ChatGPT loves to collapse
ChatGPT keeps using the word "collapse" when I talk to it. Why is that? Someone even said it to me at work the other day and I noticed, and I began to wonder if that was from being exposed to so much "collapsing" being fed to us from the models.
Why are all concepts seemingly being needed to collapse into themselves? My search history alone in ChatGPT for the word "collapse" returned 51 hits since mid-2023. I'm not that heavy of a user, but damned if it doesn't keep saying it.
The answer to this seems to be a fairly simple one: the way that these models are trained rewarded this word. It must've popped up in a response and the raters (contractors and/or internal folk whose job is to sit and decide if an output is good or not) saw this word and thought "wowie, that word fucking hits in this sentence" and clicked on the button that says "good work, Chat".
Words like "collapse" are words that have high valence, meaning it's highly attractive (or the opposite) and leaves you feeling like something serious has just dropped. It also just feels good to read: collapse. Oooh, something big just happened.
I went back into my ChatGPT history and dug out all the occurrences of the word "collapse" to figure out if it was metaphorical or literal usage.
Here's the stats:
Chart showing occurences of the word "collapse" split by metaphorical or literal usage
After looking at all the ways in which it was used, it seems to take on a metaphorical edge around the end of 2025, at least for me. It also seems to really ramp up usage of the word "collapse" as 2026 gets underway.
I am an equal parts user of ChatGPT, Claude, and Gemini. Neither Claude nor Gemini showed any results for the word "collapse", but this is almost certainly due to me not paying for either of those and therefore not being able to search inside of chats themselves. I'll try to keep tabs on it, but I suspect it does use it a bit.
What struck me on all this is that I do indeed seem to value seeing this word (enough to write about it here, obviously) but I do wonder what this means for how we are deciding to use language ourselves. It all comes back to this reverse-centaur thing I've brought up from Cory Doctorow a couple of times: that we are the ones ending up doing the bidding of the AI.
Yes, it's a mild example, but I feel my brain being formed to receive its delivery. I feel that I want to see "collapse" in my messages, so I know we're talking about deep shit. So I know that I am taking the many threads of a concept and compacting them down into a digestible, understandable nugget that ChatGPT can just plant in my brain in a way that I can regurgitate them later, likely claiming them as my own. Shit, I'm probably doing this already without even realising it.
With an example as innocuous as "collapse", who's to say that all manner of concepts not limited to the singular readable word are not already being slid into our metaphorical DMs without us knowing it. Cue the tinfoil hat, but it's already being studied, with effects as expected.
I like using AI for a lot of brainstorming, coding, history discussion, all manner of nerdy things, and because I love reading, I guess it I value its outputs really highly. The written word to me does sit on a pedestal and likely that'll end up being to my own detriment.