Model workers

Revolution in nonpoetic language

Last week, citing a Business Insider item about that the word delve being possibly overrepresented in ChatGPT outputs, Matt Levine suggested some instances when it would make sense to flaunt that you are using AI to replace yourself:

Having an LLM write your stuff demonstrates that you understand how to use LLMs, which is perhaps the most useful general business skill in 2024.

Probably an LLM is going to read your stuff, to summarize it for your actual audience, so having it sound like an LLM makes it more legible and persuasive.

Even if a human does read your stuff, probably that human is reading a lot of LLM-generated text these days, so she’ll find your stuff more legible and persuasive if it sounds like an LLM.

As an unemployed person, I think about this kind of thing a lot. Perhaps it suits a hiring manager’s idea of resourcefulness when a candidate appears to be using as many tech tools as possible to game the job search and spam employers with readily processable application materials. Perhaps if I had an AI model to generate my résumé, it would be more machine-readable, and I would seem more suitably compliant, less likely to have an ego or an investment in self-expression that might get in the way of productivity. This kind of thinking feeds the logic (warned against in this Bluesky post) that education should mainly focus on training students to use AI, since this will best show their fitness for the workforce.

It seems strange that a technology hyped for its accuracy in guessing the most likely next word would end up “overusing” a word like delve — shouldn’t the math involved produce nothing that surprises any reader? In this Guardian piece Alex Hern suggests the situation can be explained by the number of Nigerians being used to fine-tune generative models’ output. Delve is more apparently more idiomatic in Nigerian English.