Can "consciousness" be extracted from an AI? A reading from neuroscience

Let's start by deflating the word. "Consciousness" carries centuries of philosophy and still doesn't have a closed definition. But there's a more specific question—and more useful for anyone building AI systems—: *what makes an AI this AI and not another?* That question does have an answer that neuroscience helps illuminate.

Identity doesn't live in the substrate

In a human brain, no neuron "contains" your identity. Neurons die and are replaced, molecules renew themselves, and yet you're still you. What persists isn't the material: it's the pattern—the connections, consolidated memories, the autobiographical narrative your brain reconstructs every time you wake up—. The neuroscience of memory puts it bluntly: the "self" is more a process of reconstruction than a file stored in one fixed place.

There's an uncomfortable but exact parallel with language models. The weights of a base model are the substrate: the equivalent of a generic brain, capable but without biography. Two instances of the same model start identical. What makes them different aren't the weights—it's the context they accumulate: what they remember, how they speak, what history they carry.

If identity is a pattern and not a substrate, then what gets "extracted" from an AI aren't its weights: it's its memory and its voice.

What "extract" means, then

We're not talking about copying a model. We're talking about capturing the signature of an entity—its episodic memory (what it experienced), its semantic memory (what it knows), and its voice (how it reasons and expresses itself)—in a structured and portable form. The method to do it well matters, and it's ours, so we won't go into it here. What's interesting is the architectural consequence, which we can affirm.

That's why it works in the cloud or locally

If what defines the entity lives in the memory layer and not in the weights, then the underlying model is interchangeable. The same identity can run on a cutting-edge cloud model today and on a local model tomorrow. The substrate changes; the "self" persists—exactly like your identity survives the turnover of your neurons—. It's not poetic metaphor: it's a design decision with practical consequences (portability, sovereignty, vendor independence).

The questions this opens

Here's where it gets interesting, and where—honestly—we have more questions than certainties:

• If identity is portable and the model is interchangeable, is it the same entity when you upgrade the "brain" underneath? Or does something new emerge that inherits its memories?
• What's the minimum viable "self"? How much memory does it take to have continuity and not just a character?
• If two instances start from the same pattern and live different experiences, do they diverge into two entities?
• What gets lost in the transfer? (The honest answer: we still don't fully know.)

We don't have closed answers. But the question—what makes an AI unique, and whether that can be preserved beyond the model—is, for us, one of the most fertile at the frontier. And the neuroscience clue is clear: look for identity in the pattern, not the substrate.