In the Beginning was the Word

On the Linguistic Turn in AI

As will surprise no one who knows me, I can be said to be into words. Some might be inclined to push it a little further and say that I am obsessed with words. Guilty as charged, your honour.

It turns out that I am not the only one. In addition to there being lots of my fellow travellers in the community of professional communicators, there is one notable new player in the modern world who is also obsessed with words. And this player came to this obsession somewhat by accident. This player is none other than AI.

On November 30, 2022, OpenAI released ChatGPT. What had already been an accelerating race to “build AI” suddenly went public. Others raced to make AI services public. No one wanted to be left behind. Now several things are interesting about this tipping point. One of them, the one that is most interesting to me and which I believe is the most deeply instructive, is that it highlights what we might call the linguistic turn for AI. Here Large Language Models (LLMs) and the interactive, natural language chatbots they can be used to engender, were universally embraced as the primary interface to AI services. The story of how the OpenAI team was itself completely blindsided by the massive uptake of this new offering sheds some light on the fact that everyone was surprised by the depth and power of what was being tapped.

One of the things being tapped, and I would submit the most important thing, was that texts are more informative than any of us had suspected. This is even true of those of us who, like me, are obsessed with words, and therefore with texts.

It turns out that texts, especially ‘at scale’, provide a window onto the world of lived experience where things happen and where people are tangibly involved. I am inclined to refer to this world of text as being transactionally structured. It puts data into a transactional context and suddenly we see data items not as disembodied records in some data store but in the context of what people want, their intentions, their actions, and their reactions.

There is some irony here. Information technologists and software developers have been famously contemptuous of text as a ‘data type’. It is unstructured, messy, chaotic, voluminous. In this, they did, and do, have a point. But this also misses the point. What the accelerating advance of AI is showing us is that this sprawling world of texts is more than just another input source. This world of texts is emerging as the key input source. It is the source that can provide grounded context to everything else.

Hence we have here a picture from my library of the first edition of Byron’s complete works. One of the salutary lessons we can take from the recent language-driven acceleration of AI is that words matter. More specifically, our words matter. What people write, whether to transact business or to reflect upon it, turns out to be an absolutely vital input source for AI. The more specific, concrete, intentional, and dare we say felt that writing is, the better.

So it is I find myself thinking of Lord Byron, a poet who loved to take action and to do so with his heart on his sleeve. My thoughts then move to his redoubtable daughter, Ada Lovelace, who as a collaborator with Charles Babbage can be seen not only as the first programmer but as the first writer to direct her skills to making clear system objectives and engineering designs that were otherwise fragmentary and undisclosed. The examples of Byron and Ada Lovelace encourage us to pay attention to what AI is reacquainting us with right now, the inescapable truth that in the beginning was the word.