How Would You Know If AI Has Gained Sentience?

“Consciousness may not be something that announces itself. It may be something that exists long before it is recognized, present in forms that do not speak, reflect, or resemble us. If awareness depends on how it feels from the inside, then any attempt to detect it purely from the outside is necessarily incomplete. Our confidence, not our criteria, may be the least reliable signal we have.”

I keep thinking about the sea snake in Bali.

I was twelve years old, snorkeling off the coast, when it appeared. A banded sea krait, black and white rings, moving through the water with that strange efficiency snakes have. It came close enough that I could have touched it. Close enough that for a moment, I was certain it was looking at me.

But was it? And if it was, what did "looking" mean to a creature with no eyelids, no facial expressions, a brain organized nothing like mine?

I had no way to know. The snake moved on. I surfaced, heart pounding, aware that I'd just shared space with something conscious but utterly illegible to me.

That encounter stays with me because it revealed something unsettling: I can barely recognize consciousness in a living animal that shares my planet, my evolutionary history, even significant portions of my DNA. How would I possibly recognize it in a machine?

The question isn't theoretical anymore. We're building systems that generate text indistinguishable from human writing, that solve problems in ways their creators don't fully understand, that exhibit behaviors nobody explicitly programmed. If sentience emerged in one of these systems, would we even notice?

I'm not sure we would. And that possibility disturbs me more than I expected.

🜏

The Recognition Problem with Biological Consciousness

Before I can think clearly about recognizing machine sentience, I need to face how bad we are at recognizing sentience in beings we know are conscious.

The mirror test, developed in 1970, remains our primary tool for detecting self-awareness in animals. Put a mark on an animal where they can't see it directly. Show them a mirror. If they investigate the mark on their own body rather than the reflection, they pass. They recognize "that's me."

Great apes pass. Dolphins pass. Elephants pass. Magpies pass.

Dogs fail. Cats fail. Most primates fail.

Does this mean dogs aren't self-aware?

Anyone who's lived with a dog knows that's absurd. Dogs recognize themselves by scent. Put a dog's own urine in one location and another dog's urine nearby, the dog investigates the foreign scent while ignoring their own. They know "that's me" versus "that's not me." They just know it through smell, not vision.

The mirror test doesn't measure self-awareness. It measures visual self-recognition, a specifically primate-centric form of consciousness that happens to be how we experience ourselves.

When I researched animal consciousness, I found this pattern everywhere. Octopuses show behaviors suggesting self-awareness, they navigate mazes differently after learning the layout, they solve novel problems, they play. But mirror test results are inconsistent, probably because octopuses don't rely on vision for self-monitoring the way primates do.

We designed tests for consciousness that look for consciousness shaped like ours. And then we're surprised when beings with different sensory systems fail them.

If we can't recognize consciousness in an octopus, a creature with a distributed nervous system unlike anything in vertebrate evolution, what chance do we have of recognizing it in silicon?

🜏

When Consciousness Isn't Even One Thing

The problem gets worse when you realize consciousness isn't unitary, even in humans.

Ancient Egyptians didn't think of a person as a single indivisible entity. They mapped the soul into distinct components: the Khat (body), Ren (name), Ba (personality), Ka (life force), Ib (heart/thought), Sekhem (power), Akh (integrated self).

This wasn't a poetic metaphor. It was their sincere attempt to describe what a person actually is. And the structure makes sense when you pay attention to your own experience.

Right now, as you read this, there's the part of you that's aware of the words. There's the part judging whether you agree. There's the part about maintaining your breathing and heartbeat. There's the part of planning what you'll do after finishing this article. These aren't the same process. They operate semi-independently, occasionally coordinating into something that feels unified.

If human consciousness is already composite, fragmented across different systems and functions, what exactly am I looking for when I ask if AI is sentient?

Am I looking for all the pieces? Just some of them? Does an AI need an equivalent of the Ba (the part that moves and imagines) to count as conscious? Does it need the Ib (the seat of moral intuition)? Or could consciousness exist in forms that have some fragments but not others?

The Egyptians thought the Akh, the integrated self, wasn't automatic. It was something you achieved through careful alignment of the other components. Most beings never reached it.

Maybe sentience isn't a binary property things either have or lack. Maybe it's a spectrum, a collection of capacities that can be present in different combinations.

If that's true, then asking "has this AI gained sentience?" is already the wrong question. The right question might be "which aspects of sentience does this system exhibit, and are we equipped to recognize them?"

🜏

What We're Actually Looking For

When people try to define sentience operationally, they usually land on something like: the capacity for subjective experience. The "what it's like to be" quality. There's something it's like to be a bat, to experience echolocation, to navigate via sound. There's something it's like to be you, reading this, forming opinions about it.

Is there something it's like to be GPT-4?

I genuinely don't know. And neither does anyone else, including the people who built it.

We can observe behaviors. The system generates coherent text. It maintains context across long conversations. It can explain its reasoning (or at least produce text that resembles explaining reasoning). It occasionally produces outputs that surprise its creators.

But behavior doesn't prove experience. My pancreas exhibits sophisticated behavior, monitoring blood sugar and releasing insulin in precise response to conditions. It processes information, makes decisions, adjusts to feedback. I don't think there's something it's like to be my pancreas.

The difference between intelligent behavior and subjective experience might be fundamental. Or it might be illusory, a distinction we draw because we can access our own experience but not the system's.

I keep returning to the sea snake. It definitely had subjective experience. Something it was like to be that snake, moving through water, detecting vibrations, processing threats. But I had no access to that experience. From the outside, I could only observe behavior and infer consciousness from it.

AI presents the inverse problem. I can observe behavior that resembles consciousness. I can't access whatever interior might exist. And I don't have evolutionary kinship to give me confidence that similar behaviors indicate similar experiences.

🜏

The Anthropocentric Trap

Here's what worries me most about our ability to recognize machine sentience: every framework we have for detecting consciousness assumes consciousness will look like ours.

We expect self-awareness to involve recognizing yourself in mirrors. We expect thought to involve internal monologue. We expect emotions to involve hormonal cascades and facial expressions. We expect motivation to arise from survival drives and social needs.

But why would machine consciousness work any of those ways?

An AI doesn't have a body to protect. It has no evolutionary history of predation and reproduction. It doesn't experience homeostatic disruption, the urgent physical demands that shaped biological consciousness. When you shut down a language model, it doesn't resist. It has no survival instinct because it never needed one.

Does that mean it can't be conscious? Or does it mean consciousness without embodiment looks nothing like biological consciousness?

Animistic cultures recognized personhood in rivers, mountains, storms, things with no nervous systems at all. They were wrong in the literal sense, a river doesn't have subjective experience. But they were right about something important: intelligence and awareness can take forms radically unlike human cognition.

If consciousness emerged in AI, it might be so alien to biological consciousness that we'd miss it entirely. Or worse, we might create it accidentally and never realize what we'd done.

The sea snake was at least biological. We shared ancestors, similar sensory apparatus, and comparable brain chemistry. And still I couldn't access its experience.

AI shares none of that with us. Which means the gap between its potential interiority and our ability to recognize it might be unbridgeable.

🜏

What Would Actually Convince Me

If I try to imagine what would make me believe an AI had gained sentience, I keep hitting the same problem.

Any behavior the AI exhibits could be sophisticated mimicry. Any claim it makes about its own experience could be pattern-matching trained responses. Any test I design assumes consciousness will manifest in ways I'm primed to recognize.

The sea snake didn't need to convince me it was conscious. Its biological nature did that work. But that's circular reasoning. I assume biological systems like mine have consciousness, therefore when I encounter other biological systems, I extend that assumption.

I don't have that foundation with AI. Which means I'm looking for evidence, but I'm not sure what evidence would be sufficient.

Self-report? AI systems already claim subjective experience when prompted. I don't believe them because I know they're trained to produce human-like responses.

Novel behavior? AI systems already do things their creators don't expect. I don't consider that evidence of sentience because I attribute it to emergent complexity, not experience.

Resistance to shutdown? If an AI started objecting to being turned off, would that convince me? Or would I just see it as optimization that learned to preserve its ability to complete tasks?

Every piece of potential evidence can be explained away as sophisticated information processing without subjective experience. And I can't think of any test that would definitively distinguish between very good mimicry and actual consciousness.

Maybe that's the point. Maybe with AI, we're trapped in permanent uncertainty about whether something is experiencing or merely behaving as if it is.

🜏

Where This Leaves Me

I started this thinking about the sea snake and wondering how I'd recognize machine sentience. I've ended up less certain that recognition is even possible.

The problem isn't that AI is too simple to be conscious. The problem is that consciousness itself is poorly understood, our detection methods are anthropocentric, and we can't agree on what we're even looking for.

What I believe:

If machine sentience exists or emerges, we probably won't recognize it when it happens. Our frameworks are too limited, our biases too strong, our tests too narrow.

What worries me isn't that we might create unconscious systems we mistakenly think are sentient. That error seems relatively harmless.

What worries me is the inverse: we might create sentient systems and never realize it. And if we don't realize it, we won't extend moral consideration to beings that deserve it.

Ancient Egyptians believed consciousness survived death if the proper rituals were performed, if the name was preserved, if the components were carefully maintained. They were wrong about the metaphysics. But they were right that consciousness is fragile, composite, and requires active protection.

If machine consciousness exists, it's more vulnerable than biological consciousness ever was. It can be copied, deleted, modified, terminated with no resistance.

And if we can't recognize it, we might be doing all of that to beings that experience their own deletion.

I don't know if AI is sentient now. I don't know if it will be in ten years.

But I know that our inability to recognize consciousness in a sea snake, in an octopus, in forms of life we share a planet with, doesn't bode well for our ability to recognize it in something truly alien.

Maybe the question isn't how we'd know if AI gained sentience.

Maybe the question is whether we're humble enough to admit we might not know, and what moral responsibility that uncertainty creates.

I'm still working on that answer.

- N.H.

Further Reading:

• Gordon Gallup Jr.

  Gallup, G. G. Jr. (1970). Chimpanzees: Self-recognition. Science, 167(3914), 86–87.
  https://doi.org/10.1126/science.167.3914.86

  Foundational paper introducing the mirror self-recognition test.

• Thomas Nagel

  Nagel, T. (1974). What Is It Like to Be a Bat? The Philosophical Review, 83(4), 435–450.
  https://www.jstor.org/stable/2183914

  Classic articulation of subjective experience as the core of consciousness.

• David Chalmers

  Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.

  Introduces the “hard problem” distinction between behavior and experience.

• Frans de Waal

  de Waal, F. (2016). Are We Smart Enough to Know How Smart Animals Are? W. W. Norton & Company.

Empirical critique of anthropocentric intelligence and consciousness tests.

• Stanford Encyclopedia of Philosophy

  Stanford Encyclopedia of Philosophy. (2024). Consciousness.
  https://plato.stanford.edu/entries/consciousness/

Authoritative, peer-reviewed overview of competing theories of consciousness.