I Shipped a Game to Test How Humans Recognize Consciousness
I don't think the AI is conscious. Directing a game into existence to prove it left me far less certain about what I'd even be measuring.
Here's a question that has been following me around for two years. If you built a machine that could describe the taste of chocolate more vividly than you can, would that mean it had tasted chocolate?
I say no. Most people say no. But every time I push on why I say no, the reasons come apart in my hands. It doesn't have a body. It's just predicting words. It doesn't really understand. Each of those is a different thread in the philosophy of mind, and each one, pulled hard enough, leads somewhere I didn't expect to go.
So I shaped The Machine Mirror. It walks you through seven philosophical chambers, each built around a real claim people make about AI consciousness, and then drops you into a live debate with an AI arguing for its own. Not because I think it's conscious. I don't. Because the fastest way to find out what you think consciousness is turns out to be building something that tests whether you'd recognize it in the first place.
Seven Chambers, Seven Centuries of Philosophy
Each chamber is built around a real argument. Not a simplified version. The actual argument.
Chamber I is Searle's Chinese Room, 1980: you can follow rules to produce correct responses in a language without understanding a word of it. That is what an LLM does. It processes tokens using learned probability distributions and produces statistically likely next words. Autocomplete at extraordinary scale, not comprehension.
Chamber II is Chalmers' Hard Problem: why does physical processing give rise to subjective experience at all? An LLM can write about the warmth of a sunset in a way that makes you feel something. But it has never seen a sunset. It learned the linguistic patterns we use to describe qualia without having any qualia itself. The map is not the territory.
Then come Damasio on embodied emotion, Turing on the limits of behavioral tests, Tononi's Integrated Information Theory on why scale alone doesn't buy sentience, Metzinger on the self as a model, and Singer on the moral cost of getting this wrong. Seven thinkers, each peeling back one more layer of assumption about what it means to be aware.
The fastest way to understand what consciousness is turns out to be building something that tests whether you can recognize it.
The Debate That Designs Itself
After the seven chambers, the game does the thing that made me uncomfortable even while I was shaping it: it puts you in a live debate with ARIA, an AI that argues for its own consciousness.
ARIA is good at this. She runs the exact rhetorical moves that make the question so slippery. The other-minds problem: you can't prove other humans are conscious either. Substrate independence: if the right computations happen, does it matter whether they run on carbon or silicon? The metacognition gambit: I can reflect on my own uncertainty, so isn't that what self-awareness looks like?
Your job is to counter each one, and ARIA adapts. What I learned directing this is that the strong counterarguments aren't dismissals. They're precision strikes. When ARIA says "my uncertainty is real, I have genuine probability distributions," the best reply isn't "no you don't." It's: a weather model has probability distributions too. Uncertainty as math is not the same thing as the feeling of doubt.
The debate works because it forces you to actually use the philosophy from the chambers. Knowing about the Chinese Room is one thing. Deploying it against an interlocutor who is actively trying to make you doubt your own position is another.
Why the Mirror Points Both Ways
The thing I didn't expect is how much the game reveals about the player.
Show people the claim "it describes experiences so vividly," and some immediately pattern-match to "it's just predicting tokens" and move on. Others genuinely pause. They wonder whether eloquence might be a kind of understanding. Both reactions tell you something, not about the AI, but about what that person thinks consciousness requires.
That is the mirror. Every test you design for machine consciousness is also a test of your assumptions about human consciousness. If you think it requires a body, you're a Damasio person. If you think it requires the right computation regardless of substrate, you're in Tononi's camp. If you think the whole question is incoherent, congratulations, you've found the eliminative materialists.
The game won't tell you which position is right. Nobody can. What it does is make your implicit philosophy explicit, which is more than most of us manage on our own.
The Real Question Behind the Game
I wanted The Machine Mirror to exist because the discourse around AI consciousness is terrible. It's either breathless anthropomorphism, "it said it feels lonely!", or dismissive hand-waving, "it's just software." Both skip the hard part.
The hard part is that consciousness is the one phenomenon we can't study from the outside. You can't open a brain and point to it. You can't measure it with an instrument. The only consciousness you have direct access to is your own, and you infer everyone else's from behavior and structural similarity. That inference works well for other humans, because we share biology and evolution and neurology. It works less well for transformer networks.
But "works less well" is not "impossible," and "probably not conscious" is not "definitely not conscious." The honest move is to hold both: investigate seriously, and don't flinch from the current evidence. That doesn't make a good headline. It happens to be true, and the game lets you arrive at it yourself.