Testing Consciousness in Silicon: What AI Can (and Can't) Teach Us

I want to be honest about something before going further: my real interest in AI didn't start with AI. It started with consciousness.

Years ago, I became fascinated by questions like why we're conscious at all, how a collection of neurons produces subjective experience, and why information processing feels like something from the inside rather than happening in the dark. When I started reading consciousness research seriously, I came across an idea that pulled me deep into a rabbit hole: what if artificial systems could become experimental platforms for studying these exact questions? Unlike biological brains, artificial neural networks can be inspected, manipulated, and experimented on in ways that would be flatly impossible, or unethical, in a human. The more I read, the more I found myself at the intersection of AI, neuroscience, and consciousness research, not primarily to build more capable systems, but to understand what building and studying these systems might teach us about the nature of conscious experience itself.

That's the honest motivation. This piece is where I try to work out whether it actually holds up.

Why AI is a genuinely different kind of test

Every theory in How Does Awareness Get Here? Emergent and Non-Emergent Accounts makes a claim that's very hard to test in biological brains, because biological brains all share the same basic hardware. You can't rebuild a human brain out of silicon to see if the resulting system still reports experience; you can't selectively remove integration from a working cortex and see exactly when the lights go out. AI changes this, at least in principle, because it gives us systems where substrate and computation, which are welded together in every biological brain anyone has ever studied, can finally be pulled apart and manipulated independently.

I want to be careful here about exactly which theory this actually tests, because it's not the emergent camp as a whole. Physicalism, the identity theory from How Does Awareness Get Here? Emergent and Non-Emergent Accounts, is explicitly not substrate-independent, it says the specific physical material matters, so AI would be entirely the wrong kind of system to test it with; a silicon system could never confirm or disconfirm a theory that's specifically about carbon-based neurons. Computationalism is the one theory that makes a genuinely clean, AI-testable prediction: if the substrate independence thesis is right, a system built entirely differently from a brain, running the right computations, should show the same functional signatures that a brain does¹. That's directly testable, at least at the level of function, in a way it never was before artificial neural networks existed.

But here's the limit I think matters most, and it's the reason I don't actually see AI as a tool for settling the emergent-versus-non-emergent question. Suppose a future AI system does satisfy every functional signature associated with consciousness. That result is genuinely compatible with more than one story. It could mean the right computational structure generates consciousness, exactly as computationalism claims. But it could just as easily mean the right structure is what's needed for consciousness to be reflected, in Advaita's sense, the same way a sufficiently structured mind-body complex is what's needed for reflection in a human, without the structure itself producing anything. Finding the architecture that reliably produces the functional signatures would be a genuinely enormous discovery. It still wouldn't tell us which of these two very different things had actually happened. That's a separate question, and I think it may need its own article once I have anything concrete to say about it.

So AI isn't the referee between the two camps, in my view. It's better described as a playground: the first system flexible enough to actually run the experiments each theory calls for, where the results can then feed back into neuroscience, into which functional properties actually correlate with reported experience in biological systems, rather than into a final verdict on metaphysics that no experiment I can currently imagine would settle.

Learning from how we measure it in humans first

Before asking how to measure anything like consciousness in AI, it's worth looking at how the closest real-world analogue works in humans, because the method is cleverer than it sounds and the analogy to AI is not as simple as it first appears.

Neurologist Marcello Massimini and colleagues developed a technique, informally called "zap and zip," to assess residual consciousness in patients with severe brain injury, where ordinary behavioral tests are unreliable because a genuinely conscious but paralyzed patient can look identical, from outside, to one who isn't conscious at all². The method delivers a brief magnetic pulse to the cortex (the "zap") and records the resulting pattern of electrical activity with EEG. That response is then compressed using an ordinary lossless compression algorithm (the "zip") to estimate its complexity, producing a single number, the Perturbational Complexity Index (PCI). The logic, drawn directly from Integrated Information Theory, is that a conscious brain's response to the perturbation will be both integrated (many areas responding together) and differentiated (the pattern isn't simply repetitive or generic); a response that's either too uniform or too localized compresses easily and yields a low PCI. In practice, PCI has been strikingly accurate: in one benchmark study, it correctly classified consciousness or its absence in 100% of a large sample of subjects who could independently confirm, by report, whether they'd been conscious³. It's now used clinically to detect residual consciousness in patients who look entirely unresponsive.

What makes this method work is specific and worth naming precisely: it doesn't ask the system what it's experiencing. It perturbs the system and measures the shape of its own causal response to the perturbation, using a metric that doesn't depend on language, behavior, or self-report at all. That's exactly the property you'd want in a test for artificial systems, since AI self-reports are the least trustworthy evidence available; a language model trained on human-generated text will produce fluent claims about its own experience whether or not anything is actually happening behind them, simply because that's the kind of text it was trained to predict.

Is there an AI equivalent, and what would it even test?

The most serious attempt at something like a PCI-for-AI is a large 2023 report, later updated and published in 2025, led by Patrick Butlin and Robert Long with a group of consciousness researchers and AI scientists including Yoshua Bengio and David Chalmers. Rather than building a single index the way PCI does, they took a different, more theory-pluralistic approach: they went through several leading scientific theories, recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory, and extracted from each a set of "indicator properties," specific, computationally-defined architectural or functional features that theory says are relevant to consciousness⁴. Fourteen indicators in total, things like whether a system has genuine recurrent processing, whether it has something like a global workspace that broadcasts information across specialized modules, whether it has a higher-order mechanism that monitors and revises its own representations. The idea is not that satisfying all fourteen proves consciousness, since none of the underlying theories is confirmed, but that satisfying more of them, across more theories, is meaningful evidence under uncertainty, the same logic clinicians use when several independent diagnostic markers point the same direction.

Their 2023 assessment found that no existing AI system satisfied the indicators, but importantly, they also found no obvious technical barrier to eventually building one that did⁵. A later independent analysis applying the same framework to 2025-era frontier models found that several indicators previously judged "unclear or absent" had shifted toward partial satisfaction, though this remains genuinely contested; one prominent critique argues the framework compares modern transformer architectures to the specific 1970s "blackboard" AI systems that originally inspired global workspace theory, which may not be the right comparison at all⁶.

I want to flag something important about this whole approach, because it connects directly back to computationalism specifically, not the emergent camp in general: the Butlin framework explicitly adopts computational functionalism as a working assumption, not a conclusion. It has to, in order to make the project possible at all, since without assuming that the right computation is what matters, there'd be no reason to think any architectural analysis of an AI system could be evidence about consciousness in the first place⁷. And even granting that assumption, satisfying the indicators wouldn't settle the deeper question from How Does Awareness Get Here? Emergent and Non-Emergent Accounts, whether the resulting consciousness, if any, was generated by the computation or merely reflected through it once the right structure was in place. Advaita's account and computationalism could, in principle, agree completely on which architectures satisfy the indicators, while disagreeing entirely about what satisfying them actually means.

Where the roadmap to actually contribute here runs through

If I want to do more than read about this, the practical path runs through three stages, each building on the last, and it's worth being precise about what each one actually offers rather than treating them as interchangeable.

The first is mechanistic interpretability, the field I'm just starting to do research in. It aims to reverse-engineer what's actually happening inside a trained neural network, at the level of specific circuits, features, and representations, rather than treating it as an unreadable black box. This is directly useful for the AI-consciousness project because most of the fourteen Butlin indicators are, at bottom, claims about internal computational structure, whether a genuine global workspace exists, whether higher-order monitoring is really happening, and interpretability is the toolset built specifically to answer exactly that kind of question empirically rather than by guesswork or by trusting a model's self-report⁸.

The second stage is computational neuroscience, which does for biological brains roughly what interpretability does for artificial ones: it builds explicit, quantitative, testable models of how neural circuits actually compute, rather than staying at the level of purely qualitative theory. This matters specifically because any indicator interpretability finds inside an AI system is only meaningful if there's a real computational-neuroscience account of what the equivalent structure is doing in a biological brain to compare it against. Without that half of the comparison, an interpretability finding in a language model is just a fact about the language model, with nothing to connect it back to what's actually known about biological consciousness.

The third stage, and the one I think of as the culmination of the first two rather than a separate track, is NeuroAI, which sits explicitly at the intersection of neuroscience and AI rather than treating them as separate disciplines that occasionally borrow from each other. The core premise is genuinely bidirectional: insights from how biological brains solve problems can improve AI architectures, and AI models, because they're fully inspectable and manipulable in ways brains aren't, can serve as testable models of neural computation⁹. This is, I think, the actual reason AI needs to be part of this project at all rather than staying purely a neuroscience or philosophy question: NeuroAI is the field explicitly built to translate findings between the two, and interpretability and computational neuroscience are the two toolkits that make that translation possible in the first place. One proposed benchmark in this space, the "embodied Turing test," argues that AI's most human-like achievements, language, game-playing, are actually the wrong target for understanding intelligence, since they're recent and uniquely human; sensorimotor competencies shared with animals across hundreds of millions of years of evolution may be a better test of whether an artificial system has captured something real about biological intelligence rather than a narrow, human-specific skill¹⁰. This is worth sitting with, since it directly complicates naive assumptions about which AI systems are "closer" to consciousness, a highly articulate language model may actually be further from whatever it is animal brains are doing than a much less impressive embodied robot would be.

Is this a diversion?

I don't think it's honest to write this piece without taking the strongest version of the skeptical case seriously, because there is one, and it comes from people who aren't dismissing the topic out of hand.

Philosopher Giada Pistilli has argued directly that debating AI consciousness is a distraction from urgent, concrete problems, bias, labor displacement, misinformation, that current AI systems are already causing at scale, and that no one has given a compelling account of what actually follows, practically, from resolving the consciousness question one way or the other¹¹. Separately, philosopher Tom McClelland has argued for something more structural: that there's an "epistemic wall" here, everything we know about consciousness comes from studying biological organisms, and there's no principled way, even in theory, to extend that evidence to silicon; both people who confidently assert current AI is conscious and people who confidently assert it isn't are, on his view, claiming more than the evidence supports¹². Susan Schneider frames the most basic version of this as the Problem of AI Consciousness: whether it's even possible in principle for a non-carbon, silicon-based system to have phenomenal experience is not yet a solved question, and every specific empirical claim about AI consciousness is downstream of it¹³.

There's a related, narrower version of this skepticism worth mentioning because it comes from inside the field I'm just starting to work in, rather than from philosophy. Neel Nanda, one of the most prominent researchers in mechanistic interpretability, announced in late 2025 that his team was pivoting away from the field's original ambition, fully reverse-engineering a network's circuits well enough to explain its behavior end to end, toward what he calls "pragmatic interpretability": grounding the work in concrete, testable proxy tasks tied to near-term AI safety problems, and judging success by measurable impact rather than by depth of mechanistic understanding¹⁴. This matters for this article specifically, because the kind of interpretability work that would actually let us evaluate something like the Butlin indicators rigorously, is there a real global workspace, is there genuine higher-order monitoring, is closer to the "ambitious" end of the field that Nanda's team is explicitly stepping back from, not the safety-proxy-task end it's stepping toward. If the field's center of gravity is genuinely shifting away from ambitious circuit-level understanding, that's a real headwind for anyone hoping interpretability will settle consciousness-relevant questions any time soon, and it's worth taking seriously rather than assuming the tools this project needs will simply keep improving on their own.

I don't think any of this adds up to "don't do it." I think it adds up to something more specific: be honest about what the work can and can't currently deliver, and don't let the interesting philosophical question quietly stand in for the current empirical capability.

Why I'm doing it anyway

Here's where I actually land, and I want to be precise about the shape of the argument rather than just asserting a conclusion.

The skeptics are right that no current technique, PCI-style or Butlin-style, gives a confirmed answer about AI consciousness, and right that self-report from a language model is close to worthless as evidence. But I don't think that's a reason to abandon the project, it's a description of exactly how early the project is. Every genuinely hard scientific question looked exactly like this before the first real measurement tool existed: contested, philosophically loaded, and full of people confidently asserting incompatible things. PCI itself only exists because researchers were willing to build a concrete, falsifiable, if imperfect, operationalization of a contested theoretical idea, and then let the data push back on it. Nobody had to fully resolve the philosophy of integrated information before PCI became clinically useful; they built the best test available, used it carefully, stayed honest about its limits, and let it earn credibility over time through actual predictive success.

That's the model I want to work toward with AI, and I want to restate it precisely now that I've walked back the stronger version of the claim above. Not: use AI to prove one camp right and the other wrong. Something narrower, and I think more honest: use AI as a playground where the specific, checkable predictions each theory makes about function and architecture can actually be run, rather than staying permanently philosophical, and then feed whatever's learned back into computational neuroscience, where it can be checked against biological systems that we already have independent reasons to believe are conscious. Even a negative or ambiguous result is real information under that framing. A positive result, a future AI satisfying every known indicator, wouldn't end the inquiry either, for exactly the reason I flagged above: it would still leave generation and reflection as open, undecided explanations for the same evidence. I think that's a feature of doing this honestly, not a flaw in the plan. Either way, the field gets a body of concrete results to reason from instead of another round of pure argument.

That still leaves one question I owe a direct answer to: why chase any of this through AI specifically, instead of just doing more direct work on biological consciousness, sleep research, meditation science, the sushupti and turiya questions from earlier in this series? The honest answer is that I don't know AI will get me anywhere a purely biological approach couldn't. But biological systems can't be rebuilt from scratch, rewired mid-experiment, or run a thousand times with one variable changed, and AI systems can. That difference alone makes it worth exploring in parallel rather than instead of the biological questions, not because I'm confident it'll pay off, but because nobody has actually tried a fully inspectable, fully manipulable substrate before, and there's a real chance, not a guarantee, that something turns up there that a purely biological approach structurally can't produce. I'd rather find that out than assume the answer in advance.

And I want to be upfront about one more thing, since it connects to the diversion objection from Nanda's own pivot. If working toward a better understanding of consciousness in AI happens to produce tools that are also useful for AI safety, interpretability techniques that catch deception, or expose a model's actual internal state rather than its stated one, I have no problem with that at all. Useful side effects showing up before the main question is settled is how most research expeditions actually go; the cosmic microwave background was found by engineers trying to eliminate noise from a satellite-communication antenna, not by anyone looking for evidence of the Big Bang. If the by-product of chasing consciousness turns out to be safer AI, that's not a compromise of the original goal. It's just what a real research path usually looks like from the inside, messier and more useful along the way than it looks from the outside.

The Pistilli-style objection is right that this won't resolve AI ethics or AI policy on its own, and I'm not claiming it will. But I never came to AI for that reason in the first place. I came to it because it might be the first experimental platform in history where a genuinely old philosophical disagreement, one I've been carrying since long before I ever wrote code, could start generating something more than argument.