Extended Cognition and Artificial Intelligence

Are LLMs like ChatGPT or Claude already an extension of our mind?

An Unexpected Question

The Extended Mind & AI — Corrected Version

You're reading this article. Maybe on a phone, maybe on a laptop. The screen glows, your eyes move, and somehow the ideas in these words make their way into your head. But here's a question that might seem strange at first: Where exactly does your mind end?

Most people would say: "In my brain, of course. The brain thinks."

But what if I told you that philosophers have been arguing for decades that your mind might extend beyond your skull? That the notebook on your desk, the calculator on your phone, even this text on your screen might be part of your thinking?

And now, with the rise of Large Language Models like ChatGPT, Claude, and Gemini, this question becomes even stranger: Can an AI be part of your mind?

The Question That Started Everything

In 1998, philosophers Andy Clark and David Chalmers published a paper called "The Extended Mind" that would change how we think about thinking.

Their opening question was deceptively simple:

"Where does the mind stop and the rest of the world begin?"

They noticed that we use external tools all the time:

  • We write notes so we don't forget
  • We use calculators instead of doing math in our heads
  • We check our phones for information we used to memorize

But here's the key insight: when you use a notebook to remember something, is the notebook just a tool, or is it part of your memory?

Clark and Chalmers said: sometimes, the notebook is your memory.

Otto and Inga: A Thought Experiment

To understand this, imagine two people:

Inga sees a museum poster and remembers the address using her biological memory. She thinks, recalls the address, and knows where to go.

Otto has Alzheimer's. He can't remember addresses anymore. But he carries a notebook everywhere. When he needs an address, he checks his notebook, thinks, and knows.

Here's the question: Is there any meaningful difference between how Inga "remembers" and how Otto "remembers"?

Both:

  • Have access to the information when needed
  • Use that information to guide their actions
  • End up knowing the address

The only difference is that Inga's "memory" is in her brain, while Otto's "memory" is in his notebook.

Clark and Chalmers argued: if both work the same way, both count as memory. The notebook isn't just a tool; it's an extension of Otto's mind.

The Three Criteria

Not every tool makes the cut. Clark and Chalmers proposed three conditions for something to be part of your mind:

1. Reliability
The tool must be reliably available when you need it. Otto's notebook is always with him, just like Inga's brain.

2. Automatic Endorsement
You trust the tool without constantly questioning it. When Otto checks his notebook, he believes what it says. Inga believes her memory the same way.

3. Direct Access
You can retrieve information from the tool as easily as from your brain. No special effort required.

By these criteria, many things become part of our extended mind:

  • Notebooks and calendars
  • Smartphones with contacts and notes
  • Spreadsheets and documents
  • Even other people (when we delegate memory to them)

The Deeper Point: Cognitive Scaffolding

The philosopher Andy Clark, who continued developing this idea, argued that humans are natural-born cyborgs. We evolved not just brains but brain+tool systems. Our skulls contain only part of our cognitive apparatus.

Think about what happened when humans invented writing:

  • Before: everything had to be memorized
  • After: we could offload memories to clay tablets, papyrus, paper
  • Now: our smartphones store more information than any human could remember

This isn't weakness. This is cognitive scaffolding — using external resources to support and extend our thinking.

The interesting consequence: we've never truly been "just our brains." Our minds have always been distributed systems.

Enter Large Language Models

Now enter LLMs. Systems like ChatGPT, Claude, Gemini, and dozens of others.

You've probably used them. You type a question, they respond. You iterate, they adapt. You might even feel like you're having a conversation.

But here's the uncomfortable question: Are these just fancy autocomplete machines? Or is something more happening?

How LLMs Actually Work

Before answering that, let's understand what's happening under the hood.

The Simple Version:

An LLM is a system trained to predict the next word. That's it. Given "The capital of France is", it predicts "Paris."

This might sound trivial. But here's the thing: to predict text accurately across all possible topics, the model has to develop something that looks remarkably like understanding.

The Process:

  1. Tokenization: Your text is broken into small pieces called "tokens" (roughly 3–4 characters each)

  2. Embedding: Each token becomes a number (a vector) that captures its meaning

  3. Attention: The model figures out how each word relates to every other word in context

  4. Prediction: The model calculates probabilities for what comes next

  5. Generation: It picks the next token and repeats until done

The "attention mechanism" is crucial. It lets the model understand relationships — like how "bank" in "river bank" means something different from "bank" in "bank account."

The Architecture: Transformers

Modern LLMs use something called the Transformer architecture, introduced in 2017.

Before Transformers, language models processed text sequentially — one word at a time. Transformers changed everything by processing all words simultaneously through attention.

Imagine reading a sentence but seeing every word at once, with invisible threads connecting "she" to "Mary" and "her" to "sister." That's what attention does.

Transformers can be stacked: GPT-4 has around 120 layers, each refining the understanding of the text. The early layers handle grammar; later layers handle meaning and reasoning.

The Training: Learning from Humanity's Text

LLMs are trained on enormous datasets — billions of words from books, websites, code, and more.

The training objective is simple: predict the next token. Show the model "The cat sat on the" and it learns that "mat" is a likely completion.

But here's where it gets strange: to predict accurately across all of human knowledge, the model must implicitly understand:

  • Physics (how objects behave)
  • Psychology (how people think)
  • Logic (how arguments work)
  • Social norms (how we interact)

This is why the simple task of "next token prediction" produces something that seems to understand.

The Big Debate: Understanding or Imitation?

Here's where things get philosophical — and contentious.

The "Stochastic Parrots" View

In 2021, researchers Emily Bender, Timnit Gebru, and colleagues published a paper called "On the Dangers of Stochastic Parrots."

Their argument:

  1. LLMs are trained on statistical patterns in text
  2. They learn which words tend to follow other words
  3. When they generate text, they're just recombining patterns
  4. This isn't real understanding — it's sophisticated imitation

The metaphor: a parrot can say "I love you" without understanding what love means. LLMs might be doing something similar with language.

Evidence for this view:

  • Hallucinations: LLMs confidently state false facts
  • No grounding: They've never seen a sunset or felt rain
  • Shallow objectives: The goal is just next-token prediction
  • Sensitivity to wording: Small prompt changes affect outputs dramatically

The "Real Understanding" View

Others argue that dismissing LLMs as parrots oversimplifies the issue.

Evidence:

  • Functional competence: LLMs answer novel questions, solve new problems
  • Reasoning chains: They work through multi-step problems
  • Theory of mind: They model what other agents believe
  • Compression insight: To predict accurately across all text, you must understand the world

The researcher Ilya Sutskever (cofounder of OpenAI) argued something profound: to predict the next token accurately across all of human text, the model must implicitly simulate the physics, psychology, and logic of the world that generated that text.

Prediction at scale requires understanding.

A Spectrum, Not a Binary

Here's what might be true: understanding isn't all-or-nothing.

Think of different levels:

  • A lookup table (no understanding)
  • Pattern matching (shallow understanding)
  • Structured knowledge (medium understanding)
  • Human expert understanding (deep understanding)

LLMs might have real but limited understanding — more than parrots, less than humans.

They might have structural understanding (how concepts relate) without experiential understanding (what it feels like).

The Extended Mind Meets AI

Now comes the interesting convergence.

Andy Clark, who co-authored the original Extended Mind paper, recently wrote about LLMs in Nature Communications (2025). His position: generative AI extends our minds in new ways.

The key insight: we've always built hybrid thinking systems. Now we're building a new kind.

Are LLMs Extended Minds?

Some philosophers argue that certain LLMs already qualify as extended cognitive systems — not just tools, but parts of the cognitive circuit.

Consider a system called "Digital Andy": an LLM augmented with a database of Andy Clark's own writings. When you ask Digital Andy about philosophy, it retrieves relevant passages from Clark's work and generates responses sensitive to those ideas.

By the Clark-Chalmers criteria:

  • Reliability: The database is always accessible
  • Automatic endorsement: We trust what it retrieves
  • Direct access: Information flows when needed

Some argue this isn't just extended cognition — it's amplified cognition. LLMs don't just store information like Otto's notebook; they generate new possibilities that didn't exist before.

The "Amplified Mind" Thesis

Philosopher David Matta proposed a new concept: the Amplified Mind.

The distinction:

  • Extended: You offload existing cognition to tools
  • Amplified: The tool generates novel cognitive possibilities

With a notebook, you store what you know. With an LLM, you explore what you don't know yet.

The interaction becomes a "cognitive dance" — you propose, it expands, you redirect, it elaborates. Neither party alone achieves what the conversation enables.

The Metacognitive Mirror

Perhaps the strangest effect: interacting with LLMs changes how we think about thinking.

To engage effectively with these systems, we must:

  • Articulate thoughts more precisely
  • Consider multiple perspectives explicitly
  • Engage in structured reasoning

The LLM becomes a mirror for our metacognition — our ability to think about thinking. It reflects not just our thoughts but the patterns and processes of thought itself.

What This Means for You

Let's bring this home.

When you use an LLM to:

  • Brainstorm ideas for a project
  • Debug code
  • Explain a concept you don't understand
  • Work through a problem

Ask yourself: Is this just a tool? Or is it becoming part of my thinking?

The honest answer might be: both. And that's not new.

Every time you used a search engine, a calculator, or even another person to think through something, you were extending your mind.

The new part is that LLMs don't just retrieve. They generate. They propose. They challenge.

The Practical Implications

If extended cognition is real, several things follow:

1. Memory changes
We offload more to digital systems. This might reduce what we memorize — but it might not matter if we can access what we need.

2. Thinking becomes more distributed
Your thoughts aren't just in your head. They're distributed across you, your tools, and now your AI assistants.

3. Agency becomes shared
When an LLM suggests an idea, whose idea is it? The question might not have a clear answer.

4. Trust becomes technical
Extended cognition requires trusting your cognitive extensions. With LLMs, this means understanding their limitations: hallucinations, biases, context window limits.

The Limits of the Metaphor

We should be careful not to overextend the metaphor (pun intended).

A notebook doesn't generate new notebook entries. An LLM generates new text. The cognitive relationship is different.

Some philosophers argue that current cloud-based LLMs aren't truly extended because:

  • They have "session amnesia" (they forget between conversations)
  • They lack the continuity of a personal notebook
  • They're not reliably available (internet required)

Future systems with persistent memory and personalization might change this.

The Future: Proximal Integration

Philosopher Michael Wagner proposes an interesting concept: Proximal Integration.

Currently, LLMs exist as "distal" tools — we access them from a distance, like consulting an encyclopedia.

A truly extended mind partner would be proximal — integrated into your cognitive flow as seamlessly as your own thoughts.

Imagine:

  • An AI that remembers your projects across months
  • That learns over time and adapts to your thinking style
  • That can identify gaps in your reasoning and proactively address them
  • That becomes genuinely yours through continuous interaction

This might be where we're heading.

Conclusion: The Hybrid Mind

Here's what I think is happening:

We've always been hybrid cognitive systems. Biological brains extended by cultural technology. Writing, printing, computing — each extended what thinking could be.

LLMs represent the next step: not just storage and retrieval, but generation and collaboration.

Whether we call this "extension," "amplification," or something else, the phenomenon is real: human-AI interaction creates cognitive patterns that neither party achieves alone.

You are not just your brain. You never were.

And now, increasingly, you're not just you — you're you-plus-AI, thinking stuff that neither you nor the machine would think independently.

The question isn't whether this is "mind extension." The question is: what kind of extended mind do we want to become?

References

  • Clark, A., & Chalmers, D. (1998). The Extended Mind. Analysis, 58(1), 7-19.
  • Clark, A. (2008). Supersizing the Mind: Embodiment, Action, and Cognitive Extension. Oxford University Press.
  • Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.
  • Clark, A. (2025). Extending Minds with Generative AI. Nature Communications, 16, 4627.
  • Smart, P., Clowes, R., & Clark, A. (2025). ChatGPT, Extended: Large Language Models and the Extended Mind. Synthese.
  • Bender, E., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT '21.
  • Mahowald, K., Ivanova, A., et al. (2024). Dissociating Language and Thought in Large Language Models. Trends in Cognitive Sciences.
  • Pérez-Verdugo, M. (2025). Generative Midtended Cognition and AI. Synthese.
  • Wagner, M. (2025). The Gift of Language: Large Language Models and the Extended Mind. PhilArchive.