A wonderful Article/Opinion Piece on LLMs’ relationship to human languages. I highly recommend it!
“Denning, Peter, and B. Scot Rousse. 2024. “Can Machines Be in Language?” Commun. ACM 67(3):32–35. doi:10.1145/3637629.”
After a short introduction into how LLMs work and what they are (“basically ANNs”), the authors suggest a distinction between LLMs “suprising capacity to participate in human-like conversations” and “sharing other human abilities conferred by language”. They propose: “LLMs cannot match the ways we humans shape and are shaped by language.”
From a philosophical point of view (here: continental philosophers of language, from v. Humboldt, over Schleiermacher, Schlegel, Ast to Heidegger, Gadamer, Wittgenstein and Adorno), the latter part: “how humans are shaped by language” is the most important part, because it does not posit language as something that only exists when people open their mouths or use their hands to write. Instead, they propose that language exists, in a peculiar way, independently of human beings. We are born “into languages,” we could say. They have an existence that transcends generations, simultaneously binding and glueing different generations together, allowing for exchange and intergenerational and even transcultural understanding.
The authors then continue by examining several ways in which language shapes human life:
- Care (humans care, machines cannot care)
- Shared Spaces of Concern (we actively and cooperatively shape our worlds through conversations; through languages we belong to a “bigger & shared world” / and again, machines do not care, they are not concerned about anything)
- Through languages we make and deliver on commitments which structure our world (e.g. through promises and requests). This reminds us of the ordinary language philosophy & speech act theory regarding insights about how we *do things with words*. Based on the fact that we do things with language, we have future expectations. Words are not simply expressions of statistical patterns, but they convey what, how, who somebody is and what that person will do. Which also means that, as embodied, and as sharing a world, we are responsible for what we say and do. Machines “cannot yet participate in this all-important dance of human language”. LLMs cannot make commitments. We do not hold LLMs responsible for failures; rather, we turn to the designers/programmers.
- Moods and Emotions: “Language permeates our emotional life.” – and through these emotions, we experience the world. Moods are “embodied dispositions that shape the possibilities we can see” (33). This might be one of the most interesting and also most important aspects, because we can see the primordial function of languages the best in this regard. Language are not “added” on secondarily and arbitrarily to an otherwise non-lingual experience of the world. No. We experience our worlds “through languages” and languages “shape the possibilities” of our experiences (cf. 33). To attest this claim, the authors describe how we can get offended by the way someone talks to us (34). Words are never neutral. They carry meaning, Emotions, Feelings, Significance; they are expressions of our commitments and concerns, as well as of the fact that we do care about existence (point 1, 2 and 3). The authors also point to the fact that “competent leaders” “read and flow with moods” when trying to persuade people. A similar point was made by Aristotle, Cicero, and other great thinkers who wrote about Rhetorics, Speeches, and the Link between Languages and Feelings. The Authors conclude: “Only an agent capable of moods and emotions, and the standards and concerns they express, can enter the common space of language as we live it.” (34) To be fair, the authors add that LLMs can “signify emotions,” but they claim that these are only “statistical constructions” that are not based or substantiated by concerns, embodiment, moods, or any concern/care for the world.
- The Background: Languages are inherited, transgenerationally. They are manifestations and are dependent on historical backgrounds. Importantly, what’s in the background, the “sedimentations” of our relations to the world, CAN be made explicit through language. However, as sedimentations, they’re not formalized, listed – they are not even explicit. Which is why we can, by engaging and analyzing and reflection on our language use, bring something to the fore. We can “bring forth new meanings and emotions” – the authors say, claiming that this is what poets do “professionally, by revealing our shared background and transforming our sense of it.” (33) This seems to be an inherent limitation of LLMs that, to me, seems worth mentioning: “it seems unlikely these machines can infer text that has not been written or recorded” (33-34)– How’s that important? It’s important because many essential “facts” or “assumptions” that we share and that we attribute to common sense or to our background knowledge, or inherited and embodied knowledge (tacit knowledge), does simply not exist in “text-form.” It has not been “expressed” or made explicit. And perhaps, I might add, with reference to H. Dreyfus, most of these aspects cannot even be “translated” into text; they cannot be formalized. Ask an expert Jazz pianist to express what they have been doing in their improvisation. Which rules did they apply? What were they thinking and planning? You will get very vague answers, but nothing that would allow a machine to copy exactly that behavior. We have machines that can generate jazz music only because recordings exist. But as the authors themselves write on p. 35: “Although LLMs have generated some surprisingly imaginative poetry, it is more likely that these are unexpected statistical inferences rather than genuine creations relative to the back- ground. This question deserves more exploration.” This would be a big detour for the authors who focus on LLMs and human languages, not on creativity vs. generative AI.
- Embodied Action Beyond Language (continue here)