For those unfamiliar with the tokenpocalypse (TechCrunch), the term has been around for about a month and the story begins when Microsoft announces a change in the way it would charge for GitHub CoPilot. Unsurprisingly, the service didn’t become cheaper. In fact, it has set off a tokenpocalypse.

Basically this is about AI companies needing to start charging users for the tokens (i.e., compute) that they are using. As Cory Doctorow as put in the several recent interviews, these companies have been handing out $100 bills for $5. And now they need to stop doing that. This 404 video podcast (YouTube) reports on a leaked Accenture meeting recording. I won’t go into the details here, but, TL;DR they realize that they (and their corporate customers) are going to need to figure out how to restrict the ways that AI is being used in the workplace.

In turn, if AI companies are charging by the token then they have a different market incentive. Verbose responses chew through tokens. The overuse of automation in workflow as when someone prompts a bot to turn a folder of files into a slide deck summary of those files. Reasoning inflation, when users select an over-powered model for their task, or when the machine just runs off on its own, as agents obviously do. Interface-prompted iterations as the chatbot functions in an incessant mode of upselling: how about more? And, of course, the entire YOLO AI agent mode where the machine runs through the data version of your life to make it increasingly aligned to you. It turns out that level of sycophancy doesn’t come cheap.

At the same time, it will likely turn out that alignment doesn’t come cheap either. Why would the cheap version of Gemini serve user interests more than Google search does? Meanwhile those who can afford the compute can buy their way into better relationships, provisionally. But that’s not exactly what happens. Even the super-user with a wealth of resources must now manage token use. That person shifts toward the reverse-centaur position as they become responsible for what the AI cannot (or will not) carry.

As the title suggests. it turns out that water is wet. But in this context, whither the AI-forward university?

It’s a little hard to say, obviously, because the new economics are emerging and we haven’t been paying much attention to token use as users I don’t think. So, given that caveat, I’ll divide this into two broad categories.

1. Institutional-administrative uses.
   • This would start with how we encourage the general administrative staff to use AI as part of their workflows. We might also want to throttle (not literally) various dean/provost types who have jobs that beg to be vastly data intensive.
   • Then there’s the premise of an AI interface layer for students, that is, the various FAQ+ helpers. If they stay roughly as helpful as Clippy that would be fine, but if they start trying to offer individualized advice that’s not going to work. It will either be too expensive or grossly underpowered for the task. I mean, we see these “helpful” AIs across platforms now (e.g. Amazon). Does anyone use them?
   • As a silver lining maybe this means we will see a reduction in the AI slopification of institutional discourse. Maybe it turns out not to be economical to generate and spam people with emails.
2. Faculty and student–research and pedagogical uses.
   • We should anticipate tiered access to AI models. Right now access is jagged, but as AI becomes institutionalized it will have to be tiered. At a guess, I’d expect this stratification to resemble the history of such stratifications in terms of who gets access to the most powerful tools.
   • This dramatically alters the concept of AI literacy as it has been broadly discussed. The workplace preparation argument must shift, as what it means to interact with AI will be increasingly different among professions. More importantly, we can dispense with the empowering and democratizing narratives. Instead we can teach AI the way we have been teaching algorithmic culture, surveillance capitalism, etc.

I am not going to go over the implications of an AI bubble. There are plenty of places to read about that debate, and I am not an economist. That said, this is not good news for those insisting a bubble isn’t approaching.

At what point do we say that if we are preparing students for the future workplace that that also means preparing them for a workplace post an AI-bubble? I’m talking about a recessionary economy in which there is a significant withdrawal of investment into data center buildout and related projects. In this future, AI will still exist. It will be corporate boss ware. And it will be social media slop. There will still be super-powered AI for those who can afford it: national defense and policing departments; super-powered research centers; the wealthy.

Of course that tale is more speculative fiction than economic forecasting.

That said, the AI bubble is no more fictional than AGI (or even AI, unless you are sure you can tell me what AI is). In this context, what are the non-cynical reasons for universities to continue to sell their students on some fairly pollyanna visions of an AI future. Which university can bear to not proclaim itself “AI-forward”? After all, we certainly can’t proclaim ourselves to be AI backwards. I think it will be interesting to see what kinds of stories different universities begin to tell around artificial intelligence. Maybe everything will work out “fine.” Regardless, from the perspective of the humanities, the task isn’t to teach students to become “proper” users of AI but rather to provide opportunities for investigating and practicing how to live in an AI-pervasive mediascape.