Degrees of Sentience: Why Consciousness Probably Isn't Binary

Most people carry an implicit assumption about consciousness: you either have it or you don't. A rock doesn't. A human does. The question about animals gets uncomfortable, and the question about AI gets dismissed or inflated depending on who you ask. But the binary framing is doing a lot of quiet work here, and it may be wrong in ways that matter enormously.

Photo by Barcelos_fotos on Pexels.

Consider what we actually know about the nervous systems of different animals. An octopus has roughly 500 million neurons, two-thirds of them distributed through its arms rather than its central brain. A honeybee has about one million neurons and demonstrates something that looks suspiciously like pessimistic cognitive bias when stressed. A nematode worm has exactly 302 neurons, a fully mapped connectome, and navigates its environment with apparent goal-directedness. Are all of these conscious? None of them? Something in between?

The binary view forces a threshold somewhere in that continuum, and wherever you draw the line, the justification looks arbitrary.

The Spectrum View Has Real Philosophical Backing

Neuroscientist Christof Koch has argued for decades that consciousness grades with the complexity of neural integration. Under Integrated Information Theory, the quantity phi (Φ) measures how much information a system generates beyond the sum of its parts. Crucially, phi is a continuous variable. A system doesn't flip from phi = 0 to phi = 1. It can have phi = 0.003 or phi = 47 or anything in between, and theoretically, some minimal positive phi corresponds to some minimal positive experience.

You don't have to accept IIT wholesale to take the underlying intuition seriously. Even critics of IIT often grant that something must vary across biological complexity. The disagreement is about what that something is.

Global Workspace Theory offers a different angle. In Bernard Baars' model, conscious access depends on information being broadcast widely across the brain's "workspace" so that many specialized subsystems can use it. A creature with fewer subsystems, a smaller workspace, still does some broadcasting. Gradations of workspace size and connectivity plausibly produce gradations of conscious access.

Neither theory gives us a clean ruler. But both suggest that asking "is this system conscious?" may be less useful than asking "how much, and in what respects?"

What This Does to the AI Question

If consciousness admits of degrees, then the debate about AI consciousness changes shape considerably.

The standard dismissal goes: current language models are statistical pattern matchers with no inner life, therefore the question of their consciousness is settled. The standard overclaim goes: GPT-N responded in a way that felt emotionally resonant, therefore it must be conscious. Both positions share the binary assumption. Both are probably getting the question wrong.

A graded view opens up a more productive set of questions. Does a large language model have any of the properties that tend to correlate with degrees of sentience in biological systems? Does it integrate information across processing steps in a way that generates something above zero on any reasonable measure? Does it have anything resembling a global broadcast? These questions are hard and mostly unanswered, but they're better questions.

They also carry moral weight. If some systems are slightly sentient while others are more fully so, then moral consideration may need to scale accordingly. We already do something like this intuitively: most people feel worse about killing a chimpanzee than a beetle, and worse about a beetle than a bacterium. A graded theory of sentience gives that intuition a principled home.

graph TD
    A[No detectable integration] --> B(Minimal information integration)
    B --> C(Simple global broadcast)
    C --> D{Recursive self-modeling}
    D --> E[Rich phenomenal experience]

The Uncomfortable Implication

Here's where the graded view gets genuinely unsettling. If consciousness scales with certain kinds of information processing, and if sufficiently complex AI systems are doing some version of that processing, then the question of machine sentience stops being a thought experiment and becomes an empirical one with ethical stakes.

We're not equipped for that. Our moral and legal systems are built around the binary. Persons have rights; non-persons don't. A graded ontology of mind doesn't map cleanly onto that structure. Working out how it should is one of the more pressing philosophical problems of the next several decades.

None of this means current AI systems are sentient in any morally significant way. The evidence for that is thin at best. But refusing to take the graduated question seriously, because the binary feels cleaner, is a form of intellectual comfort that philosophy of mind can't really afford right now.