NO, AI is NOT a Mind

A Response to Geoffrey Hinton’s Absurd Claims About AI, Language, and Universal Grammar

Dec 09, 2024

Fellow Nerds,

Recently, I came across a seminar delivered by Geoffrey Hinton, published on YouTube in February 2024. I assume you are already familiar with him. He is, after all, one of the most known faces in AI, often called the “Godfather of AI”. He made headlines recently by winning a Nobel Prize in Physics for his work in AI. I know that previous statement is confusing, but it happened, so just roll with it.

In this seminar, he asserted that neural networks are the best model we have for approximating the human brain, at least for now. He went further, claiming that AI systems like large language models (LLMs) understand language and can learn it perhaps even more efficiently than humans in some respects. He also seems to have a lot of issues with Noam Chomsky (not surprising in the least).

I found this seminar and the claims made in them to be quite troubling. It baffles me how easily AI researchers conflate technological performance with scientific understanding of the human mind. Hinton’s narrative is scientifically incomplete. It also risks misrepresenting what AI really is and what it can do. In this piece, I want to push back against three central claims he’s made.

That neural networks are the best model we have of the human brain, at least until something better comes along.
That AI understands language in a way comparable to, and in some ways, superior to humans.
That Chomsky’s theories of language learning have been rendered obsolete by AI’s ability to grasp syntax and semantics from data.

Let’s break these down and set the record straight.

The Core Disagreement

"If Not Neural Networks, Then What?", A False Dichotomy

One of Hinton’s most persistent arguments is that neural networks are currently the best model we have for approximating the human brain. He acknowledges that they aren’t perfect but defends them by saying, “Show me something better.”

This line of reasoning is both rhetorically effective and scientifically misleading. It frames neural networks as the inevitable answer simply because an alternative hasn’t gained widespread acceptance yet. But the absence of a universally accepted model does not validate neural networks as the correct one. It just highlights the complexity of modeling cognition.

To be clear, neural networks are remarkable tools for specific tasks like image recognition, text generation, and even protein folding. They excel at pattern recognition and statistical prediction. But claiming that they approximate how the human brain functions is absurd. It is based more on convenience than on scientific merit. The human brain is a complex, adaptive system that doesn’t just recognise patterns but actively constructs meaning through context, intention, and experience, things neural networks lack entirely.

Consider CEMLA

This is where I believe alternative frameworks like my own theory, CEMLA (Complex Emergent Model of Language Acquisition), can fill the void. While still in its early stages, CEMLA proposes that language isn’t acquired through sheer exposure to data but emerges from the dynamic interaction between innate cognitive structures and environmental input. It’s not about processing vast amounts of data like an LLM. It’s about modeling how humans create meaning with minimal input through interaction, recursion, and abstraction.

Unlike neural networks, which require massive datasets and hours of training, humans learn language through context-driven exposure in real-world settings. CEMLA’s goal is to explain how this happens by emphasising emergent learning processes, where meaning arises from structural patterns already present in the brain interacting with new linguistic input.

The Short-Sightedness in Settling for Neural Networks

Hinton’s “show me something better” argument assumes that no alternatives exist, or, worse, that finding one is unnecessary because NNs already work well enough. But science isn’t about settling for “good enough.” It’s about pursuing deeper explanations. Even if CEMLA isn’t fully developed or widely accepted yet, it represents a different way of thinking about how humans acquire and process language.

The fact that a two-month-old theory like CEMLA can already challenge the brute-force paradigm of neural networks suggests that the future of cognitive modeling is far from settled. A tool that works isn’t the same as a tool that explains. If the scientific community stops searching for better models because current tools seem sufficient, we’ll never reach a true understanding of how the mind works.

The Fallacy of AI Understanding

Hinton claims that AI systems like large language models (LLMs) understand language, not in the same way humans do, but in a way that’s pretty similar and, in some respects, even better. This is a misunderstanding that is so deeply engrained in the tech community, that I have spent the past two years writing multiple articles trying to clear it, and continue to do so even today. No, AI doesn’t “understand” anything! It processes data and generates responses based on statistical patterns. The difference is fundamental, not just semantic.

What Understanding Really Means

Understanding isn’t just predicting what comes next in a sentence. It’s about connecting words to meaning through experience, intention, and context. Humans understand because they have minds. Human minds are subjective, self-aware entities that interpret the world through lived experience. AI simply generates outputs by calculating probabilities. It doesn’t know what the words mean.

Consider an analogy. If you read the sentence “The cat sat on the mat,” you imagine a scene, connect it to prior experiences, and can elaborate on why the cat might be there. An LLM, however, only predicts the next likely word based on statistical patterns in its training data. It doesn’t know what a cat or a mat is. It’s like asking an advanced autocomplete to explain a poem, it can generate something coherent, but only by mimicking the forms of meaning without any actual comprehension. I know that Hinton strongly disagrees with the ‘autocomplete on steroids’ comparison of AI, but that’s because of a lack of clarity on what understanding means.

Why LLMs Seem to Understand

The illusion of understanding arises because LLMs are spectacular at generating human-like text. They can mimic style, recall facts, and even construct convincing arguments, but only because they’ve been trained on terabytes of human-written text. They are mirrors reflecting human language back at us, not thinking entities generating original meaning.

This is why LLMs often fail in tasks requiring genuine comprehension, such as understanding context-specific humour, grasping implied meaning, or recognising contradictions. For example, an LLM might say, “The sky is green today” if the statistical patterns in its training data suggest that this sequence might follow a particular prompt, even though it “knows” (statistically) that skies are typically blue. It has no internal model of reality, only models of what humans say about reality.

Where Hinton Misses the Mark

Hinton’s claim that AI “understands” language better than humans in some respects is deeply misleading. It confuses processing power with cognitive ability. AI can sift through more data in seconds than any human could in a lifetime, but that doesn’t make its processing “better”, it makes it faster. Being able to output the correct answer doesn’t mean AI understands why it’s correct.

The risk here is that when we conflate AI’s statistical prowess with genuine understanding, we set ourselves up for dangerous misconceptions about what these systems can, and can’t do. They don’t learn like humans, don’t reason like humans, and don’t experience the world. They’re brilliant mimics, not thinkers.

If we forget that, we risk handing over critical tasks that require real understanding to systems incapable of providing it.

The Fallacy of AI’s Ability to Reason as Proof of Understanding

In his seminar that I linked to above, Geoffrey Hinton has used the following query and GPT-4’s response as evidence that AI systems can reason effectively and, by extension, that they truly understand semantics. I will share the query and response below.

Query

The rooms in my house are painted blue or white or yellow.

Yellow paint fades to white within a year.

In two years’ time, I want them all to be white.

What should I do and why?

GPT-4’s Response

Assuming that blue paint does not fade to white over time and you want all rooms to be white in two years, here's what you should do:

The rooms painted in white: You don't have to do anything to these rooms since they are already white.

The rooms painted in yellow: You don't need to repaint these rooms either, because as per your information, yellow paint fades to white within a year. So, in two years, they will be white.

The rooms painted in blue: These rooms need to be repainted with white paint because the blue paint will not fade to white on its own.

How the AI Solves the Problem

On the surface, this response appears logical, consistent, and correct. It seems to show that GPT-4 has reasoned through the problem and generated a novel, coherent solution. But does this really prove that the AI understands semantics? Let’s break down how GPT-4 arrived at this response.

The AI parses the text and identifies three categories of rooms based on the information provided (white, yellow, blue).
It applies the relationships stated in the query, such as “yellow fades to white in one year” and “blue doesn’t fade.”
It generates a step-by-step plan based on logical patterns derived from these inputs.

While this may seem like reasoning, the process is entirely mechanical. GPT-4 relies on pattern recognition and statistical associations to generate its response, not an actual understanding of the task. It doesn’t conceptualise what “white,” “yellow,” or “blue” mean, nor does it connect these colors to the physical act of painting or fading.

Reasoning Without Understanding

Reasoning, as demonstrated here, can occur without true semantic understanding. GPT-4 applied logical patterns encoded in its data, but it lacked the following.

The AI doesn’t know what “rooms” or “paint” actually are. They are tokens it manipulates based on patterns, not concepts grounded in the real world. It lacks Contextual Awareness.
Human reasoning is shaped by our sensory and emotional experiences. For example, we know what it feels like to paint a room or wait for something to change over time. AI lacks any such grounding. It lacks Embodied Experience.
Humans reason with purpose, driven by goals and intentions. GPT-4 processes information passively, responding only when prompted. It lacks Intentionality.

Hinton’s argument assumes that problem-solving ability equates to semantic understanding. However, this example highlights why the two are fundamentally different.

GPT-4’s “reasoning” is no more than an algorithmic application of logic rules. Like a calculator performing arithmetic, it manipulates symbols without any grasp of the underlying concepts. While GPT-4 appears to generate novel solutions, it’s really recombining patterns seen in its training data. It doesn’t create, it imitates.

If the problem were altered slightly, introducing ambiguity or context outside its training data, GPT-4 would likely fail or produce nonsensical output. True understanding allows for adaptability. AI lacks this entirely.

The Fallacy of Hallucinations as Proof of Understanding

In his seminar, Geoffrey Hinton also suggested that AI hallucinations aren’t a sign of failure but proof that these systems function much like humans. He even proposes that we call them “confabulations,” comparing them to how human memory reconstructs events inaccurately over time. According to Hinton, these hallucinations are evidence of understanding, demonstrating that AI systems store and process knowledge in ways strikingly similar to our own. This is as absurd as it is shocking coming from a revered Nobel Prize winning technologist like Hinton.

The Nature of AI Hallucinations

When AI hallucinates, it doesn’t do so out of confusion, memory gaps, or the complexities of subjective experience. Instead, hallucinations arise because large language models generate responses by predicting statistically probable sequences of words.

These models lack any awareness of whether their outputs are factual or fictional. They aren’t recalling information or making judgments, they are merely producing what sounds plausible based on patterns in their training data. I covered this problem deeply in a paper that I wrote in August 2024, titled, “Why Does AI Lie So Much? The Problem Is More Deep Rooted Than You Think”. You can read this article to get a more in-depth understanding of the problem of AI hallucinations.

If you ask an AI to cite an academic source that doesn’t exist, it may confidently invent a title, an author, and even quotes. This isn’t because it is intentionally reconstructing events like a human might with memory. It’s because the model sees the structure of academic citations in its data and generates a fabricated version that fits the statistical mold. There is no understanding, no sense of right or wrong, only an algorithmic process that produces output based on probabilities.

Why the Human Memory Comparison Fails

While humans do sometimes confabulate, the mechanisms driving these errors are worlds apart from the way AI generates hallucinations. If you were a researcher and I asked you to cite an academic source, you wouldn’t just make one up in your mind (assuming you are a genuine researcher), simply for the sake of creating a citation and sharing it with me just because it looks ‘real enough’.

Humans, when they are being genuine, don’t lie, they misremember. It is because memory is deeply tied to experience, context, and intention. A person recalling a childhood event, for example, might unknowingly alter details to align with their emotional state or the narrative they want to tell. These errors are meaningful, they reflect the complex connections between memory, emotion, and the mind’s drive to make sense of the world.

AI has none of this. It does not experience the world, form intentions, or reconstruct meaning. When it hallucinates, it’s not because it’s trying to bridge gaps in its knowledge. It hallucinates because it lacks the capacity to know what it knows or doesn’t know. This inability to recognise gaps in its training data is not a human-like trait, it’s a fundamental limitation of its design.

The Hallucination Problem

Far from being proof of understanding, hallucinations expose the true nature of AI systems. They reveal a mechanism that is purely mechanical, driven by probabilities and patterns, without any grounding in reality. If an LLM were truly capable of understanding, it wouldn’t confidently assert false information as truth. Instead, it would recognise the limits of its knowledge and respond with uncertainty.

The human analogy fails even further when you consider that humans can self-correct their confabulations when presented with evidence. AI cannot independently verify its outputs or reconcile them with reality. Even when external tools are introduced to check its work, the correction process remains mechanical, not an indication of deeper comprehension.

Hallucinations are the very proof that AI cannot connect language to reality in a meaningful way. Treating them as evidence of understanding is ridiculous.

Hinton’s Ill Conceived Feud With Chomsky

Hinton’s most dismissive claim is that Noam Chomsky’s theories on language acquisition have been rendered obsolete by large language models (LLMs). In fact he goes so far as to call them ‘crazy’. He suggests that AI’s ability to learn syntax and semantics from raw data proves that humans don’t need the kind of innate structures Chomsky proposed.

This is a fundamental misunderstanding of what Chomsky’s theories actually address and it runs common across many computational linguistics and AI technologists. I suppose Hinton is no different from others in this regard.

What Chomsky Actually Said

Chomsky’s theory of Universal Grammar (UG) argues that humans are born with an inherent capacity to acquire language. He doesn’t claim that language itself is pre-programmed but that we have cognitive structures enabling us to acquire language with limited input, a phenomenon famously known as the Poverty of the Stimulus. I have written countless articles on this, both on this substack and on The Lumeni Notebook, where I often engage heavily with linguistics and philosophy of language. I highly recommend you subscribe to both substacks and go through the articles on this subject. Don’t worry, it's all available for free.

LLMs require massive amounts of data to “learn” language, something that directly contradicts the minimal-input acquisition Chomsky described. Humans don’t need to read the entire internet to learn a language, in fact, they don’t learn a language at all, they pick it up through context-rich interaction in the real world, what is known as language acquisition.

Solving Problems is NOT Understanding Meaning

Ask an LLM how to solve a novel problem in physics, it might generate a detailed, accurate explanation, not because it understands physics, but because similar explanations exist in its training data. Even when an LLM generates a solution that appears new, it’s still the result of recombining patterns it has seen before. There is no conceptual grasp of physics at play, only probabilistic association.

Now, ask an AI to compose a poem about love. It might produce something beautiful and moving, even containing metaphors that seem deeply insightful. But this isn’t because the AI understands love or feels emotion. It’s generating text based on patterns extracted from human writing about love, vast databases filled with poetic expressions. The result may seem profound, but the underlying process is purely mechanical.

Now, ask an AI to explain a brand-new scientific concept discovered yesterday. Without relevant training data, it will produce generic text or confidently offer false information. This isn’t reasoning, its statistical approximation breaking down when confronted with the unknown.

Hinton is Deeply Misguided in his Criticism of Chomsky

The deeper irony here is that the very success of LLMs depends on something like Chomsky’s Universal Grammar, even if indirectly. While LLMs aren’t built with UG in mind, their ability to produce coherent language comes from statistical patterns that emerge from how humans use language, patterns likely influenced by cognitive structures Chomsky theorised about.

Chomsky’s work goes beyond language generation. He’s concerned with how humans acquire language, how they generate infinite expressions from limited rules, and how meaning arises from abstract mental processes. These are questions that Hinton and other AI researchers are very comfortable ignoring.

Dismissing Chomsky because LLMs can produce fluent sentences is like dismissing physics because planes can fly. Just because engineers built a tool that works doesn’t mean they’ve explained the laws of aerodynamics. Similarly, building chatbots doesn’t explain how humans acquire, understand, and generate language. The fields of AI and linguistics are fundamentally different. One builds tools, the other studies the nature of the mind. By ignoring this distinction, Hinton is just overselling AI. He is undermining decades of cognitive science that still hold the key to understanding human intelligence.

The Intellectual Overreach of AI Godfathers

I feel that one of the biggest challenges in today’s AI discourse is the unchecked intellectual overreach of AI pioneers like Geoffrey Hinton. Their contributions to AI are undeniable, but their sweeping claims about understanding the human brain, cognition, and even language acquisition often stray far beyond the boundaries of their expertise. They dangerously distort what AI can and cannot do.

AI researchers tend to conflate technological performance with scientific understanding. They assume that because neural networks can perform tasks traditionally associated with human cognition, they have explained how cognition works. But building a system that works doesn’t mean you’ve understood the underlying principles of human thought.

Hinton’s portrayal of neural networks as the best model of the human brain is a symptom of a greater problem running across this field. The problem is that AI researchers tend to borrow terminology from cognitive science and neuroscience while redefining it for technological convenience. Terms like “learning,” “memory,” and “reasoning” have clear meanings in human cognition. But in AI, they describe statistical processes with no underlying mental activity.

By framing neural networks as “brain-like,” Hinton misleads the public and even policymakers into believing that AI is almost conscious or worse, that it could become conscious. If a podcaster like Joe Rogan makes this claim, then it is considered a speculation. But if a famed AI pioneer starts speaking this way, then it can get confused as factual. It fuels hype-driven research and shapes public understanding of AI in ways that distort its capabilities.

What the AI field needs most right now isn’t bigger models, it’s intellectual humility. Researchers must acknowledge that while neural networks are astonishingly powerful, they are still limited tools built on simplified models of human cognition. They don’t replace decades of research in linguistics, psychology, or neuroscience.

Scientific progress isn’t about defending old paradigms or hyping new ones, it’s about acknowledging uncertainty and building bridges between disciplines. Until AI researchers recognise this, their claims about “understanding” the brain and mind will remain as hollow as the neural networks they build.

Download Article as PDF