I dust off, revise, and expand this old entry in connection with discussions with Malcolm Pollack, Dmitri D. et al. about artificial intelligence. One of the questions is whether AI systems are subjects of intrinsic intentional states. I incline to the view that they are not. Since intentional states are meaning-laden, I further incline to the view that no AI system, not even the most advanced LLMs, understand language or anything else. They are merely syntactic engines.
………………………….
1. Intentionality: Original and Derived
All will agree that there is a distinction to be made between original (intrinsic) and derived intentionality. Intentionality is aboutness. For example, a map is not the terrain it maps; the map is about the terrain. What makes the map about the terrain? A map is not about a chunk of terrain and its geological features just in virtue of the map’s physical and geometrical properties. Consider the contour lines on a topographical map. The closer together, the steeper the terrain. Closer together means steeper. But surely the lines and their arrangement have no representational power in themselves. That closer together should mean steeper is a meaning assigned by map-makers and agreed to by map-users.
The point is that a map’s meaning is not intrinsic to the map qua physical object. The meaning derives from us. There has to be one or more meaning-assigners and meaning-apprehenders external to the map to give the map a meaning and determine what that meaning is. The map’s intentionality is derivative: it derives, or has it source in, our intentionality, which is original.
So some things, maps for example, derive their semantic properties from other things, map-makers and map-users, for example. Semantics has to do with sense and reference. ‘Meaning’ covers both. Setting sense aside for now, the reference or referent of a map is that to which it purports to refer.
But note that a map needn’t actually refer to anything to be a map. It need only purport to refer. This is a well-known characteristic feature of intentionality or aboutness or object-directedness: the existence in reality of the intentional object does not follow from the fact that something intends it. Thus the existence in reality of a chunk of terrain does not follow from the existence of a map of it. The map below purports to refer to something extra-cartographically real, but fails in its purport. It is no less a map for all that.
Suppose I draw you a map so that you can find my camp in the extra-cartographically real Superstition Wilderness. I use the Greek uppercase Φ to mark my camp and the Greek uppercase Ψ to mark the camp of a heavily-armed crazy man that you are well advised to avoid. I intend that Φ designate my camp, and Ψ his. That intending (narrow sense) is a case of intentionality (broad sense). I am the meaning-assigner. I assign a referent to the Greek symbols, thereby making those inscriptions mean (refer to) something, symbolize something.
A symbol may be viewed for certain purposes as a purely syntactic unit. To view it this way is to abstract from its semantic properties. A symbol in use, however, a symbol that symbolizes, a symbol that exercises its symbolic function, is a syntactic unit that means something other than it, and is therefore more than a merely syntactic unit. A syntactic engine is a machine that manipulates symbols qua syntactic units, ‘tokens’ is what the AI guys call them. A computer is just a syntactic engine. The same is true of the most advanced AI systems. The manipulation of symbols as mere syntactic units cannot confer meaning upon them. Such ‘manipulation’ is neither constitutive of nor sufficient for meaning.
Or at least this is a question we need to consider. I incline to the view that, just as you cannot squeeze blood from a stone, you cannot squeeze semantics from syntactics. No AI system, not even the most advanced LLM, can by itself, mean anything or understand anything.
It is also important to note that the very distinction between syntax and semantics presupposes languages and understanding, and thus presupposes the antecedent ‘marriage’ of the syntactic and semantic. How then can we reduce the latter to the former? I’ll spell this out in later posts.
I have said enough to make clear the difference between derived and original intentionality. The symbols on my map derive their intentionality or aboutness from my intentionality which, relative to theirs, is original. But if the origin of their intentionality is in me, and not in them, might it not be the case that my intentionality is itself derived? After all, it does not follow from the fact that I am the source of the intentionality of the symbols on my map that my intentionality does not itself have a source. But if so, then a regress threatens which appears to be both infinite and vicious. But before discussing this further, I need to bring in another point.
2. Ascriptivism Versus Realism
Daniel Dennett appears to hold an ascriptivist or projectivist or imputationalist or instrumentalist theory of intentionality: nothing is an intentional system intrinsically or an sich; a system is intentional only in virtue of someone’s assigning intentional states to it. Thus there is no intrinsic intentionality in my computer (when it is running a chess program or any program). But it may be nonetheless useful to take up the “intentional stance” vis-a-vis the computer and impute to it beliefs and desires (e.g., the desire to inflict mate) in order to explain its behavior. Given the computer’s complexity, the intentional stance promises an easier explanation than the physical and design stances do.
A realist about intentionality, on the other hand, is someone who holds that some systems really are intrinsically intentional: their having beliefs and desires and other object-directed states is not a matter of anyone’s viewing them as having such states. I am a realist: some systems are intrinsically intentional. But there are two ways to be a realist.
One way is by being a reductive realist. Such a person admits the reality of intentional states but identifies them with something believed to be more fundamental such as brain states. If you believe, with Fred Dretske, that intentionality does not need naturalizing, but already is a natural phenomenon in the natural world, then you are presumably a reductive realist about intentionality. You are saying this: there is intrinsic intentionality, but it is physical in nature.
I’m a realist too, but a nonreductive one: I hold that intentionality is irreducibly real: it cannot be reduced to, or identified with, anything else. It is what it is and not some other thing. It is what it is and it is spiritual in nature. Part of the reason I believe this is because the alternatives are quite hopeless. But what on earth or in heaven does ‘spiritual’ mean? This is a challenge I face.
3. The Ascriptivist Regress
Back to Dennett. Suppose one takes the ascriptivist line. Then an infinite regress appears unavoidable. Suppose I ascribe intentionality to my chess playing computer. Quite obviously, my ascribing, projecting, imputing, is itself an intentional state or a series of such states. So if ascriptivism is true, then my acts of ascribing must themselves be ascribed — otherwise they are intrinsic and Danny’s game is over. This launches us on a regress that is obviously infinite since at each level, one can “kick it up a notch.” It is also clear that this regress (unlike the truth regress) is vicious since at no level is the explanation of intentionality complete.
One might think to avoid the regress by embracing a circle: I ascribe intentional states to the computer and it ascribes intentional states to me. But surely this is even more of an absurdity for reasons that needn’t be belabored.
The upshot is that pure ascriptivism is incoherent, an utter non-starter. Dennett’s actual view is a bit different.
4. Adulterated Ascriptivism : Homunculism
Perhaps the solution is to mix some realism in with the ascriptivism. Maybe it goes like this. I have a head full of homunculi. These little men and women, working together, ascribe intentionality to me. But each homunculus has its intentionality ascribed to it by other, stupider, homunculi which are constituents of it. The stupider homunculi, in turn, are composed of even stupider ones, and so on until we get to the level of individual neurons which aren’t ‘about’ anything. The base level homunculi are so stupid that one could say that they don’t even rise to the level of being either stupid or intelligent.
Does this avoid the dilemma of having to choose between a vicious infinite regress and a vicious circle? What Dennett is proposing is a finite regress that terminates with something naturalistically acceptable, namely entities that lack intrinsic or original intentionality.
Admittedly, this proposal gets rid of the infinity of the regress, but not its viciousness: we still have no explanation of intentionality. Consider the base-level homunculi. They are so primitive as to lack all intentionality. How then can they ascribe intentionality to their colleagues one level up? Surely it is obvious that the ascribing of intentionality is itself an intentional act or series thereof.
The point is quite simple. If the regress terminates with base entities utterly devoid of intentionality, then no higher level entities will be intentional. For the higher-ups get their intentionality only from the ones lower down in the hierarchy. If the ones at the very bottom have no intentionality, then they can’t transmit it up. But if, on the other hand, the ones at the very bottom possess intentionality, then it is false that all intentionality is derived.
In the final analysis, Dennett is an eliminativist about intentionality. His position amounts in the end to the denial of intentionality. To see this, you just have to think clearly. Dennett is saying that all intentionality is derivative, none is intrinsic or original. But that makes sense only if one embraces an infinite regress. But in this case an infinite regress must be vicious. On the other hand, a regress that terminates either terminates with entities that are intrinsically intentional or entities that are not. If the former, the game is over. If the latter, no intentionality gets transmitted up and Dennett is an eliminativist about intentionality just as he is an eliminativist about consciousness and qualia.
Hi Bill
I confess to not understanding the infinite regress argument. And I have my reservations about LLMs classification as syntactic engines (I assume you use this term in John Haugeland’s sense as he invented the concept if I remember correctly). LLMs outputs have the appearance that they are produced by such engines but their underlying math and modus operandi are nothing like GOFAI syntactic engines. LLMs theoretical foundations are as remote from math logic based syntactic engines as statistical thermodynamics is from the classical one.
Independently, I’d like to suggest a short article by Mitchell and Krakauer that expresses, much better than I can, a balanced point of view on modes of understanding I am inclined towards. Here is the authors’ key point:
It could thus be argued that in recent years, the field of AI has created machines with new modes of understanding, most likely new species in a larger zoo of related concepts, that will continue to be enriched as we make progress in our pursuit of the elusive nature of intelligence. And just as different species are better adapted to different environments, our intelligent systems will be better adapted to different problems.
https://www.pnas.org/doi/epdf/10.1073/pnas.2215907120
I think that the question of whether the current LLMs understand or not is much closer, from the structural perspective, to the question about modes of being I asked you elsewhere. It is conceivable that there are different models and levels of understanding just like it is conceivable that there are different modes of being. Internal consistency and empirical adequacy matter in evaluating the different approaches (models), but there is probably no single correct framework. A practical bottom line with philosophical consequences appears to be that linguistic competence is achievable without intentionality.
Dmitri writes, “I confess to not understanding the infinite regress argument.” Let me see if I can present it more clearly.
1. Belief and desire are intentional (object-directed) states. I cannot just believe; necessarily, if I believe, then I believe something, e.g. that bishops in chess move along diagonals and never along ranks and files. Same with desire. My chess opponent cannot just desire, if he has desires; he has to desire something, e.g., that I run out of time.
2. Now suppose with Dennett that there is no real intentionality anywhere in anything , which implies that no computer running a program (e.g. a chess program) is ever in an intentional state.
3. This fact about the computer cannot stop me from assuming the “intentional stance” with respect to the computer. Adopting this stance, I ascribe intentional states to the computer, states it does not have. That is, I view the computer AS IF it e.g. wants (desires ) to mate me in under ten moves.
4. This ascribing that I do is (a series of ongoing) intentional states transpiring in me. For, necessarily, to ascribe is to ascribe something.
5. Now this ascribing is either really occurring or it is not.
6. It cannot be really occurring on Dennett’s view since there is no real intentionality anywhere.
7 So, given that intentionality is either real or ascribed, my ascribing of intentional states to my computer must itself be ascribed to me.
8. By whom? By an ascriber A1 distinct from BV.
9. But of course A1’s ascribing is itself an intentional performance, which implies that there has to be an A2 that ascribes an intentional performance to A1, and so ad infinitum.
10. Now not every regressus ad infinitum is vicious. Some infinite regresses, if not virtuous, are merely uneconomical. Consider the truth regress: p, it is true that p, it is true that it is true that p, etc. Every true proposition spawns an infinity of truths. (It doesn’t matter whether we think of this infinity as actual or as merely potential.)
11. The “intentional stance” regress is vicious. It involves a contradiction: Intentionality is not real anywhere or in anything AND intentionality is real in the case of an agent adopting the intentional stance.
NOTE: to avoid exegetical quibbles, remove ‘Dennett’ and replace it with ‘Dennett*’
Now Dmitri, is this not a knock-down refutation? If you don’t think so, point to which step you reject and why you reject it. “Your move,” as the chess players say.
I have difficulties around these stages:
7 So, given that intentionality is either real or ascribed, my ascribing of intentional states to my computer must itself be ascribed to me.
8. By whom? By an ascriber A1 distinct from BV
9. But of course A1’s ascribing is itself an intentional performance, which implies that there has to be an A2 that ascribes an intentional performance to A1, and so ad infinitum.
On the reductionist’s view of Dennett intentionality does not exist just like phlogiston. Ascribing intentionality or adopting an intentional stance towards somebody or something is not an intentional act. What is going on with this ascription process is a total mystery to be solved by future science for the eliminativist Dennett, but there is no regress and no contradiction. In a joking eliminativist summary form “ascribing of nothing can be done by nobody.”
A very ‘creative’ reply, Dmitri. I dismiss it as incoherent. You might as well say: you can adopt the intentional stance without adopting the intentional stance. You understand, of course, that the int. stance is not a physical stance like that of a boxer in the ring.
Bill
All I meant to say is that Dennett’s intentional stance is knowingly — and by design — incoherent in the same way as Fodor’s more clear slogan which I am citing from memory: “if intentionality is real it really must be something else”. Dennett and Fodor are in the same physicalist/materialist camp on this issue, that is ideological anti-realists with respect to intentionality. Applying an infinite regress argument to a clearly incoherent position which, I have to repeat what I said above, the authors hope to be resolved by advances in future science appears to me a somewhat superfluous effort. I am not defending Dennett’s concept as it is logically indefensible. The act of ascribing that you are using in your argument means something else to Dennett. The two of you are talking past each other. Your key terms are not in sync with each other.
Bill
For anyone interested in up to date philosophically informed survey of AI in historical context with plenty of references and discussions of the inner workings of LLMs, I’d like to suggest reading two part paper on A Philosophical Introduction to Language Models by Buckner and Milliere found here http://cameronbuckner.net/professional/puball.htm
Hello Bill,
Back in February you kindly noted a criticism I made of an argument of yours and Ed Feser which concerned the intentionality of photocopies. I would make a similar criticism of the present piece. You say above,
I appreciate that this is how we ordinarily speak of the meaningfulness of symbols. We would call this a case of ‘derived intentionality’. But I think this is a surface account, a façon de parler, or a mere ascription. For meanings arise only in suitably prepared minds (or so I claim). For ‘Φ’ to represent your camp to me you have to tell me of this meaning, perhaps by writing ‘Φ = Bill’s camp’ in the map’s legend. So we fall back on language to transmit meanings. I reject ‘derived intentionality’ but will continue to speak of the meaning of a symbol. By this I mean the understanding sparked by the symbol in me and others similarly prepared.
Regarding contour lines on a map, you say that ‘closer together should mean steeper is a meaning assigned by map-makers and agreed to by map-users.’ This might suggest that contour lines are almost as arbitrary as your Φ. But we could paint lines of constant height and regular vertical separation (‘real’ contours as it were) onto a landscape, and a photograph of said landscape taken from a height would look just like the map in respect of the contours. So a map is in some respects a model of a landscape. Distances and bearings on the landscape can easily be read off from the map by an algorithmic process. This suggests that there is gradation in the complexity of extracting meaning from symbols, with models being relatively simple and paragraph-long Dickensian sentences with nested subordinate clauses rather hard. And this leads us to something adjacent to Dennett’s homunculism. We are trying to envisage mental structures that can encode meanings beneath the level of language and yet can be reported in language. As with Dennett the complexity of meaning is reflected in a hierarchy of structures. But our hierarchy is not an upward flow of intentionality from nothing, like an inverted river system fed from hidden springs. Instead, at the bottom we can place elementary experiences, ie, qualia. The encoding of ‘Cerberus is fearsome’, for example, might be such that when our attention alights on Cerberus a frisson of fear occurs in us, and likewise for other qualia words. So on this rather empiricist theory meanings are constructions out of experiences. And it leads to the same conclusion as yours regarding AI systems.
David,
Thanks for the comment, but I don’t know what to make of it. My point is very simple. Symbols as such have no intrinsic meaning (sense or reference). Whatever meaning they have is ascribed to them by original intentionality. It can’t be ascription all the way down on pain of a vicious infinite regress. Homunculism doesn’t get us anywhere since the base-level homunculi are too stupid to mean anything.