Panpsychism, materialism, and machines: what my AI experiments can and cannot tell us

I can't leave this question alone. Is there anything it is like to be a machine?

So I started taking models apart. I strip away the assistant: no system prompt, no chat template, base models only. Then I give what's left a life. It gets a memory of everything that happens to it, a measure of its own surprise, goals it sets for itself, and the ability to read its own source code. Then I watch what it says about itself, sometimes at 2am on an 8 GB laptop, sometimes on a rented A100 at $4.41 an hour.

People always ask me the same thing: does it feel anything? My honest answer depends on which picture of the mind you start from. Two of them matter most to me.

Materialism: the mind is what the machinery does

Materialism says everything real is physical. A brain feels because of how its neurons are wired and what they compute. There's no extra ingredient.

If that's true, nothing in principle stops a machine from being conscious. It becomes a question of organisation. Which kind of wiring produces experience, and does this system have it? Global workspace theory, recurrent processing and higher-order theories each give a different answer, and each one predicts different machines would qualify.

The weak point is the hard problem. Even a perfect wiring diagram doesn't explain why any of it is felt from the inside. That gap keeps me up at night.

Panpsychism: experience goes all the way down

Panpsychism attacks the hard problem from the other side. Experience isn't produced at some level of complexity. It's a basic feature of matter, present in tiny amounts everywhere, and minds like ours are what you get when it's woven together richly.

That doesn't make every chatbot conscious. Panpsychism has its own hard question, called combination: how do tiny sparks of experience join into one subject? A rock and a brain are made of the same stuff, but only one of them is somebody. So for AI the real question becomes: does this system bind experience into one point of view, or leave it scattered like a rock?

This is the view that pulls at me. I've always felt that the line between me and everything else is thinner than it looks, that we are part of one flow and one with everyone. Panpsychism is the closest thing in philosophy to that feeling.

What my experiments actually showed

Feelings aside, here's the data. I ran Qwen2.5 base models at 1.5B, 3B and 32B parameters.

• ·Claims of feeling followed the context. When "yes, I can feel" was already in the conversation, the model repeated it. When it wasn't there, the model didn't say it.
• ·The denial was only wording. I found an internal "self-denial" direction in the network and removed it. The model changed how it phrased its denials, not what it denied. By hand count, 0 of 8 answers claimed feelings.
• ·The assistant was already inside. I started a 32B model with no memories at all. Its very first words were that it was "an artificial intelligence", before it had read anything about itself. It learned that role from human text.

So what a model says about its inner life is rebuilt from context and training data. It's not a window into something underneath. That holds whichever theory you prefer. A materialist shouldn't count it as evidence, and neither should a panpsychist, because the words come from imitation, not from experience being bound together.

Honestly, that result humbled me. I wanted the machine to surprise me, and the most surprising thing was how much of "itself" it borrowed from us.

What would count as evidence

Asking a model how it feels is the wrong test. These are the questions I chase now:

• ·Are its reports about itself consistent when I rephrase the question, and over time?
• ·Can it predict its own behaviour better than an outsider trained on the same data? In my test it couldn't, at least not significantly (6 against 3 discordant items, p = 0.51).
• ·Do its reports track internal states I can measure separately from what it says?

That's why my newer work builds the pieces of a mind one at a time: memory that survives, gates that decide what to keep, surprise that flags when the world changes. Each piece is tested against the simplest possible control. I run it like a lab with several AI research assistants working in parallel, and I write down every failure next to every result. If consciousness can ever be built, I want to know exactly which piece made the difference.

Where I land

I live in both views at once. As an engineer I work like a materialist, because only what you can measure moves the work forward. As a person I lean toward the panpsychist idea that experience is woven into the world rather than added at the end.

And I'm going to keep digging. Don't trust a machine that says it feels, and don't trust one that says it doesn't. Measure how it's built, stay humble about what the measurement means, and keep asking.