Don't Hold Your Breath For Digital Consciousness
Do you ever get so fed up with the nutjobs who insist their favorite chatbot is conscious, that it makes you want to call them names and prove them wrong? Well, I know a special name you can call them: "strong computationalists". Now, this isn't necessarily a bad word -- some strong computationalists (unlike those aforementioned) were serious people: Hilary Putnam, Jerry Fodor, Daniel Dennett etc. This makes the prospect of proving them all wrong a bit shakier, but today, I'm going to try.
What is strong computationalism?
Put simply, strong computationalism is the idea that the right computation, implemented in the right way, equals consciousness. Proponents differ in the constraints they impose on the structure of the computation, or the manner of its implementation; however, they typically agree on the principle of multiple realizability: the same abstract architecture can be implemented by different physical systems, to the same effect. Due to the tension between the demands for both multiple realizability and a "right implementation", strong computationalism has diverged into quite a few variants, each with its own nuances. Instead of trying to address them all, I will temporarily throw nuance out of the window and define a fairly general version -- let's call it 'naive computationalism' -- by way of two commitments:
- Consciousness is a computational process, or multiple interacting processes, realizing the dynamics of the mind
- Consciousness is implementation-independent -- any physical means of performing computations can make consciousness
This position intersects with many flavors of strong computationalism, but what makes it "naive"? Primarily, its second commitment: it embraces multiple realizability to the fullest extent, omitting special criteria for what counts as an implementation. By accepting that a computation is a computation, naive computationalism opens itself up to absurd conclusions (soon to be explored) that more sophisticated computationalists strive to avoid.
So why set up a proposition doomed to absurdity? Well, instead of dismissing it out of hand, I will proceed by the principle that 'absurd' does not equal 'impossible', and use its example to derive more stringent criteria to judge by. Those criteria, I believe, can shed some light on the most controversial aspects of strong computationalism as a whole.
But first, let's talk about processes.
The Merriam-Webster dictionary provides the following two definitions of the word 'process':
- A natural phenomenon marked by gradual changes that lead toward a particular result
- A series of actions or operations conducing to an end
The dictionary is rarely my favorite place to start, but in this case, it touches on the core of a contentious issue: the word 'process' comes in two different flavors, each with different implication, but computationalism may intentionally conflate them. For physicalists who reject illusionism, consciousness falls under the first definition; but for the naive computationalists among them, it may fall under both definitions simultaneously.
As those familiar with Turing machines know, any computational structure, parallel or otherwise, can be broken down into a series of steps.
If naive computationalists are right, such a series of steps also constitutes a phenomenon, so long as it's implemented by physical operations. When the system performing those steps is a biological brain, the result is the natural phenomenon of consciousness. Thus the two distinct meanings of the word 'process' are unified, but this is not necessarily desirable.
Such a hypothetical phenomenon -- computational consciousness -- is capable of transcending the spatial and temporal constraints that bind all normal phenomena. In principle, each computational step can be completed in a different time and place: so long as the intermediate state is recorded somehow, there can be arbitrary intervals of space and time between them. Combined with unbounded multiple realizability, this "feature" can be exploited to construct seemingly absurd scenarios:
For instance, I could buy a big notebook and start manually computing a mind in it.
I would likely die before making any significant progress, but my grand-grandchildren may accidentally find the notebook in the attic and feel inclined to compute a few more steps, continuing the process. Once they get bored with it, the associated mind will presumably hibernate -- indefinitely, or perhaps forever. This possibility violates some intuitions, at the very least.
Still, a devotee of the theory will merely shrug at this: yes, the process looks haphazard from the outside, but the simulated mind is none the wiser -- the computation maintains for it an independent, internal frame of reference. This presents an interesting challenge: if 'absurd' doesn't equal 'impossible', can this be properly refuted?
Maybe.
To my knowledge, no uncontested natural phenomenon can be freely distributed across time and space.
Yes, some phenomena go through "dormant" phases, while others cover great distances with apparent discontinuities; these may rely on elaborate, oscillating patterns of activity in both time and space, sometimes creating the appearance of a arbitrary stop-and-go. But there are patterns, which are characteristic of, and integral to, each phenomenon: they are direct evidence of propagational constraints.
In contrast, a computational process -- even when physically implemented -- is an abstract collation of events. This may not be obvious with a continuously operating computer, but the notebook example showcases it clearly: to recognize the relevant process, one must pinpoint and connect the dots between a few "special" events, scattered randomly in a vast ocean of them -- an ocean that spans all four dimensions. From this perspective, the process seems to violate the basic physical principle of locality: one event is taken to lead into the next, without any clear or necessary path of mediation -- as if by spooky action at a distance, or at least predestination.
Since no other physical phenomenon relies on such a perspective for its analysis, this serves as a reasonable heuristic for a lack of physical viability. At the very least, "computational consciousness" becomes an extraordinary claim, demanding extraordinary evidence that can scarcely be provided. But this merely shifts the burden of proof based on empirical observations. Why consider this heuristic reliable? After all, even in the notebook scenario, the computational events are, in fact, mediated by a chain of physical causality. It just happens to be contrived, consisting of links that are mostly irrelevant to the logical trajectory of the computation (e.g. the sequence of events leading up to my descendant's birth, then to his eventual finding of the notebook etc).
Why associate some causal chains with phenomena, but not others?
Because the causality of a phenomenal process follows a more or less distinct trajectory amid the interconnected jumble of the entire world's events.
It has its own kind of momentum that drives it forward, supporting the continuity of its trajectory, which enables it to be traced. This is precisely the quality that makes a phenomenon objectively recognizable: its characteristic propagation can be traced from start to finish, both in terms of manifestation and causality.
The activity of a true phenomenon may fluctuate, on its own terms, but it can't be stopped and restarted arbitrarily: if interrupted, it quickly decays, losing the causal momentum. Even if we could somehow "freeze" it, we would be faced with something akin to the problem of cryonics. A true phenomenon lives continuously -- if by obscure means, or in an attenuated way -- or else it dies. A computation likewise has a "momentum" of sorts -- its state -- however, this is abstract and independent from the causal chain that helps it unfold; instead, the computation relies on external intervention, or sheer chance, in order to progress.
I propose that phenomena are distinguished by their intrinsic causal momentum, giving them a distinct causal trajectory.
If that sounds murky, consider Conway's Game of Life as a microcosm to illustrate the point:
The game's rules are applied uniformly, all across the board, with each cell's subsequent state depending on the current state of its neighbors. This can be framed in terms of causality: cell state transitions are caused by each cell's local neighborhood triggering one of the Game's three rules. The cause of a given cell's state can be traced back in time, first by considering its neighbors from the previous step, then the local neighborhoods of those neighbors two steps ago, and so on and so forth -- going backwards in time and outwards in space.
Thus a causal account can be given for each cell, involving a growing number of factors the further back it goes. Such an account is indistinct in principle: the exact state transitions may vary between cells, but the overall causal structure is uniform, following the topology of the grid. Any emergent shape could be accounted for in this manner by composition.
Indeed, some shapes can only be accounted for by considering ever-expanding areas of the grid, which happened to converge into this outcome by chance. Sounds familiar? In the notebook scenario, causally connecting the computational steps would involve not only an account of the incidental circumstances behind my descendant's finding the notebook, but also the secondary circumstances that facilitated those primary circumstances and so on.
However, there are other shapes, like the "block" or the "beehive", which are stable and self-sustaining: save for destructive interference, they can be accounted for independently from the environment. More strikingly, shapes like the "glider" or the "loafer" are not only self-sustaining, but dynamic: they cycle through a number of distinct configurations as they propel themselves forward, traveling across the grid.
These could be considered proper phenomena in Conway's minimalistic universe: despite the ultimate interconnectedness of all events on the grid, they exhibiting a distinct causal trajectory, allowing for a parsimonious and characteristic causal explanation. Moreover, they exhibit causal momentum: their configuration are either self-sustaining or imply an identifiable sequence of followups.
This demonstration substantiates the qualitative distinction between phenomena and computations on one hand, but seems to undermine its purpose on the other:
Not only is the Game of Life itself a computation; it is, in fact, a Turing-complete computer. By carefully choosing the initial state, one can interpret the changing patterns to compute anything. A sufficiently clever choice may even result in a computation that gets interrupted and then, via an unlikely sequence of intermediate states, resumes. To the underlying system, a glider is exactly as real (or unreal) as the activity behind some contrived computational setup: if the Game of Life were the universe, there would be no ontological difference between the two cases.
It may be tempting to extrapolate from this observation and elevate computations to the status of phenomena in reality: if the universe is a collection of trivial elements, interacting locally according to a set of rules, perhaps the same logic applies. It could then be argued that the issue of causal momentum is epistemological rather than ontological. But such an attempt can backfire: reading too much into the Game of Life invites skepticism regarding all phenomena, save for the most elemental. If the universe itself is agnostic to the qualitative difference between phenomena and computations, it is likewise agnostic to the one between meaningful computations and computational pareidolia. Moreover, the omnipresence of incidental cyber-Boltzmann brains would be impossible to disprove.
Nonetheless, since I've opened this can of worms, I now have to close it, and doing so may even get us closer to the root of the original issue.
So, the Game of Life is similar enough to Real Life to provide insights about the causality of phenomena, but is it different enough to dismiss the doubts about their reality? The universe it presents is discrete in every aspect: it evolves in quantized steps, where discrete cells undergo discrete state changes, all of which are perfectly explainable in terms of discrete causal sequences.
Humanity's scientific traditions, especially Classical Physics, tend to view the universe in a vaguely similar fashion: as a big Rube Goldberg contraption, made up from a collection of separate parts, whose distinct interactions can be sorted into causal chains explaining emergent phenomena. Granted, the fabric of spacetime itself is taken to be continuous, but the world within it is explained using spatial and temporal dissections. On the more cutting edge, quite a few theories suggest that the universe is indeed a cellular automaton, not so simple as Conway's, but of the same basic nature; these question even the continuity of space and time.
This tendency for reductionist division implies its own dual: if a whole can be nearly deconstructed into parts, can't it also be constructed from them?
While deconstruction is the attitude of physics, construction is the attitude of engineering. Engineering could even be viewed as the enterprise for the construction of artificial phenomena. As it grows more sophisticated, it erodes the intuitive distinction between process as an evolving phenomenon and process as an enacted sequence of operations. Ultimately, it's this side of the reductionist coin, that gives rise to the computationalist hunch. After all, biologists also love to dissect an organism into crude parts, but they know better than to think they could put it back together!
So perhaps it's this other kind of "whole" -- the constructed whole -- that deserves to be doubted. It is merely put together, and then deemed whole, by virtue of combined function, or by convention. If it enacts a process, that process is not a preexisting fact interpreted through the framework of causality, but one specifically built within the extent of that framework's applicability -- however far that goes. Hume's famous skepticism about causality itself -- and its characterization as a mental habit -- comes to mind.
A natural whole, even if theoretically deconstructible, would only be similarly suspect once put back together to demonstrate an equivalence. This can only rarely be done, perhaps never for an organic or dynamic phenomenon.
Suppose, then, that the mark of a true whole is a kind of irreducibility, which allows discrete and causal approximations to arbitrary degrees of precision, but bars perfect replication. If so, the quality of causal momentum is not the cause of a ontological difference, but a consequence of it: the inherent continuity of a true whole simply lends itself to parsimonious causal explanations. Consequently, artificial constructions with a clear causal trajectory can blur the line empirically, even if they lack the intrinsic causal momentum needed to cross the ontological barrier.
So what about a computer based on a self-sustaining mechanism? Wouldn't it have causal momentum? This would perhaps be akin to the glider from the Game of Life: a collection of inherently discrete elements in close proximity, acting together continuously as a self-propelling unit, under the influence of reality's physical laws. It can still be distinguished from a true phenomenon by invoking the criterion of irreducibility directly -- if one accepts that criterion. But even elsewise, the computer is not the computation: the computational flow may mirror the causality of a phenomenon, but the causality driving the computer itself would likely differ significantly, so that the computational states would not truly lead into each other, the way the physical states of the real phenomenon do.
At last, this brings us to the matter of more sophisticated computationalists -- the ones who would single out the notebook scenario as incorrectly implementing the process.
For instance, consider connectionists: they don't think in terms of Turing machines or the Von Neumann architecture, but in terms of neural networks -- often of the familiar kind, which may well run on a scalar processor (or a notebook). However, they will typically insist that it be physically implemented using distributed, real-time processes -- the way they think a brain works. This helps them avoid the absurdities of naive computationalism, but they struggle to explain why such criteria are necessary, in a principled way. In effect, they wish to emulate the brain's characteristic causal trajectory without causal momentum; or, more deeply, to mandate the appearance of irreducability within a perfectly reducible architecture. This merely amounts to a cargo-cult-style attempt to summon consciousness.
A more interesting case is that of certain computational functionalists, who propose that the causality driving the implementation of a computational consciousness, should somehow mirror the computation's abstract logic. But how can that be accomplished, if true a phenomenon is dependent on causal momentum? A computation can describe a phenomenon that propagates itself, but it doesn't propagate itself, except from a perspective that any given thing is an analog computer of its own dynamics. Accepting that, and respecting multiple realizability, the computational model becomes a description of a class of genuinely brain-like analog computers -- a kind of formal analogy. This bars digital computers, whose state changes only suggest the simulated phenomenon's causality, while the machine itself operates according to its own principles -- a high-resolution fluid dynamics simulation may melt your processor, but it won't turn into a pool of water.
So what to make of all this?
If we live in a universe where causal momentum is an ontological status symbol, general computations are second-class citizens, making naive computationalism untenable. If we live in one where it's only an empirical criterion, then either unrecognized computations are omnipresent to the point of making computation trivial, or computation is a matter of subjective interpretation (both suggesting illusionism with respect to consciousness as well as computation).
A philosophical framework where irreducibility implies causal momentum, which in turn implies a distinct causal trajectory, provides a principled basis for skepticism against strong computationalism in general. It establishes an ontological difference between true phenomena, blatant non-phenomena, and subtle imitations, as well as empirical criteria to tell them apart. The seemingly arbitrary conditions sophisticated brands of computationalism impose on the way conscious computations are to be implemented, can now be interpreted and classified with respect to those criteria.
Moreover, this framework suggests a plausible root cause for the differences in intuition between strong computationalists and skeptics, or even among different brands of computationalism: on a surface level, the debate pits the constructed and enacted against the emergent and evolving, as modern engineering continues to diminish the perceptible gaps between the two; on a deeper level, it invites reflection on the ontological difference between them (or lack thereof), which may ultimately be rooted in the question of reductionism versus holism.
Even if this framework is not accepted in its entirety, causal trajectory and causal momentum can prove to be useful tools for analysis. Or at least for beating Sam Altman's marketing stuntmen over the head with.