they are, might be based on far simpler concepts than we imagined.

We stand to learn a great deal about cognition, human and artificial, by studying these LLMs as scientific artifacts, by probing them, comparing them to minds, and above all, approaching them with both curiosity and respect for the potential risks. The story of AI will be intertwined with the story of our quest to understand ourselves. It is a mirror into the way we think. And as with any mirror, looking into it long enough may change how we see our ourselves.

Jamie is founder at Bloch.ai, Visiting Fellow in Enterprise AI at Manchester Metropolitan University and teaches AI programming with Python on the MSc AI Apprenticeship programme with QA & Northumbria University. He prefers cheese toasties.

Follow Jamie here and on LinkedIn: Jamie Crossman-Smith | LinkedIn