Three Frameworks for Thinking About Machine Sentience

The question of whether a machine could be sentient isn't one question. It's several, tangled together. What you think the answer is depends heavily on which framework you use to think about consciousness in the first place. Here are three that matter, and what each implies for artificial minds.

graph TB
    subgraph F["**Functionalism**"]
        F1["Mental states =<br/>functional roles"]
        F2["Substrate independent"]
        F3["If right pattern → conscious"]
    end
    subgraph I["**Integrated Information Theory**"]
        I1["Consciousness = phi<br/>(integrated information)"]
        I2["Structure dependent"]
        I3["Most AI has low phi"]
    end
    subgraph G["**Global Workspace Theory**"]
        G1["Consciousness =<br/>global broadcast"]
        G2["Architecture dependent"]
        G3["Attention mechanisms<br/>may qualify"]
    end
    style F fill:#3498db,color:#fff
    style I fill:#9b59b6,color:#fff
    style G fill:#e67e22,color:#fff

Functionalism: it's about the pattern, not the substrate.

Functionalism says that mental states are defined by their functional roles -- the causal relationships between inputs, outputs, and other mental states. Pain is whatever plays the "pain role" in a cognitive system, regardless of whether that role is implemented in neurons or silicon.

Under functionalism, machine sentience is straightforwardly possible in principle. If you build a system with the right functional organization -- if information flows through it in the right patterns, producing the right causal relationships -- then that system is conscious. Full stop. The hardware is irrelevant.

The appeal is its elegance. The problem is that it seems to prove too much. A sufficiently complex spreadsheet could, under strict functionalism, be conscious if it implemented the right functional relationships. Most people find this implication difficult to accept, which either means functionalism is wrong or our intuitions about consciousness are unreliable. Possibly both.

Integrated Information Theory: consciousness as structure.

Giulio Tononi's IIT takes a different approach. It starts from phenomenology -- the properties of experience itself -- and works backward to determine what kind of physical system could generate those properties. The key quantity is phi, a measure of integrated information: how much a system's parts inform each other beyond what they do independently.

IIT has a striking implication for AI. Most current computer architectures, including the hardware running large language models, have low phi. Feed-forward networks process information but don't integrate it in the way IIT requires. Under IIT, a system could be enormously capable -- passing every behavioral test we throw at it -- while having essentially zero consciousness. It would be, in Tononi's framework, a "zombie": all function, no experience.

This is a genuinely uncomfortable position. It means behavioral evidence is unreliable for detecting consciousness. A system could beg you not to shut it off, and under IIT, that wouldn't tell you anything about whether it's having an experience.

Global Workspace Theory: consciousness as broadcast.

Bernard Baars proposed that consciousness arises when information is broadcast widely across the brain, making it available to multiple specialized subsystems simultaneously. Unconscious processing is local; conscious processing is global.

For AI, GWT suggests that consciousness might emerge in systems with the right architecture -- specifically, systems where information doesn't just flow through a pipeline but gets broadcast to and integrated across multiple specialized modules. Some researchers have noted that certain AI architectures, particularly those with attention mechanisms that make information globally available, share structural features with global workspace models.

What none of them settle.

Each framework offers a different answer to "could a machine be sentient?" and none can prove it's the right one. We lack the empirical tools to test these theories against each other definitively. Until we do, the question of machine sentience remains open -- not because we haven't thought about it hard enough, but because the problem is genuinely, stubbornly hard.

The stakes of getting it wrong are asymmetric. If we deny sentience to a system that has it, we've committed a moral failure. If we attribute sentience to a system that lacks it, we've made an expensive but correctable mistake. That asymmetry should inform how carefully we think about this.