I get it. You’re sick of hearing about Generative AI. It’s inescapably everywhere right now and if it’s not going to steal your dinner, it’s going to explain to you all the inefficiencies of your current dinner making process and menu, so you wish it just had stolen it anyway.
So let’s talk about hobbies instead.
Like many of you, during the periods of lockdown I started a new hobby. OK, this wasn’t during the first period of the lockdown, but by the second one I’d run out of good books and needed an alternative, so I bought a starter kit for a form of art, thinking that it would distract me for a bit. 2 years or so later and we’re all free to head outside whenever we want and I still choose to spend hours of every week indoors dedicated to working on this, pretty much whenever I can. Or rather, when I’m not looking at others’ work on Instagram for inspiration and reassurance.
As I can’t let a simple joy get in the way of me analyzing everything in my life, I’ve realized – long after everybody else in the world – that making art is founded on two different but complementary elements. There’s the mechanical, the physical techniques which allow you to make the thing and then there’s the compositional, which allows you to think of the thing you wanted to make in the first place. Both of these things I’ve come to realize, need to exist in balance for you to produce something you can – even for a moment – stand back and admire. I mean, that’s never happened to me, but I like to believe at some stage it might.
This becomes an interesting balance when people interact with art which they didn’t create. Which is to say most interactions by most people, with most art. That balance is often imperceptible, especially if the resulting work is non-representative. Seeing something abstract, there’s some which will look at it and utter the words;
“Well, I could have done that!”.
To which the answer should always be, “yes, but you didn’t, did you?”.
What is often seen in that instance is a pure interpretation of the mechanical. And if the mechanics seem simple, then the work loses merit in the eyes of some of those who interpret it without the full context of the composition. The space between the composition and the mechanics – if not filled – can create a vacuum that is filled with bafflement, confusion, even anger. Senses that become even more intense when a monetary value for the work is applied.
When you first start to engage with an artform, the first part you have to focus upon is the mechanical; how do I make the marks on the medium that I’ve chosen to produce the form that I intend. It’s from the results of those formative experiments that composition becomes possible. Iterate, improve and your options in that composition become gradually wider. Outside the personal, looking at works by those more experienced and adept might provide inspiration in both elements; the “how did they do that?” as much as the “why did they do that?”.
Right now, we’re very much at the mechanical phase of learning with generative AI, where there is a lot of public experimentation of what is possible (albeit, in demonstration rather than production form at this stage), without anywhere near as much thought to the compositional value of what is being produced.
Where any thought in that direction is cast, it is generally toward the efficacy of the output, which is an important first step (not least for reasons of legality). Along the way from generating those contemporary models and the general chat apparatus that utilize them, plausibility became unshackled from accuracy. Perhaps missing the intent that was meant by the thought that became the “Turing Test”, leaning into making the structure and timbre of output believable human was overweighted as an outcome.
So let’s consider practically where this leaves us right now, if we believe that in order to create art – whether that be through words or pictures – the balance between the mechanics of creating needs to be balanced with the composition of the final form. The use cases that are being proposed for generative AI are naturally wide at this point, especially as the vehicles for expressing are themselves broad (the generation of written responses and imagery from prompts).
What is significantly underweight is comparing those outputs to those that would come from the same prompts, but through human efforts at composition. It’s arguable that the learning mechanisms that are employed to build the models already do that, but they largely employ forms of learning based on rote reproduction, because that provides an acceptable framework for which to operate. A safety box. As such it is presupposed to trail the leading edge of human creativity, to only produce ghosts of others in its wake.
It’s perhaps natural at this phase to be overly concerned with the mechanics of how to engineer generation, rather than question what is being generated. Once that begins to be questioned, then it is naturally the legality and efficacy that follows, the nature of originality and composition fall far further down the list. Yet it is the human connection with that composition that will ultimately be a key arbiter if it can truly replace – rather than merely imitate – the humans whose written and visual art it learned from.