What follows is a reflection on how popular generative AI *things* like ChatGPT and Stable Diffusion are playthings, and why treating them as tools or work-mediating technologies is misunderstanding them. Like many others, I am writing as things happen, and not everything will stick. But I think that thinking in public has its value, and therefore here I am. Before I get into this topic, let me state a fundamental premise of this newsletter: the current crop of popular generative AI systems are at best morally dubious. They are built on stolen data and labor, and they (tend to) create unethical results that require underpaid ghost work to reach us in forms that we find savory. There are ways of using these deep learning techniques that are less unsavory, but even the most ethical approaches only confirm that the master's tools will never dismantle the master's house. Generative AI as we know it today is a colonialist and imperialist project based on extractivist and exploitative politics. Thanks for reading Playing Software! Subscribe for free to receive new posts and support my work. (Researchers like Abeba Birhane and Timnit Gebru show ways of thinking about machine learning in different ways - read and cite their work!). That said, the fact is that we are living in a culture and society in which these systems are being made available for (almost) everybody to use. So we need to think about what these systems are, and how we're using them. And what we are doing with them is playing. Playing Software proposes that play is a way of making sense of computational agency, of what computers can do. When we play, we create relations to others in a temporary world with its own rules and logics. Part of the pleasures of play derive from being with others, in that world, keeping it going together. That is exactly what is happening with these machine learning systems. Think about the prompt. It is not just an interface element to provide some instructions. It is an invitation. The prompt is a gateway to the agent on the other side. Writing a prompt means relating to the system, creating a world, together, intertwined. Writing a good prompt is playing well; writing a prompt injection is creative, maybe even dark play.
Prompts are not just instructions; they are windows in which we show ourselves to the AI. Who we are, how we think, what we desire and what we find fun, or useful, or needful, all of that is contained in that prompt. And solicitous as ever, the system returns us not just an image, or a paragraph, but a part of itself, a partial vision of what it is and what it could be. And there we are, together, in that world we've just created. This worldness explains why these early attempts of using ML systems as work or productivity tools fail. BingSearch is more interesting because it "shows its feelings" (it doesn't) than because it shows us actual good search results. Dall-E and Midjourney and Stable Diffusion are still a long way from substituting actual artists in everything except their cost.
We think about the prompt as if it was a drive-in window through which we can command orders. But it’s not. The prompt is not an instrument to extract some kind of result that we will find worthy. We use the prompt not just to tell these systems to do something, but to let them show themselves to us. The prompt is an opening towards another being, an invitation to relate to it, and to together create something that was not there before. Not a creative process, but a reconfiguration process, following rules, breaking some rules, learning from each other. Playing. Generative systems are not tools for work or productivity. In fact, they are not tools at all. Are they games? I mean, we play with them, right? I don’t think they are games. I think we are witnessing the emergence of a new cultural object that plays with us, that we play with, and that doesn't fall in the traditional categories of games and toys. That’s why they are so exciting, and so scary - because we cannot really make sense of what they are. (yes, I know what these systems are technically, but reducing them to technical objects, now that they are in the wild, is misunderstanding what generative AIs are doing to our world right now). In Playing Software I propose the concept of playthings to describe those things we relate to by playing with them. Games and toys are playthings. So is ChatGPT, or Midjourney. These are systems we make sense of, we interact with, we relate to, by playing. And therefore, we won't be able to understand their role in our society if we don't see them primarily as playthings. I don't know how long this situation will last, but right now, at the beginning of 2023, any attempt of using these systems for producing anything like work is a waste of time. These things are not made for working, they are made for playing. We should embrace their useful uselessness, we should let them say things about us and them like we let games or toys model our world, and we should enjoy them. And if we need to take them seriously, we should do so with the seriousness of play.
On the next newsletter: computational agency and play, and the risks of working with playthings. (and: I’m still finding my rhythm with this public writing things, so: all typos are chatGPTs fault, and all half-baked arguments and pending reflections are to blame to the fact that I am, after all, an unsupervised neural network) Thanks for reading Playing Software! Subscribe for free to receive new posts and support my work. |