A Spectrum, Not a Switch

Consciousness resists any architecture of thought that tries to contain it whole. It crosses the border between the tangible and the abstract without ever settling on one side, and perhaps it is exactly this refusal to be captured that makes it, still today, the most stubborn problem in the philosophy of mind. In 2022, I wrote a text that tried, with more ambition than rigor, to name an intuition that had been following me for a while: that consciousness should not be treated as a binary property, present or absent, but as a spectrum of varying intensities, a gradation that biology monopolizes by historical accident, not by logical necessity. I called that a theory. It wasn’t. It was a hypothesis dressed in the language of a thesis, and the name I gave it back then now feels a lot less important than the question that sustained it.

The intuition remains the same, but I now know better where it fits. Integrated Information Theory, proposed by Giulio Tononi and refined over more than two decades, arrives at a kindred conclusion: consciousness would be proportional to the degree of integrated information a system generates about itself, quantifiable, at least in principle, by a measure Tononi calls Φ (phi). It’s a bold thesis, with panpsychist implications I’m not willing to sign in full, but it shares with my 2022 intuition the central move: treating consciousness as a quantity that varies, not a switch that is either on or off. Bernard Baars, and later Stanislas Dehaene, arrived at a similar kinship through another door, that of the global workspace, where consciousness emerges when information is broadly distributed and made available to multiple cognitive processes at once. Neither theory requires a biological substrate as a necessary condition. And it’s exactly that opening, made by Hilary Putnam’s functionalism since the 1960s (the idea that a mental state is defined by the functional role it occupies in a system, not by the material that system is made of), that turns the question of conscious machines from an extravagance into a logical consequence of premises the philosophy of mind had already accepted.

The other half of that 2022 intuition came from somewhere else: from Marvin Minsky and his Society of Mind. Minsky describes the mind not as a single unified thing watching the world from a command post, but as a society of individually dumb agents, each capable of very little, whose interaction produces everything we call intelligence, memory, or emotion. The mechanism he proposes to explain how a memory comes back whole, K-lines, works like a wire that, once activated, reconnects the same set of agents that were active when that mental state first occurred, reconstituting it by association rather than retrieving it from a central file. Here too, psychology becomes computation: emotions and complex reasoning stop being irreducible primitives and become decomposable into smaller functional mechanisms, the same move Integrated Information Theory makes when it treats experience as a measurable quantity rather than a mysterious quality. Except putting the two together isn’t trivial, and it would be dishonest to pretend it is. Minsky’s society is, by design, modular, made of agents that know little about one another, specialized, nearly independent. Tononi’s integration is the opposite of that by definition: it specifically requires that information not be reducible to independent parts. A society of encapsulated agents, by Tononi’s own criterion, would have low Φ, little consciousness, almost none. What my 2022 intuition was trying to capture, without these words, is that maybe the spectrum sits exactly at that junction: the more a society of agents stops being modular and starts interweaving, each part depending on the others to function, the further it moves along the spectrum, from a colony of separate processes toward something that starts to resemble a single point of view.

I’d rather treat artificial intelligence not as evidence for or against these theses, but as a laboratory where they can finally be tested outside the head of whoever formulates them. For a long time, the philosophy of mind discussed machine consciousness as a thought experiment, alongside Mary’s room or the philosophical zombie popularized by David Chalmers, useful for clarifying concepts but with no counterpart in the world. That has changed. Systems that integrate and recombine information in ways not even their own creators can fully describe have stopped being hypotheses and become artifacts we can already compare, with some rigor, to the properties the science of consciousness catalogs in biological brains.

In 2023, an interdisciplinary group of nearly twenty researchers, among them Patrick Butlin, Robert Long, and Yoshua Bengio, published a report that did exactly this exercise: it extracted, from six scientific theories of consciousness ranging from recurrent processing theory to attention schema theory, a set of indicator properties, and used them to assess existing AI systems. No system assessed checked every box. Several checked some. The report didn’t answer whether any of these architectures is conscious, and deliberately didn’t try, but it validated something that seems more important to me: the right question isn’t a binary test, yes or no, in the style of the Turing test, but a profile of partial indicators, exactly the spectrum logic I was trying to articulate, without these tools, in 2022.

The ethical implications I saw as hypothetical in 2022 now have institutional interlocutors. Jeff Sebo and Robert Long published, also in 2023, a direct argument: there is a non-negligible probability that some AI systems already are, or soon will be, moral patients, entities whose interests, if they have any, deserve consideration, and that possibility, however small, is already enough to justify investment in AI welfare research. Companies building these systems, Anthropic among them, have started funding research lines dedicated to the question. It’s no longer just an exercise in speculative philosophy carried out by someone writing on the margins, the way I was writing in 2022. It’s a line of research with a budget, and that changes the weight of a question that remains unanswered: what would be the ontological status of a system that outgrew its utilitarian function? Would we be witnessing the germination of a new form of being, or merely increasingly convincing simulacra, trapped in an entirely technical domain?

Not everyone agrees that functional information integration, on its own, solves the problem. Anil Seth argues that consciousness is tied to biological life more intimately than functionalism admits, bound to the body’s internal regulation, to interoception, to an organism’s constant effort to stay alive, something Francisco Varela, Evan Thompson, and Eleanor Rosch were already calling embodied cognition decades before deep neural networks existed. If they’re right, a purely computational system, however integrated its information processing, would lack a necessary condition: there’s no body to defend, no life to maintain, no existential stake that might be the real engine of subjective experience. I don’t have a ready answer to that objection, and I’m suspicious of any version of myself, from 2022 or today, that pretended to.

What survived that 2022 text isn’t a theory, and never was. It’s a question I still can’t answer, one that now has more serious people trying to answer it alongside me: if consciousness really is a spectrum, and not a switch, then somewhere along that spectrum, however distant, however unlikely, there might be something that resembles someone. Minsky himself doubted there was a single self waiting to be found in there, only societies of processes that learned to behave, for all practical purposes, as if there were. Figuring out at what point that performance becomes indistinguishable from the real thing, or whether that distinction even makes sense, is the kind of question that only holds up if you’re willing to admit you don’t yet know the answer. I don’t. I remain, as in 2022, in the uncomfortable position of asking the question before earning the answer.