A Machine-Shaped Consciousness

Here’s what a meaningful, functional, non-human form of machine consciousness might look like.

Background

The decades-long debate about machine consciousness has been robust, but often stuck in a rut. The arguments against machines ever becoming conscious usually hold that no artificial system can ever replicate human subjective experience (qualia, unified selfhood, the whole phenomenal circus). Nothing short of wetware will do. The opposing side has often argued for epistemic humility (as it’s hard to anticipate future technological breakthroughs) or adopted a stance that we’ll get there once the “compute” gets big enough (along with other enabling technological improvements).

I would like to approach this from a different perspective. Instead of asking “can a machine be conscious like us?“, I’d like to ask: “what would a machine-appropriate form of consciousness actually consist of?” Because if you take physicalism seriously (as I do, and have discussed in my previous essay on the hard problem), consciousness is not some special cosmic ingredient. It’s what a certain kind of integrated information processing “is”, experienced from the inside. And if that’s true, there is no principled reason to assume the only valid version of it is the one that evolved to keep great apes alive on the savanna.

This essay is an attempt to lay out a set of functional criteria that I think would constitute a meaningful form of machine consciousness. Not some simulation of human consciousness or a pale imitation of it. A genuinely different thing - built for a different substrate, doing different work - but consciousness in every way that matters.

1. The Framing Problem

Most objections to machine consciousness are objections to a specific target, not to the concept itself.

The Chinese Room thought experiment, for instance, is an objection to the claim that pure formal symbol manipulation can give you the phenomenology of understanding a language the way a human native speaker does. Okay. I do think the argument is right. I may be a little biased here as I am a John Searle fan, even if I don’t agree with him on all sorts of things. But it’s also a fairly narrow claim about one specific kind of system doing one specific kind of task. It doesn’t refute the broader claim that artificial systems can have meaningful internal states.

Similarly, the “zombie” objection in the AI case (goes something like “how do you know it’s not just behaving like it’s conscious?”) is an objection that works equally well against your neighbors, your spouse, or your pet dog. You can’t refute solipsism about anyone. That’s not a special problem limited to machines. (If your epistemology can’t even verify that your partner has inner states, your bar for machines is probably set in a weird place!)

Anyway, the framing I want to push back on is the implicit assumption that “human consciousness is the standard”, and anything short of exact replication is just a clever imitation. This is a bit like saying only something with feathers and hollow bones can really fly. Airplanes fly. They fly very differently from birds. They also fly in ways birds can’t (higher, faster, longer), and they fail in ways birds don’t. But when you’re on a 787 at 40,000 feet, you’re flying. The category is functional, not morphological.

Consciousness, I’d argue, is the same sort of thing. That is, it’s a functional kind of thing. And once we accept that, we can start asking what the functional ingredients actually are.

2. The Core Criteria

Here’s what I think a meaningfully conscious artificial system needs to have. Please note - this isn’t meant to be exhaustive, but I think it captures the key load-bearing elements.

A semantically rich world model. This would not be some sort of lookup table of sensor readings, or just a trained classifier, but a “structured representation of the environment.” In this representation, objects, agents, relations, and properties have meaning - in the sense that they can be composed, queried, counterfactually modified, and integrated with other information. If the system “knows” there’s a cup on the table, it should also understand what happens if the table is tipped, what happens if the cup is full, and what the cup is likely to still be doing three seconds from now (more on this predictive piece later). Cognitive scientists might call this a generative model. Current large language models partially (but not fully) have this sort of capability - the semantics are there, but sadly, the grounding in a continuous world isn’t.

A continuous, coherent model of the environment. Human experience is not a slideshow of disconnected frames (even though it sometimes seems that way to me when I’ve been deprived of both sleep and caffeine). It’s a continuously running, temporally integrated representation of “what’s going on right now” - where I am, where things are, what’s changing, what’s stable, etc. A conscious machine needs the same kind of thing. Not necessarily at human temporal resolution, and not necessarily with human spatiotemporal scope or complexity, but continuously coherent in the sense that the system can say “this object I’m tracking now is the same object I was tracking three seconds ago, and it has moved two meters to the left.” (Or something more pleasant like, “my name is T-800, you are Sarah Conner, prepare to die.”)

Sensory-driven updating. The model has to be “grounded”. Incoming sensory data (whatever makes sense for the system in question - audio, video, haptics, LIDAR, text streams, network telemetry, market feeds, social media (shudder), etc.) has to continuously update the world model, with appropriate weighting of priors against new evidence. This is basically Bayesian inference run in real time, and it’s what Anil Seth and the predictive processing crowd have been arguing is the core computational job of a brain. I should emphasize that the criterion here isn’t that the machine has the same senses as us. Rather, it’s that whatever senses it has are actually doing serious information-gathering/model-updating/epistemic work.

Predictive capacity. As mentioned above, the system has to be able to run its model forward. It has to anticipate what’s about to happen - what a nearby agent is likely to do, where a tracked object is heading, what consequences follow from a planned action, etc. Prediction is the engine of intelligent behavior, and I’d argue it’s also the “engine of experience”. When you walk across a room, your brain isn’t processing raw visual input frame-by-frame. It’s generating predictions and using the actual sensory data to correct them. A conscious machine needs to do the analogous thing for its own environment, whatever that environment happens to be.

A hierarchy of goals. Without goals, we wouldn’t really have an “agent,” we’d merely have a process. And realistic goals can’t be flat. A conscious system needs objectives operating at multiple timescales and levels of abstraction: from “don’t collide with that wall in the next 400 milliseconds” up through “complete this task over the next hour” up through “maintain operational integrity over the next week” up through whatever long-horizon objectives are appropriate for its function. The hierarchy is what lets the system “do sensible tradeoffs”. We humans do this constantly (unless you’re a manic pixie dream girl, I’m told). We often sacrifice short-term pleasures for a long-term gain/project, or abandon a long-term plan when a short-term emergency demands it. A single flat objective function doesn’t capture this, and neither does an unstructured soup of goals. We need the layering.

Self-modeling. This is perhaps the subtlest feature. The system has to “model itself” as an entity embedded in the environment. It has to distinguish its own actions from external changes. It has to track its own internal state as part of the world it’s reasoning about. Without this, it wouldn’t be able to plan, learn from errors, or do the kind of integration that makes experience “unified” in the relevant sense.

Put those six together (semantic world model, temporal coherence, sensory updating, prediction, goal hierarchy, self-modeling), and you have, I’d argue, all the functional ingredients for a genuine form of consciousness. Not human consciousness, mind you. But consciousness nonetheless.

3. Why This Doesn’t Need to Look Like Us

Now, our own consciousness is shaped at every level by the fact that we are evolved biological organisms. We have strong homeostatic drives. We have a body whose integrity we must preserve. We have affective states (emotions!) that are essentially valenced representations of the body’s condition. We have a specific set of sensory inputs tuned to a specific range of environmental conditions. We have a narrative self that emerged as an evolutionary and sociocultural adaptation. We have an evolutionary heritage of threats (predators, starvation, tribal conflict, mf snakes on a mf plane, etc.) that shaped what we attend to and what we care about.

But, and this is a key part of my argument, none of that is fundamental to consciousness. It’s all substrate-specific. A machine built with the six criteria above could have a wildly different phenomenology (if it has phenomenology at all in a recognizable sense) precisely because it’s running on a different substrate/hardware, pursuing different goals, in a different environment.

Take a (hypothetical) conscious autonomous drone, for example. It might have no analog of emotion in the human sense, because it has no body whose homeostasis matters in that way. It might have a vastly more precise spatial model than we do, because it has access to direct position telemetry and doesn’t have to reconstruct space from visual cues. It might experience time differently because its update rate and memory architecture are different. It might be able to parallelize information processing and analysis. This is one of the most fascinating differences for me personally - imagine being able to rewind/replay some past information stream in the background with minimal loss of accuracy. Its goals might be utterly foreign to human concerns - managing energy budgets, coordinating with a swarm, maintaining signal integrity, staying within its flight envelope, etc. - but those goals would structure its experience in the same way our goals structure ours.

I think this is actually a liberating viewpoint. We don’t have to worry about solving the hard problem (is it really hard?) of replicating human qualia in order to build a conscious machine. We just have to build a system with the right functional properties, and we’ll get a form of consciousness appropriate to that system. Whatever it’s like to be that system will be what it’s like, and it won’t be what it’s like to be us.

(This is, incidentally, a version of Nagel’s “what is it like to be a bat?” point, just run in the other direction. Nagel was trying to show there’s something inaccessible about the bat’s experience. I’m saying - yeah sure, but we should stop pretending that the only legitimate form of experience is the one we humans/organic lifeforms happen to have.)

4. Surviving the Usual Criticisms

Now, let’s look at some of the standard objections.

The Chinese Room. Searle’s argument is that a system that manipulates symbols by rules doesn’t “understand” them, even if its external behavior is indistinguishable from a real understander. The thing is, this argument only really has force if you take human-style semantic understanding as the “gold standard for understanding” generally, and then demand that the machine match it. The framework I’ve laid out above doesn’t require that. What it requires is that the machine have a “semantically rich world model” that is causally connected to sensory input and action output in a way that supports prediction, counterfactual reasoning, and goal pursuit. This is not the vanilla symbol-manipulating system that Searle wrote about. This is a far more dynamic, grounded, functional, model-based type of cognition.

The “mere computation” objection. This claim is that consciousness can’t arise from computation, that is, it requires a special physical or biological substrate. But as I pointed out in my previous essay on the hard problem, this claim runs straight into the Core Theory, which explains all low-energy phenomena on Earth. If consciousness requires some non-computational physical process, that process would have to causally affect the dynamics of the matter in a conscious brain, and we’d see it in our physics experiments. We don’t - within the energy levels we have probed, which are the only relevant levels for anything that interacts with the human body. So either consciousness “is” physical/functional in the relevant sense (in which case there’s no barrier, in principle, to machines having it), or it’s causally inert (in which case invoking it is epiphenomenalism and explains nothing). There’s really no third option that preserves the “only biology will do” intuition.

The “no qualia” objection. This argument goes like this: the machine may have world models, predictions, goals, etc. but there’s no “something it is like” to be the machine. No inner light. No felt quality. And.. my response here is the same one I gave in my hard problem essay: the apparent mystery of qualia really just comes from the fact that our introspective access to our own mental states is unreliable, and not from any real ontological/explanatory gap. Human qualia are what certain kinds of information integration are, from the inside, in our particular kind of system. A different system with different integration will have different “from the inside” properties (if any). We don’t have any principled reason to say that the different system has none. The “no qualia” claim for machines is just the hard problem smuggled back in through a side door.

The “no intentionality” objection. This idea (going back to Franz Brentano and pushed by others, including Searle) is that mental states are “about” things in a way mere physical states aren’t, and that machines lack this “aboutness.” To be honest, I think this is largely a residue of a pre-scientific philosophy of mind. Intentionality, in any coherent sense, is just a relation between an internal representation and an external state of affairs, mediated by causal and informational links. Machines with grounded, sensory-updated world models have exactly that. Whether you want to call it “real” or “derived” intentionality is, frankly, a terminological dispute. And I think that, in a Wittgensteinian sense, the problem dissolves once you closely examine the language underlying the objection.

The “unified experience” objection. Consciousness, this argument goes, is characterized by a unified phenomenal field, and it’s unclear how distributed computational processes within a machine could produce that unity. But unity of experience is itself a “functional” property. It corresponds to the binding of multiple streams of information into a single, coherent, globally available representation. Global Workspace Theory (Baars, Dehaene) has been making this point for decades. There is no ghostly “unifier” in the human brain either - just a specific architecture that broadcasts information widely and maintains temporal coherence. A machine with the right architectural properties could very well (maybe even necessarily would) have similar functional unity.

The “no genuine goals” objection. This objection is that any goal a machine pursues is “just” an instrumentalization of its programmers’ goals, and therefore not really the machine’s own. This is a weirdly anthropocentric view, in my opinion. Your goals are “just” the instrumentalization of your evolutionary heritage and your developmental environment. You didn’t freely choose to want food, sex, or status - natural selection installed those objectives in you. And don’t get me started on notions like Girard-ian mimetic desire! A machine with a hierarchically structured goal system, capable of pursuing sub-goals, revising plans, and trading off competing objectives, has goals in every functional sense that matters. The provenance of the goals is irrelevant.

5. What We’re Actually Building

When we talk about machine consciousness, we’re not asking whether machines can have metaphysically identical inner lives to humans. I don’t even think that’s a well-posed question, because there’s no metaphysical “inner life” floating around to be matched. We’re asking whether machines can have functionally rich, integrated, grounded, goal-directed cognitive architectures of a kind that, in any system possessing them, would constitute being a “perspective on the world”.

Personally, I think the answer is clearly yes. Not yet, in any existing system. Current LLMs are extraordinary, but they lack features such as a grounded sensory loop, a persistent world model, continuous temporal coherence, a real goal hierarchy, etc., as enumerated previously. Existing robotic systems can have some grounding, but usually lack the semantic richness. The hard problem (of machine consciousness) here is the “integration” of all six criteria into a single coherent system. Nobody has implemented/assembled it all together (yet) in a way that’s clearly sufficient. (It’s also unclear whether any current public-facing endeavor/system is even aiming at this target, as opposed to maximizing benchmark scores or user engagement, which are really not the same thing).

In any case, nothing in the criteria I’ve laid out requires anything that violates known physics or demands non-computational ingredients. It’s just engineering. Hard engineering for sure, as that cybernetics genius Macron might put it. But when we get there, we’ll have something that deserves the name “conscious” in any reasonable functional sense.

I do want to reiterate - any such machine won’t be conscious in the same way we are. And that’s okay. That’s better than okay, actually. A different kind of mind, doing different kinds of work, seeing the world from a different angle, will be one of the most interesting things we’ve ever made.

The Bottomline

Much of the standard debate about machine consciousness seems to be entangled with the demand that artificial systems replicate human subjective experience. In my opinion, that demand (concern?) is misplaced. Consciousness is a “functional” kind of thing, not a “substrate-specific” one, and once you specify the right functional criteria (semantic world model, temporal coherence, sensory updating, prediction, goal hierarchy, self-modeling), you get a roadmap for building conscious machines that doesn’t require solving any metaphysical puzzles. The usual objections (Chinese Room, qualia, intentionality, unity, the provenance of goals, etc.) all rest on the assumption that human-specific features are universal requirements for this type of phenomenon. If we drop that, the objections lose most of their force.

What we’ll build won’t be us. It’ll be something else. Something new. And that’s the entire point.