There are still some material and administrative obstacles to the dream of an ongoing, free-flowing conversation with a chatbot. ChatGPT is often overburdened and tends to time out. Users, when they can access it at all, are limited to a certain number of queries per hour. This reflects the economic realities Open AI faces: Its CEO has described the costs of running the bot as “eye-watering” (one widely circulated estimate put the costs at $3 million a month, though I’m not sure if that includes the cost of periodically retraining the model), and consequently Microsoft may potentially provide $10 billion in funding to facilitate monetization strategies.
From that perspective, ChatGPT is a successful product, for which there has been established a clear pattern of demand. It doesn’t really matter what it does; we already know it “works” — i.e., it attracts interest, it could be something that people will pay for — and hence investment and expansion and integration into various products and practices will follow. (The commercialization process is as indifferent to content and meaning and purpose as the AI models themselves are in their data processing; in that sense, the models are homologous with the capitalist interests fueling their development.) Whatever rationalizations or regulations are required to sustain that momentum toward profitability will be improvised along the way as needed. Much of the commentary about ChatGPT fits into this; it hypes its capabilities and dreams up business models, it announces AI’s irresistible inevitability and postulates all the ways society must now change to accommodate it. Critiques tend to be cast in a defensive, reactionary posture that reinforces the premises of how the models are being hyped: How biased are their outputs? Can they be prevented from confabulating? Will they usurp human creativity? Will they destroy jobs?
Those sorts of questions, important as they are, can seem to accept the fact of consumer demand for AI products at face value. Sometimes critics will draw on the wonder and amazement people supposedly feel about AI to create a sense of urgency around their concerns. They point to what models can already “achieve” or report on their own fascination with the models’ capabilities, how fun it is to play with them, how it feels to get intimations of how their creative power can be harnessed. The idea that technology is a form of irresistible magic returns. It can either fully manipulate populations so that they cannot resist its takeover of society, or its possibilities are so self-evidently beneficial that no one really would want to resist them, except for the class of professional nay-sayers, worry-worts, and others on the wrong side of history who have various vested interests in registering their complaints.
I generally feel that I am on that wrong side, that my discourse is self-marginalizing at best and at worst contributes to the conditions that seem necessary to critique. Often I deal with this through a kind of disavowal, pretending that I am writing in a kind of vacuum. In fact, this is what my experience of using ChatGPT itself has felt like. Whatever I write in its prompt window will get an immediate response that can be interpreted as some sort of expression of the social average without my having to directly interact with any actual people or see myself intersubjectively constituting any part of that social average (even though assuredly all of our words are in there). Since the conversation is not reciprocal, I can see myself as above it, not average, and in a position to judge without being judged.
This is reflected in what I choose to put in the prompts: fake questions that either are so hypothetical they can have no answers (the bot is programmed to issue disclaimers when asked to respond to counterfactuals) or questions that I think I already mostly know the answer to. It seems epistemologically irresponsible to ask questions whose answers from ChatGPT you would be willing to take on faith, since by design it has no truth standard other than statistical averages of past language use, with no consideration of context or meaning or intention let alone polysemy or irony. So whatever its output, I am judging how correct it seems to me, knowing that no amount of dialogue will convince me that its responses can approach truth. A person can convince you that they know what they are talking about, that you can and should learn from them. To come to that conclusion after interacting with ChatGPT would be to fundamentally misunderstand it, or to surrender to a deeply impoverished understanding of “knowledge.”
But who knows? Maybe it will become too practical to take ChatGPT-like models at their word. Once upon a time, people were skeptical that Wikipedia could be relied upon to provide accurate information, but most people pretty much accept it as being accurate enough in most instances now, despite a methodology that can’t guarantee truth. Totalizing language models could move along the same trajectory, becoming more accepted as trustworthy as their methods for refining their outputs become accepted as good enough. With Wikipedia, the idea that people are adding information in good faith that is eventually corrected by other people when it is inaccurate is sufficient to build trust; with generative AI, the algorithms retraining themselves on more and more data will perhaps come to be seen as always approaching some total apprehension of the facts about the world and how they are connected.
It may be that people will not look to generative models for factual information but will instead primarily rely on them to generate texts that don’t participate in knowledge creation. Generative models are good at filling placeholders or producing formulaic material that fulfills formal expectations. That is, they will be good at tackling noncommunicative forms of communication— so good that the opportunities for such interchange may vastly expand: the “automated email responding to automated email” scenario. Some believe this will open up the space for “real” communication; some believe it would crowd it out. A similar question is whether generative models’ capacity for generic, mediocre productions will clarify what human creativity is capable of, or if human creativity will be overwhelmed by their automaticity. Usually, optimists paint a picture of people using “AI tools” to “jump-start” their productivity. Anything that the models can do, they argue, should be understood as commoditized, thus clarifying what it is that people need to add, and what form it should optimally take — the idea that prompt engineering is an emerging art form that sublates all the earlier forms of artistic ideation.
When I open ChatGPT right now, it is usually because I read an article somewhere that described someone using it to do some interesting or surprising thing, but when I confront the prompt window myself, it feels like staring at a blank page. I am curious about what people can make it do, but not all that interested in what I can make it do myself or anything about the specific content it makes, which just seems like effluvia. I don’t have a sense of when I will think, “A generative model will help me get this task done,” or a sense of what will motivate me to develop that sense. It still feels like I would have to change my life, limit myself in certain ways, shut aspects of the world and especially other people out, to accommodate the instrumentality of generative models. I probably won’t recognize the moment I stop feeling that way, even though I will be complaining about its imminence the whole time.
Thanks for reading!