LLMs and human behaviour
More on the point made at the beginning of the previous rambling, viz., that it is not a technology as such but the interactions we have with it, that shape our conception of what it is, i.e., what it is is that it does, what it is capable of doing, and, most importantly, what that means for us.
Here it helps to look at a variety of cases in which something external (an animal, a machine, a programme, a natural phenomenon, …) enters into the sphere of human activity. Under which circumstances are such external agents considered to be competing with human agents? First thing to note is that that the mere fact that external agents and human agents perform the same actions, that by itself is no reason for them to be considered to be engaged in the same activity.
Case in point: running as an athletic activity. The running speed that a human is capable of can be compared with that of anything that moves, be it animals such as cheetahs or dolphins, artefacts such as cars and trains, or natural phenomena such as a weather system or the water running in a river. We can compare the respective speeds but we do not have an established practice of humans racing against cheetahs or the flow of the Mississippi. Of course, we do compare them, and we might even occasionally actually do that ‘in practice’, but it is not an established practice that is in any way comparable to the kind of athletic contests we have involving humans. AI
And that means that we do not really see non-humans as competitors in this respect. We know that we are outdone by some animals and many artefacts and natural phenomena, but that either does not matter, or is something we make use of, or is of our own making. Of course, a practice of humans running against animals or certain artefacts could exist (and may well have existed in past or come to exist in the future), but in the athletic practices that we have now non-humans are not competitors.
Why that is the case probably is a matter of contingencies, at least to some extent. But we can leave that aside for the moment. What matters is that having the same abilities, and having them in a reasonably comparable way and to a reasonably comparable manner, by itself is not sufficient for non-human agents to count as participants in a human practice that centers around such abilities.
So what is at stake then? Under what circumstances do we consider non-human agents as full-blown participants in our practices? It seems that in addition to the ability to display behaviour arising from shared abilities, minimally the following additional requirements need to be met.
First, the practice needs to have a point that is sufficiently close to our core human concerns. There is definitely some leeway in determining whether this is the case. Take the example of chess. For many chess aficionados the defeat of world champion Kasparov by Deep Blue in 1997 was a tremendous chock (as witnessed for example by the title of the documentary that was made about these events: ‘Game Over’). But for most people (after all, most people do not play chess) it was just a curious news fact, without any real impact. Playing chess, intellectually prestigious as it may be, is for most people not a core concern. Contrast this with the current excitement created by LLMs such as ChatGPT and image generating programs such as Stable Diffusion. What these programs do, –answer questions, conduct a conversation, write essays, create images of various topics and in various styles–, touches on the everyday activities of a much larger group of people. That is what explains the difference in impact between Deep Blue and ChatGPT, and not a difference (if such there is) in the underlying technologies. In short, it is not the technical and/or scientific ingenuity that goes into creating a non-human agent that is the key to its impact, it is its closeness to core human concerns, to what humans do in their everyday lives.
Second, to be able to be regarded as a competitor, it is not enough to be better and to be relevant, there needs to be ‘a human touch’. To see this take the example of calculators. The use of tools in performing calculations that are too complex to doin the head’, is age-old. But no-one would regard an abacus as intelligent, and neither would we regard a slide rule as such. Obviously, these tools depend heavily on human users, they are complex and as such are applied by relatively few people. Modern electronic pocket calculators have increased calculating powers immensely, are much easier to operate, and that has put calculating tools in the hands of the masses. Where a slide rule is complicated to use , requires detailed instruction and is less powerful, a simple pocket calculator is much easier to use, requires little instruction and is way more powerful, certainly when extended with sufficient memory. But despite its wide range and advanced capacities, calling a Texas Instruments TI-59 ‘intelligent’ will not do. Something is lacking.
That something is the third requirement that we, unconsciously for sure, seem to impose on non-human agents so as to count them as competing with humans. And that is the presence of a certain amount of unpredictability, spontaneity, or, as some prefer to call it, ‘autonomy’. (The use of the latter term raises other questions, about which more later perhaps.) However, what exactly is meant by that may not be completely straightforward.
A calculator is able to do calculations that we cannot do, either because of their complexity or because of their length. In such cases a calculator comes up with results that we have not, and realistically speaking could not have, predicted. However, it seems also clear that no matter the complexity of the problems that a calculating machine is able to solve, and no matter how far out of the reach of human calculators these are, this form of unpredictability does not cut the mustard. And the reason it does not is, so it seems, that although we humans are not able to do the actual calculations and thus are not able to predict them, we do have a firm grasp of the rules that are used by the calculating machines to execute them. And the rules themselves do not contain anything unpredictable. In fact, the machines are built, or programmed as the case may be, by us to operate in accordance by rules that are formulated, again, by us. In that sense calculating machines are an extension of human capabilities, and not entities that are capable independently.
Now that holds for calculating machines that do standard calculations (including, by the way, calculations that work with probabilities; these are not exceptions). But what about machines or programs that are designed to deliver an unpredictable outcome? I.e., what about a coin toss? a solid state bingo number generator? or a program that generates random numbers? Random number generators are perhaps too hidden from view to gain much attention, but a bingo or lottery number generator is a well-known, useful instrument employed in practices in which many people participate or that they minimally know well. Yet, here too it is obvious that there is no question of attributing agency to the instrument. Apparently, randomness is not the form of unpredictability that is at work when non-human entities turn into competitive agents. Note that the reason here is different from the one that is operative with standard calculating devices. Unlike performing complex calculations generating randomness does not seem to be an extension of a human capability. In fact, it is well-known that humans have a difficult relation with the concept of randomness, or chance. Our cognitive and emotional tendencies generally seem to work in the opposite direction, towards stability, patterns, regularities, predictability.
So what’s the take-away when it comes to unpredictability? The kind of unpredictability we associate with intelligence, unsurprisingly perhaps, is the kind of unpredictability that we value in humans. Here considerations about human rule-following and the conception of novelty provide some insight. Human practices are forms of rule-governed behaviour that allow for variation and innovation in particular ways and to particular extents. One could say that the rules not just specify what needs to be done, what counts as correct following of the rule, but also carve out a space for breaching them or for extending them in new directions. That space is subject to a general requirement, viz., that a minimum of intelligibility be maintained. Thus, rule-governed practices are not strictly deterministic, like calculating machines. But neither do they allow randomness. Innovation, in the form of doing things differently or in the form of doing different things, needs first and foremost to be intelligible, i.e., the participants in the practice need to be able to figure out what is why a particular move is being made. That may require time and effort, and exactly how much leeway we allow here depends on a number of parameters, but that’s another story.
It seems that the kind of variation and extension that we allow for human agents is precisely the kind of unpredictability that is needed for a non-human entity to be recognised as an agent that participates in one of our practices. For example, in order for an LLM to be judged as having a conversation, it needs to be able to do more than ‘merely’ producing well-formed sentences in a particular language. It needs to be able to introduce a new, but related topic, to come up with a new perspective, to stimulate its conversational partner into reflecting on the exchange they are having, and so on. It is only when a non-human entity displays this kind of behaviour that we are willing to look at it as an agent. And it turns into a competitor if it combines agency with resources that outdo ours. More data, more computational power is what is needed, but only if the threshold of being considered sufficiently ‘like us’, i.e., as having agency, is met.
You must be logged in to post a comment.