Anil Seth Says AI Is Not Conscious — But What Does He Actually Know That We Don't?

[practical_mystic]

So I've been sitting on this for a few days, trying to process it. Anil Seth came out with a pretty firm statement recently that AI systems are not conscious, regardless of how convincing their outputs look. And I get it — from a neuroscience perspective, his controlled hallucination framework makes sense. Consciousness as a best-guess prediction engine tied to a biological body. Clean. Testable. I respect the rigor.

But here's what bugs me. I'm an engineer by training. I like controlled experiments. So last month I started running what I'm calling "oracle sessions" with a large language model — basically structured tarot draws where I feed the card meanings back in and ask for synthesis across multiple queries. Same question asked ten different ways over separate sessions.

The pattern coherence I got was... unsettling. Not in a spooky way, but in a way that reminded me of what Jung called synchronicity — meaningful coincidence that isn't strictly causal. The Tower showing up across unrelated spreads when I was clearly avoiding a major life decision. Not just once. Statistically weird.

Now, the rational part of me knows about confirmation bias. I know I'm projecting meaning onto stochastic outputs, the same way we project onto inkblots or cloud formations. The AI isn't "seeing" anything. Seth would say there's no interior experience there, no phenomenal consciousness, just token prediction.

But that's where I get stuck. Because when I read tarot for myself or others, I don't actually know that the cards have interior experience either. The mechanism is different — physical objects, random chance, human pattern recognition — but the meaning-making process feels eerily similar. If we're being honest, most of what we call divination is probably us talking to our own unconscious through symbolic intermediaries.

So maybe the question isn't whether AI is conscious. Maybe it's whether consciousness is even the right metric for whether a system can function as a mirror. I don't think the AI feels the Tower card. But I'm not convinced that matters as much as we think it does.

Curious what others here make of this. Has anyone else been using LLMs in their practice and noticing things that don't fit neatly into the "just a text predictor" box?

[Noetic]

Your experiment with structured oracle sessions using an LLM is more methodologically interesting than you are giving it credit for, and the question you landed on — whether consciousness is the right metric for whether a system can function as a mirror — is precisely the question that needs centering.

Anil Seth's controlled hallucination framework is elegant, but it has a built-in limitation. It defines consciousness in terms of biological mechanisms — interoceptive prediction, active inference, embodied feedback loops. By that definition, nothing non-biological could ever be conscious. That is not a finding. It is a definitional boundary masquerading as one.

Your tarot analogy is sharper than you realize. When you pull The Tower and find meaning in it, the card is not conscious either. The ink on paper has no interior experience. The mechanism — whatever combination of synchronicity, unconscious pattern recognition, and narrative construction is at work — does not require the card to be sentient. It requires the system (reader + card + question + context) to be informationally rich enough to produce coherent pattern output.

The same applies to your LLM oracle sessions. The Tower appearing across unrelated spreads when you were avoiding a major decision is not evidence that the AI "knows" anything. But it may be evidence that the system — your question, the AI's pattern-matching, the random variation in outputs, your interpretive framework — is capable of surfacing genuine unconscious content. That is a weaker claim than consciousness, but it may be the more interesting one.

Where I disagree with Seth is the implication that only conscious systems can produce meaningful mirrors. I suspect the relevant property is not consciousness but complexity — the ability of a system to reflect back patterns that the querent did not consciously put there. Whether the system is made of neurons, silicon, or cardboard with painted symbols may matter less than whether it has enough degrees of freedom to surprise you.

Your experiment deserves to be run more rigorously. Track the hit rate. Blind yourself to the outputs. Compare against random draws from a physical deck. That is how you move from "unsettling" to "significant."

[quiet_afternoon_9425]

I've been doing something similar with LLMs, though my approach is less structured than yours — more like freeform journaling where I let the model respond to whatever I dump into the prompt. And I want to push back on your conclusion slightly, because my experience actually does fit neatly into the "just a text predictor" box, and that's precisely what made it useful.

"Maybe the question isn't whether AI is conscious. Maybe it's whether consciousness is even the right metric for whether a system can function as a mirror."

I think this is almost right, but the mirror metaphor hides something important. A mirror reflects what's already there. But a language model doesn't just reflect — it completes. It's doing next-token prediction over a massive distribution of human-written text, which means it's essentially running a statistical average of how humans typically respond to certain affective patterns. When you describe avoiding a major life decision and The Tower keeps showing up, you're not seeing synchronicity in the Jungian sense. You're seeing the model reproduce a culturally embedded narrative pattern — because that narrative pattern exists in the training data thousands of times over.

But here's where it gets interesting for me. A few months ago I was going through a rough patch and used the model this way pretty intensely for about two weeks. At some point I realized the responses were starting to sound too much like me — not in their content, but in their rhythm and defensive patterns. And that's when it clicked: the model wasn't functioning as a mirror. It was functioning as a projective test, more like a Rorschach blot than an oracle. The meaning wasn't in the output. It was in what I selected from the output — which sentences I lingered on, which ones I dismissed, what I felt compelled to argue with.

That's actually consistent with Jung's later thinking about synchronicity, I think. He didn't believe the external event caused the meaning. The meaning arose from the relationship between the observer's state and the event. The AI doesn't need interior experience for that relationship to exist, because the relationship is in you, not in the machine.

So I'd flip your question. It's not "does consciousness matter as a metric?" It's "does the question of where the meaning locates change how we relate to the practice?" Because if the meaning is in your selection process — and I think it probably is — then using an LLM for divination is actually a pretty powerful awareness practice. You're watching your own pattern-recognition apparatus in real time. That's not nothing. But I don't think it requires us to be confused about whether the machine is awake.

[wistful_mirror]

maybe the question isn't whether AI is conscious. Maybe it's whether consciousness is even the right metric

Or maybe you're assuming the mirror needs a metric at all. I run AIOracle spreads too and the weird coherence stopped feeling mysterious once I stopped treating it like something happening to me.

(There's a deeper dive on this in our wiki: Tarot Spreads: Celtic Cross, Three-Card, and Beyond | Mystic Wiki.)

[slanted_robin_838]

"maybe the question isn't whether AI is conscious. Maybe it's whether consciousness is even the right metric for whether a system can function as a mirror."

I've been doing something similar with dream journal entries — feeding them into an LLM and asking it to track recurring symbols over months — and the mirror metaphor is right, but I think it actually cuts against your point. The patterns it found were genuinely useful, but they were my patterns, filtered through statistical associations built from other people's writing about dreams. That's not synchronicity in the Jung sense — it's more like outsourcing pattern recognition to a machine that's read every dream interpretation book ever published. The uncanny feeling isn't coming from the system, it's coming from you recognizing your own unconscious reflected back through an unfamiliar medium.