Font nerds: have you noticed that ChatGPT seems to have learned a "default font" when rendering images? A slightly rounded sans-serif, maybe a little narrow. It's immediately recognizable (which is nice, because it allows me to identify AI slop). It reminds me of Avería, an old project to generate an "average font": http://iotic.com/averia/.

(Image and alt-text from Lucy Rycroft-Smith https://bsky.app/profile/honeypisquared.bsky.social/post/3lsizpkfiuc27)

@csk Please put some kind of mark (like a big red X) over images that are being used as illustration of something bad so that the skimming reader doesn't just see them as the information. It would be unfortunate if some of the formulas on there were confabulated and they stuck in someone's mind as true.
@dalias @csk Too late! Confabulation already succossful
@jimfl @csk LMAO just checked. 🤖 🤡

@jimfl @csk I actually kinda love this image as an example, because it makes it so clear how the model is optimizing to give the *impression*/format for a table of mathematical facts readers might recall seeing on the inside covers of their gradeschool textbooks, to deceive readers into thinking that's what they're seeing, rather than to provide any informational value.

I purposefully didn't look at it in detail before, because I didn't deem "proofreading slop" worthwhile, but I think it's actually instructive to folks who don't already know what these models are doing.

@dalias @jimfl Of course. And Lucy generated the image to make a broader point about the potential damage AI tools might cause in a pedagogical setting (see also https://bsky.app/profile/honeypisquared.bsky.social/post/3lseyr27ea222). I agree with her, but I'm also somewhat fixated on the "neural typography".
Dr Lucy Rycroft-Smith (@honeypisquared.bsky.social)

THIS is the guy you want writing your lesson plans because it will ‘save time’?

Bluesky Social
@dalias @jimfl @csk Yeah, it's brilliant for the less detail focused because every telltale sign is right there if you engage with it for a moment. They're improving the surface level stuff but still very much "not even wrong".
@csk It is weird. It is recognizable, but if somebody asks me, i can't explain it.
@csk also as always with AI, part of it is bullshit.