I had the misfortune of working in tech back when Bitcoin came on the scene. An interesting aspect of this is that, for a summer, half the office (even then, we could already tell it was the bad half) was seeped in podcasts and video content about currency, and I remember at least a couple of tech bruvs (this was London, not Silicon Valley) discovering, somewhat astounded, the concept of fiat currency. One conversation in particular felt very "My First Social Construct".

Fiat currencies are not backed by anything, they just are. Backed currencies would always come with the temptation of issuing more than you had collateral. They were only backed as long as everyone didn't redeem the backing at the same time, so if that was good enough, why not back them on zero collateral? We think of fiat currency as a post-war thing because the US suspended the convertibility of USD into bullion in the 1970s, so everyone was like "well fine, then, let's use all that gold for the palace... cause we totally had loads, haha" or something. But instances have existed in medieval China and Early Modern Europe. It's a thing. That people do. Currency is a system of trust. Backing, traditionally, has been the material anchor for this trust, but once systems of trust grow large, complex, and interconnected enough, they can be self-sustaining through purely social, rather than material, modes. Your quid, buck, or guinea is such because it acts as such in the social world.

Hold this thought for a moment, park it aside, and consider language. In the kind of structural linguistics typified by Ferdinand de Saussure, language is a system of signs, which maps signifiers (words, spoken or written, or, I guess, tokens, because you see where this piece going, right?) to signifieds, thus held to be platonic concepts. The noun 'cat' and the verb 'jump', each representing a concept rather than specific instances, can enter into a syntagmatic relationship: 'the cat jumped'; the concepts, in turn, share paradigmatic relationships. A jumping cat is an idea which makes sense in a way that a jumping house doesn't.

So far, so neat. Except, of course, this is all theory: surely, because language represents the world, we must have signifieds, paradigmatic relationships matter, it makes sense to model it like that, but we can't prove these things actually exist. The concept evoked by the word 'cat', or 'house', may be a consensual one across a vast number of people; but I'm sure you can think of plenty of words for which this is not the case. Post-structuralist linguistics refute the existence of a 'transcendental' signified, and instead see meaning as an endless chain of references, and, if we couldn't prove the existence of signifieds, we can certainly disprove it. Certainly, this is what language models have done.

There is enough substance-in-the-structure of what we thought was representation that the meaning represented in this structure (your alleged signifieds) emerges in the linear algebra of a model trained solely on representation. The structure apparent in the training data, through un-supervised learning, is enough to model that, say, a house is not the jumping kind of noun. Like with fiat currencies, there doesn't need to be a real referent, over there, for the structure of the reference to do its job, as if there was a real referent. I'm not even talking LLMs here, that's already a thing in the early 2010s with word2vec.

Now, educational assessment, say, in higher education. For a long time, assessment in education was thought of as a specific application of psychological measurement. In fact assessment is only called that since the second world war, and the term 'educational measurement' still has stock in the anglosphere. Then even the psychometricians realised their stuff didn't really have any real-world referent: IQ is just a measure of how good you are at IQ tests, personality tests capture the personality you think you have when you fill them. But, in a paradigmatic application of the philosophical pragmatism in which it is rooted, it doesn't matter if this or that test doesn't measure a real thing, in, like, your miiind, maaan, as long as the construct is valid (a propriety exhibited by tests results), and that you can use it to make meaningful decisions. As I write this I expect it's a bit more complicated, because you wouldn't want to take meaningful decisions on completely made-up stuff, surely... but then *gestures vaguely at the nasty history of psychiatry*

Educational assessment, which can't do Cronbach-style math to validate its constructs, has taken a consequential view of validity: when "empirical evidence and theoretical rationales support the adequacy and appropriateness of inferences and actions based on test scores" (Messick, 1989, p. 13). Your theoretical rationales, these inferences, these actions, this is all, again structure. Notice that, in this definition, the word learning does not feature. Of course, it's nice if in some way assessment verifies learning, but even if assessment doesn't verify anything but how well the student tested in these particular sets of components at these particular points in time, the system would be solid enough to run. There's enough structure that the lack of referent doesn't... wouldn't matter. Surely this is to be voiced in the conditional. But, when you stop to think about it, even if a first in English at Oxford (or a PGCE in Roehampton) indicated nothing more that your erstwhile capacity to play the system well enough to get these qualifications - on these specific years, they would still be doing their job as qualifications.

The lack of referent, of signified, transcendental or mundane, in language, is not a risk of collapse for the reference system because, as a social construct, everyone losing faith in it at the same time is unlikely. Currency, though, that's another matter - the whole frailty of the social construct can show through, and when it does even looks like unravelling, things go bad. Now think of degrees, and backing them up assessment results, as the same kind of referent-less construct. Such a system would only hold as long as those who gives a degree its value, ostensibly employers these days, but to an extent all those in society for whom a degree has currency in status, in social capital, continuing to believe in the value of a degree. Surely this doesn't collapse the way a currency does... but is it even as stable as language?

References

Messick, S. (1989) ‘Validity. em r. linn (org.), educational measurement.(13-103)’, New York, NY: American Council on Education and Macmillan Publishing Company .